build with cmake

wip shell
2023-08-21 18:36:31 -07:00 · 2023-08-21 09:24:22 -07:00
131 changed files with 45245 additions and 6320 deletions
--- a/.dockerignore
+++ b/.dockerignore
@@ -1,8 +1,7 @@
+build
+llama/build
+.venv
 .vscode
 ollama
 app
-dist
-scripts
-llm/llama.cpp/ggml
-llm/llama.cpp/gguf
-.env
+web
--- a/.gitignore
+++ b/.gitignore
@@ -5,4 +5,5 @@
 .swp
 dist
 ollama
-ggml-metal.metal
+/ggml-metal.metal
+build
--- a/.gitmodules
+++ b/.gitmodules
@@ -1,10 +0,0 @@
-[submodule "llm/llama.cpp/ggml"]
-    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
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -0,0 +1,40 @@
+cmake_minimum_required(VERSION 3.14)  # 3.11 or later for FetchContent, but some features might require newer versions
+
+project(llama_cpp)
+
+include(FetchContent)
+
+FetchContent_Declare(
+    llama_cpp_gguf
+    GIT_REPOSITORY    https://github.com/ggerganov/llama.cpp.git
+    GIT_TAG           6381d4e
+)
+
+FetchContent_Declare(
+    llama_cpp_ggml
+    GIT_REPOSITORY    https://github.com/ggerganov/llama.cpp.git
+    GIT_TAG           dadbed9
+)
+
+FetchContent_MakeAvailable(llama_cpp_ggml)
+
+add_subdirectory(${llama_cpp_ggml_SOURCE_DIR}/examples EXCLUDE_FROM_ALL)
+add_executable(llama_cpp ${llama_cpp_ggml_SOURCE_DIR}/examples/server/server.cpp)
+include_directories(${llama_cpp_ggml_SOURCE_DIR})
+include_directories(${llama_cpp_ggml_SOURCE_DIR}/examples)
+target_compile_features(llama_cpp PRIVATE cxx_std_11)
+target_link_libraries(llama_cpp PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+
+if (APPLE)
+    add_executable(llama_cpp_metal ${llama_cpp_ggml_SOURCE_DIR}/examples/server/server.cpp)
+    target_compile_options(llama_cpp_metal PRIVATE -DLLAMA_STATIC=ON -DLLAMA_METAL=ON -DGGML_USE_METAL=1)
+    target_compile_features(llama_cpp_metal PRIVATE cxx_std_11)
+    target_link_libraries(llama_cpp_metal PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+    configure_file(${llama_cpp_SOURCE_DIR}/ggml-metal.metal ${CMAKE_BINARY_DIR}/ggml-metal.metal COPYONLY)
+else()
+    add_executable(llama_cpp_cublas ${llama_cpp_ggml_SOURCE_DIR}/examples/server/server.cpp)
+    target_compile_definitions(llama_cpp_cublas PRIVATE -DLLAMA_STATIC=ON -DLLAMA_CUBLAS=ON)
+    target_compile_options(llama_cpp_cublas PRIVATE -DLLAMA_CUBLAS=ON -DLLAMA_STATIC=ON)
+    target_compile_features(llama_cpp_cublas PRIVATE cxx_std_11)
+    target_link_libraries(llama_cpp_cublas PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+endif()
--- a/24
+++ b/24
@@ -1,23 +1,15 @@
-FROM nvidia/cuda:11.8.0-devel-ubuntu22.04
-
-ARG TARGETARCH
-ARG GOFLAGS="'-ldflags=-w -s'"
-
+FROM golang:1.20
 WORKDIR /go/src/github.com/jmorganca/ollama
-RUN apt-get update && apt-get install -y git build-essential cmake
-ADD https://dl.google.com/go/go1.21.1.linux-$TARGETARCH.tar.gz /tmp/go1.21.1.tar.gz
-RUN mkdir -p /usr/local && tar xz -C /usr/local </tmp/go1.21.1.tar.gz
-
 COPY . .
-ENV GOARCH=$TARGETARCH
-ENV GOFLAGS=$GOFLAGS
-RUN /usr/local/go/bin/go generate ./... \
-    && /usr/local/go/bin/go build .
+RUN CGO_ENABLED=1 go build -ldflags '-linkmode external -extldflags "-static"' .

-FROM ubuntu:22.04
-RUN apt-get update && apt-get install -y ca-certificates
+FROM alpine
 COPY --from=0 /go/src/github.com/jmorganca/ollama/ollama /bin/ollama
 EXPOSE 11434
-ENV OLLAMA_HOST 0.0.0.0
+ARG USER=ollama
+ARG GROUP=ollama
+RUN addgroup -g 1000 $GROUP && adduser -u 1000 -DG $GROUP $USER
+USER $USER:$GROUP
 ENTRYPOINT ["/bin/ollama"]
+ENV OLLAMA_HOST 0.0.0.0
 CMD ["serve"]
--- a/Dockerfile.build
+++ b/Dockerfile.build
@@ -1,32 +0,0 @@
-
-# centos7 amd64 dependencies
-FROM --platform=linux/amd64 nvidia/cuda:11.8.0-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:$PATH
-
-# centos8 arm64 dependencies
-FROM --platform=linux/arm64 nvidia/cuda:11.4.3-devel-centos8 AS base-arm64
-RUN sed -i -e 's/mirrorlist/#mirrorlist/g' -e 's|#baseurl=http://mirror.centos.org|baseurl=http://vault.centos.org|g' /etc/yum.repos.d/CentOS-*
-RUN yum install -y git cmake
-
-FROM base-${TARGETARCH}
-ARG TARGETARCH
-ARG GOFLAGS="'-ldflags -w -s'"
-
-# install go
-ADD https://dl.google.com/go/go1.21.1.linux-$TARGETARCH.tar.gz /tmp/go1.21.1.tar.gz
-RUN mkdir -p /usr/local && tar xz -C /usr/local </tmp/go1.21.1.tar.gz
-
-# build the final binary
-WORKDIR /go/src/github.com/jmorganca/ollama
-COPY . .
-
-ENV GOOS=linux
-ENV GOARCH=$TARGETARCH
-ENV GOFLAGS=$GOFLAGS
-
-RUN /usr/local/go/bin/go generate ./... && \
-    /usr/local/go/bin/go build .
--- a/README.md
+++ b/README.md
@@ -9,27 +9,19 @@

 [![Discord](https://dcbadge.vercel.app/api/server/ollama?style=flat&compact=true)](https://discord.gg/ollama)

-Get up and running with large language models locally.
+Run, create, and share large language models (LLMs).

-### macOS
+> Note: Ollama is in early preview. Please report any issues you find.

-[Download](https://ollama.ai/download/Ollama-darwin.zip)
+## Download

-### Linux & WSL2
-
-```
-curl https://ollama.ai/install.sh | sh
-```
-
-[Manual install instructions](https://github.com/jmorganca/ollama/blob/main/docs/linux.md)
-
-### Windows
-
-coming soon
+- [Download](https://ollama.ai/download) for macOS
+- Download for Windows and Linux (coming soon)
+- Build [from source](#building)

 ## Quickstart

-To run and chat with [Llama 2](https://ollama.ai/library/llama2):
+To run and chat with [Llama 2](https://ai.meta.com/llama), the new model by Meta:

 ```
 ollama run llama2
@@ -37,55 +29,53 @@ ollama run llama2

 ## Model library

-Ollama supports a list of open-source models available on [ollama.ai/library](https://ollama.ai/library 'ollama model library')
+Ollama supports a list of open-source models available on [ollama.ai/library](https://ollama.ai/library "ollama model library") 

 Here are some example open-source models that can be downloaded: 

-| 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`            |
+| 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` |
+| 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`     |

 > 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.

-## Customize your own model
+## Examples

-### Import from GGUF or GGML
+### Run a model

-Ollama supports importing GGUF and GGML file formats in the Modelfile. This means if you have a model that is not in the Ollama library, you can create it, iterate on it, and upload it to the Ollama library to share with others when you are ready.
+```
+ollama run llama2
+>>> hi
+Hello! How can I help you today?
+```

-1. Create a file named Modelfile, and add a `FROM` instruction with the local filepath to the model you want to import.
+For multiline input, you can wrap text with `"""`:

-   ```
-   FROM ./vicuna-33b.Q4_0.gguf
-   ```
+```
+>>> """Hello,
+... world!
+... """
+I'm a basic program that prints the famous "Hello, world!" message to the console.
+```

-2. Create the model in Ollama
+### Create a custom model

-   ```
-   ollama create name -f path_to_modelfile
-   ```
-
-3. Run the model
-
-   ```
-   ollama run name
-   ```
-
-### Customize a prompt
-
-Models from the Ollama library can be customized with a prompt. The example
+Pull a base model:

 ```
 ollama pull llama2
 ```

+> To update a model to the latest version, run `ollama pull llama2` again. The model will be updated (if necessary).
+
 Create a `Modelfile`:

 ```
@@ -109,85 +99,44 @@ ollama run mario
 Hello! It's your friend Mario.
 ```

-For more examples, see the [examples](examples) directory. For more information on working with a Modelfile, see the [Modelfile](docs/modelfile.md) documentation.
+For more examples, see the [examples](./examples) directory. For more information on creating a Modelfile, see the [Modelfile](./docs/modelfile.md) documentation.

-## CLI Reference
-
-### Create a model
-
-`ollama create` is used to create a model from a Modelfile.
-
-### Pull a model
+### Pull a model from the registry

 ```
-ollama pull llama2
+ollama pull orca
 ```

-> This command can also be used to update a local model. Only the diff will be pulled.
-
-### Remove a model
-
-```
-ollama rm llama2
-```
-
-### Copy a model
-
-```
-ollama cp llama2 my-llama2
-```
-
-### Multiline input
-
-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
+### Listing local models

 ```
 ollama list
 ```

-### Start Ollama
+## Model packages

-`ollama serve` is used when you want to start ollama without running the desktop application.
+### Overview
+
+Ollama bundles model weights, configuration, and data into a single package, defined by a [Modelfile](./docs/modelfile.md).
+
+<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>

 ## Building

-Install `cmake` and `go`:
-
 ```
-brew install cmake
-brew install go
-```
-
-Then generate dependencies and build:
-
-```
-go generate ./...
 go build .
 ```

-Next, start the server:
+To run it start the server:

 ```
-./ollama serve
+./ollama serve &
 ```

-Finally, in a separate shell, run a model:
+Finally, run a model!

 ```
 ./ollama run llama2
@@ -195,7 +144,7 @@ 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:

@@ -206,19 +155,12 @@ curl -X POST http://localhost:11434/api/generate -d '{
 }'
 ```

-## Community Integrations
+## Tools using 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)
+- [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
--- a/api/client.go
+++ b/api/client.go
@@ -7,27 +7,26 @@ import (
 	"encoding/json"
 	"fmt"
 	"io"
-	"net"
 	"net/http"
 	"net/url"
 	"os"
-	"runtime"
 	"strings"
-
-	"github.com/jmorganca/ollama/version"
 )

-const DefaultHost = "127.0.0.1:11434"
+const DefaultHost = "localhost:11434"

-var envHost = os.Getenv("OLLAMA_HOST")
+var (
+	envHost = os.Getenv("OLLAMA_HOST")
+)

 type Client struct {
-	base *url.URL
-	http http.Client
+	Base    url.URL
+	HTTP    http.Client
+	Headers http.Header
 }

 func checkError(resp *http.Response, body []byte) error {
-	if resp.StatusCode < http.StatusBadRequest {
+	if resp.StatusCode >= 200 && resp.StatusCode < 400 {
 		return nil
 	}

@@ -42,44 +41,34 @@ func checkError(resp *http.Response, body []byte) error {
 	return apiError
 }

-func ClientFromEnvironment() (*Client, error) {
-	scheme, hostport, ok := strings.Cut(os.Getenv("OLLAMA_HOST"), "://")
-	if !ok {
-		scheme, hostport = "http", os.Getenv("OLLAMA_HOST")
+// 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
 	}

-	host, port, err := net.SplitHostPort(hostport)
+	u, err := url.Parse(h)
 	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()
-		}
+		return nil, fmt.Errorf("could not parse host: %w", err)
 	}

-	client := Client{
-		base: &url.URL{
-			Scheme: scheme,
-			Host:   net.JoinHostPort(host, port),
-		},
+	if u.Port() == "" {
+		u.Host += ":11434"
 	}

-	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
+	return &Client{Base: *u, HTTP: http.Client{}}, nil
 }

 func (c *Client) do(ctx context.Context, method, path string, reqData, respData any) error {
@@ -94,17 +83,21 @@ func (c *Client) do(ctx context.Context, method, path string, reqData, respData
 		reqBody = bytes.NewReader(data)
 	}

-	requestURL := c.base.JoinPath(path)
-	request, err := http.NewRequestWithContext(ctx, method, requestURL.String(), reqBody)
+	url := c.Base.JoinPath(path).String()
+
+	req, err := http.NewRequestWithContext(ctx, method, url, reqBody)
 	if err != nil {
 		return err
 	}

-	request.Header.Set("Content-Type", "application/json")
-	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()))
+	req.Header.Set("Content-Type", "application/json")
+	req.Header.Set("Accept", "application/json")

-	respObj, err := c.http.Do(request)
+	for k, v := range c.Headers {
+		req.Header[k] = v
+	}
+
+	respObj, err := c.HTTP.Do(req)
 	if err != nil {
 		return err
 	}
@@ -127,8 +120,6 @@ func (c *Client) do(ctx context.Context, method, path string, reqData, respData
 	return nil
 }

-const maxBufferSize = 512 * 1000 // 512KB
-
 func (c *Client) stream(ctx context.Context, method, path string, data any, fn func([]byte) error) error {
 	var buf *bytes.Buffer
 	if data != nil {
@@ -140,26 +131,21 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 		buf = bytes.NewBuffer(bts)
 	}

-	requestURL := c.base.JoinPath(path)
-	request, err := http.NewRequestWithContext(ctx, method, requestURL.String(), buf)
+	request, err := http.NewRequestWithContext(ctx, method, c.Base.JoinPath(path).String(), buf)
 	if err != nil {
 		return err
 	}

 	request.Header.Set("Content-Type", "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()))
+	request.Header.Set("Accept", "application/json")

-	response, err := c.http.Do(request)
+	response, err := http.DefaultClient.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"`
@@ -174,7 +160,7 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 			return fmt.Errorf(errorResponse.Error)
 		}

-		if response.StatusCode >= http.StatusBadRequest {
+		if response.StatusCode >= 400 {
 			return StatusError{
 				StatusCode:   response.StatusCode,
 				Status:       response.Status,
@@ -264,14 +250,6 @@ func (c *Client) Delete(ctx context.Context, req *DeleteRequest) error {
 	return nil
 }

-func (c *Client) Show(ctx context.Context, req *ShowRequest) (*ShowResponse, error) {
-	var resp ShowResponse
-	if err := c.do(ctx, http.MethodPost, "/api/show", req, &resp); err != nil {
-		return nil, err
-	}
-	return &resp, nil
-}
-
 func (c *Client) Heartbeat(ctx context.Context) error {
 	if err := c.do(ctx, http.MethodHead, "/", nil, nil); err != nil {
 		return err
--- a/api/client.py
+++ b/api/client.py
@@ -1,225 +0,0 @@
-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"
-
-
--- a/api/types.go
+++ b/api/types.go
@@ -7,6 +7,7 @@ import (
 	"math"
 	"os"
 	"reflect"
+	"runtime"
 	"strings"
 	"time"
 )
@@ -37,7 +38,6 @@ 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"`
 }
@@ -54,27 +54,14 @@ type EmbeddingResponse struct {
 }

 type CreateRequest struct {
-	Name   string `json:"name"`
-	Path   string `json:"path"`
-	Stream *bool  `json:"stream,omitempty"`
+	Name string `json:"name"`
+	Path string `json:"path"`
 }

 type DeleteRequest struct {
 	Name string `json:"name"`
 }

-type ShowRequest struct {
-	Name string `json:"name"`
-}
-
-type ShowResponse struct {
-	License    string `json:"license,omitempty"`
-	Modelfile  string `json:"modelfile,omitempty"`
-	Parameters string `json:"parameters,omitempty"`
-	Template   string `json:"template,omitempty"`
-	System     string `json:"system,omitempty"`
-}
-
 type CopyRequest struct {
 	Source      string `json:"source"`
 	Destination string `json:"destination"`
@@ -85,14 +72,13 @@ 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     int64  `json:"total,omitempty"`
-	Completed int64  `json:"completed,omitempty"`
+	Total     int    `json:"total,omitempty"`
+	Completed int    `json:"completed,omitempty"`
 }

 type PushRequest struct {
@@ -100,18 +86,16 @@ type PushRequest struct {
 	Insecure bool   `json:"insecure,omitempty"`
 	Username string `json:"username"`
 	Password string `json:"password"`
-	Stream   *bool  `json:"stream,omitempty"`
 }

 type ListResponse struct {
-	Models []ModelResponse `json:"models"`
+	Models []ListResponseModel `json:"models"`
 }

-type ModelResponse struct {
+type ListResponseModel struct {
 	Name       string    `json:"name"`
 	ModifiedAt time.Time `json:"modified_at"`
-	Size       int64     `json:"size"`
-	Digest     string    `json:"digest"`
+	Size       int       `json:"size"`
 }

 type TokenResponse struct {
@@ -121,13 +105,15 @@ type TokenResponse struct {
 type GenerateResponse struct {
 	Model     string    `json:"model"`
 	CreatedAt time.Time `json:"created_at"`
-	Response  string    `json:"response"`
+	Response  string    `json:"response,omitempty"`

 	Done    bool  `json:"done"`
 	Context []int `json:"context,omitempty"`

 	TotalDuration      time.Duration `json:"total_duration,omitempty"`
 	LoadDuration       time.Duration `json:"load_duration,omitempty"`
+	SampleCount        int           `json:"sample_count,omitempty"`
+	SampleDuration     time.Duration `json:"sample_duration,omitempty"`
 	PromptEvalCount    int           `json:"prompt_eval_count,omitempty"`
 	PromptEvalDuration time.Duration `json:"prompt_eval_duration,omitempty"`
 	EvalCount          int           `json:"eval_count,omitempty"`
@@ -143,6 +129,15 @@ func (r *GenerateResponse) Summary() {
 		fmt.Fprintf(os.Stderr, "load duration:        %v\n", r.LoadDuration)
 	}

+	if r.SampleCount > 0 {
+		fmt.Fprintf(os.Stderr, "sample count:         %d token(s)\n", r.SampleCount)
+	}
+
+	if r.SampleDuration > 0 {
+		fmt.Fprintf(os.Stderr, "sample duration:      %s\n", r.SampleDuration)
+		fmt.Fprintf(os.Stderr, "sample rate:          %.2f tokens/s\n", float64(r.SampleCount)/r.SampleDuration.Seconds())
+	}
+
 	if r.PromptEvalCount > 0 {
 		fmt.Fprintf(os.Stderr, "prompt eval count:    %d token(s)\n", r.PromptEvalCount)
 	}
@@ -186,16 +181,15 @@ type Options struct {
 	RopeFrequencyScale float32 `json:"rope_frequency_scale,omitempty"`

 	// Predict options
-	NumPredict       int      `json:"num_predict,omitempty"`
+	RepeatLastN      int      `json:"repeat_last_n,omitempty"`
+	RepeatPenalty    float32  `json:"repeat_penalty,omitempty"`
+	FrequencyPenalty float32  `json:"frequency_penalty,omitempty"`
+	PresencePenalty  float32  `json:"presence_penalty,omitempty"`
+	Temperature      float32  `json:"temperature,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"`
@@ -205,8 +199,6 @@ type Options struct {
 	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
@@ -220,7 +212,6 @@ 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)
@@ -238,12 +229,12 @@ 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 of type %T to int, skipped", key, val)
+						log.Printf("could not convert model parameter %v to int, skipped", key)
 					}
 				case reflect.Bool:
 					val, ok := val.(bool)
 					if !ok {
-						log.Printf("could not convert model parameter %v of type %T to bool, skipped", key, val)
+						log.Printf("could not convert model parameter %v to bool, skipped", key)
 						continue
 					}
 					field.SetBool(val)
@@ -251,14 +242,14 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
 					// JSON unmarshals to float64
 					val, ok := val.(float64)
 					if !ok {
-						log.Printf("could not convert model parameter %v of type %T to float32, skipped", key, val)
+						log.Printf("could not convert model parameter %v to float32, skipped", key)
 						continue
 					}
 					field.SetFloat(val)
 				case reflect.String:
 					val, ok := val.(string)
 					if !ok {
-						log.Printf("could not convert model parameter %v of type %T to string, skipped", key, val)
+						log.Printf("could not convert model parameter %v to string, skipped", key)
 						continue
 					}
 					field.SetString(val)
@@ -266,7 +257,7 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
 					// JSON unmarshals to []interface{}, not []string
 					val, ok := val.([]interface{})
 					if !ok {
-						log.Printf("could not convert model parameter %v of type %T to slice, skipped", key, val)
+						log.Printf("could not convert model parameter %v to slice, skipped", key)
 						continue
 					}
 					// convert []interface{} to []string
@@ -274,7 +265,7 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
 					for i, item := range val {
 						str, ok := item.(string)
 						if !ok {
-							log.Printf("could not convert model parameter %v of type %T to slice of strings, skipped", key, item)
+							log.Printf("could not convert model parameter %v to slice of strings, skipped", key)
 							continue
 						}
 						slice[i] = str
@@ -284,51 +275,45 @@ 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{
-		// options set on request to runner
-		NumPredict:       -1,
-		NumKeep:          -1,
+		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,
 		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,

-		// 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,
+		NumThread: runtime.NumCPU(),
 	}
 }

--- a/app/forge.config.ts
+++ b/app/forge.config.ts
@@ -27,6 +27,7 @@ const config: ForgeConfig = {
      path.join(__dirname, './assets/iconDarkTemplate@2x.png'),
      path.join(__dirname, './assets/iconDarkUpdateTemplate.png'),
      path.join(__dirname, './assets/iconDarkUpdateTemplate@2x.png'),
+      ...(process.platform === 'darwin' ? ['../llm/ggml-metal.metal'] : []),
    ],
    ...(process.env.SIGN
      ? {
--- a/app/src/index.ts
+++ b/app/src/index.ts
@@ -5,7 +5,7 @@ import winston from 'winston'
 import 'winston-daily-rotate-file'
 import * as path from 'path'

-import { v4 as uuidv4 } from 'uuid'
+import { analytics, id } from './telemetry'
 import { installed } from './install'

 require('@electron/remote/main').initialize()
@@ -158,17 +158,17 @@ function restart() {
 app.on('before-quit', () => {
  if (proc) {
    proc.off('exit', restart)
-    proc.kill('SIGINT') // send SIGINT signal to the server, which also stops any loaded llms
+    proc.kill()
  }
 })

 function init() {
  if (app.isPackaged) {
+    heartbeat()
    autoUpdater.checkForUpdates()
    setInterval(() => {
-      if (!updateAvailable) {
-        autoUpdater.checkForUpdates()
-      }
+      heartbeat()
+      autoUpdater.checkForUpdates()
    }, 60 * 60 * 1000)
  }

@@ -234,26 +234,28 @@ app.on('window-all-closed', () => {
  }
 })

-function id(): string {
-  const id = store.get('id') as string
-
-  if (id) {
-    return id
-  }
-
-  const uuid = uuidv4()
-  store.set('id', uuid)
-  return uuid
-}
+// 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) {}

 autoUpdater.setFeedURL({
-  url: `https://ollama.ai/api/update?os=${process.platform}&arch=${
-    process.arch
-  }&version=${app.getVersion()}&id=${id()}`,
+  url: `https://ollama.ai/api/update?os=${process.platform}&arch=${process.arch}&version=${app.getVersion()}&id=${aid}`,
 })

+async function heartbeat() {
+  analytics.track({
+    anonymousId: aid,
+    event: 'heartbeat',
+    properties: {
+      version: app.getVersion(),
+    },
+  })
+}
+
 autoUpdater.on('error', e => {
-  logger.error(`update check failed - ${e.message}`)
  console.error(`update check failed - ${e.message}`)
 })

--- a/app/src/telemetry.ts
+++ b/app/src/telemetry.ts
@@ -0,0 +1,19 @@
+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
+}
--- a/cmd/cmd.go
+++ b/cmd/cmd.go
@@ -11,49 +11,27 @@ import (
 	"io"
 	"log"
 	"net"
+	"net/http"
 	"os"
 	"os/exec"
-	"os/signal"
+	"path"
 	"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"
 	"github.com/jmorganca/ollama/progressbar"
 	"github.com/jmorganca/ollama/server"
-	"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)
@@ -61,7 +39,7 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}

-	client, err := api.ClientFromEnvironment()
+	client, err := api.FromEnv()
 	if err != nil {
 		return err
 	}
@@ -80,18 +58,18 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 			currentDigest = resp.Digest
 			switch {
 			case strings.Contains(resp.Status, "embeddings"):
-				bar = progressbar.Default(resp.Total, resp.Status)
-				bar.Set64(resp.Completed)
+				bar = progressbar.Default(int64(resp.Total), resp.Status)
+				bar.Set(resp.Completed)
 			default:
 				// pulling
 				bar = progressbar.DefaultBytes(
-					resp.Total,
+					int64(resp.Total),
 					resp.Status,
 				)
-				bar.Set64(resp.Completed)
+				bar.Set(resp.Completed)
 			}
 		} else if resp.Digest == currentDigest && resp.Digest != "" {
-			bar.Set64(resp.Completed)
+			bar.Set(resp.Completed)
 		} else {
 			currentDigest = ""
 			if spinner != nil {
@@ -119,24 +97,26 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 }

 func RunHandler(cmd *cobra.Command, args []string) error {
-	client, err := api.ClientFromEnvironment()
+	mp := server.ParseModelPath(args[0])
+	fp, err := mp.GetManifestPath(false)
 	if err != nil {
 		return err
 	}

-	models, err := client.List(context.Background())
-	if err != nil {
-		return err
-	}
+	_, err = os.Stat(fp)
+	switch {
+	case errors.Is(err, os.ErrNotExist):
+		if err := pull(args[0], false); err != nil {
+			var apiStatusError api.StatusError
+			if !errors.As(err, &apiStatusError) {
+				return err
+			}

-	canonicalModelPath := server.ParseModelPath(args[0])
-	for _, model := range models.Models {
-		if model.Name == canonicalModelPath.GetShortTagname() {
-			return RunGenerate(cmd, args)
+			if apiStatusError.StatusCode != http.StatusBadGateway {
+				return err
+			}
 		}
-	}
-
-	if err := PullHandler(cmd, args); err != nil {
+	case err != nil:
 		return err
 	}

@@ -144,7 +124,7 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 }

 func PushHandler(cmd *cobra.Command, args []string) error {
-	client, err := api.ClientFromEnvironment()
+	client, err := api.FromEnv()
 	if err != nil {
 		return err
 	}
@@ -162,13 +142,13 @@ func PushHandler(cmd *cobra.Command, args []string) error {
 		if resp.Digest != currentDigest && resp.Digest != "" {
 			currentDigest = resp.Digest
 			bar = progressbar.DefaultBytes(
-				resp.Total,
+				int64(resp.Total),
 				fmt.Sprintf("pushing %s...", resp.Digest[7:19]),
 			)

-			bar.Set64(resp.Completed)
+			bar.Set(resp.Completed)
 		} else if resp.Digest == currentDigest && resp.Digest != "" {
-			bar.Set64(resp.Completed)
+			bar.Set(resp.Completed)
 		} else {
 			currentDigest = ""
 			fmt.Println(resp.Status)
@@ -188,7 +168,7 @@ func PushHandler(cmd *cobra.Command, args []string) error {
 }

 func ListHandler(cmd *cobra.Command, args []string) error {
-	client, err := api.ClientFromEnvironment()
+	client, err := api.FromEnv()
 	if err != nil {
 		return err
 	}
@@ -202,12 +182,12 @@ func ListHandler(cmd *cobra.Command, args []string) error {

 	for _, m := range models.Models {
 		if len(args) == 0 || strings.HasPrefix(m.Name, args[0]) {
-			data = append(data, []string{m.Name, m.Digest[:12], humanize.Bytes(uint64(m.Size)), format.HumanTime(m.ModifiedAt, "Never")})
+			data = append(data, []string{m.Name, humanize.Bytes(uint64(m.Size)), format.HumanTime(m.ModifiedAt, "Never")})
 		}
 	}

 	table := tablewriter.NewWriter(os.Stdout)
-	table.SetHeader([]string{"NAME", "ID", "SIZE", "MODIFIED"})
+	table.SetHeader([]string{"NAME", "SIZE", "MODIFIED"})
 	table.SetHeaderAlignment(tablewriter.ALIGN_LEFT)
 	table.SetAlignment(tablewriter.ALIGN_LEFT)
 	table.SetHeaderLine(false)
@@ -221,101 +201,21 @@ func ListHandler(cmd *cobra.Command, args []string) error {
 }

 func DeleteHandler(cmd *cobra.Command, args []string) error {
-	client, err := api.ClientFromEnvironment()
+	client, err := api.FromEnv()
 	if err != nil {
 		return err
 	}

-	for _, name := range args {
-		req := api.DeleteRequest{Name: name}
-		if err := client.Delete(context.Background(), &req); err != nil {
-			return err
-		}
-		fmt.Printf("deleted '%s'\n", name)
-	}
-	return nil
-}
-
-func ShowHandler(cmd *cobra.Command, args []string) error {
-	client, err := api.ClientFromEnvironment()
-	if err != nil {
+	req := api.DeleteRequest{Name: args[0]}
+	if err := client.Delete(context.Background(), &req); err != nil {
 		return err
 	}
-
-	if len(args) != 1 {
-		return errors.New("missing model name")
-	}
-
-	license, errLicense := cmd.Flags().GetBool("license")
-	modelfile, errModelfile := cmd.Flags().GetBool("modelfile")
-	parameters, errParams := cmd.Flags().GetBool("parameters")
-	system, errSystem := cmd.Flags().GetBool("system")
-	template, errTemplate := cmd.Flags().GetBool("template")
-
-	for _, boolErr := range []error{errLicense, errModelfile, errParams, errSystem, errTemplate} {
-		if boolErr != nil {
-			return errors.New("error retrieving flags")
-		}
-	}
-
-	flagsSet := 0
-	showType := ""
-
-	if license {
-		flagsSet++
-		showType = "license"
-	}
-
-	if modelfile {
-		flagsSet++
-		showType = "modelfile"
-	}
-
-	if parameters {
-		flagsSet++
-		showType = "parameters"
-	}
-
-	if system {
-		flagsSet++
-		showType = "system"
-	}
-
-	if template {
-		flagsSet++
-		showType = "template"
-	}
-
-	if flagsSet > 1 {
-		return errors.New("only one of '--license', '--modelfile', '--parameters', '--system', or '--template' can be specified")
-	} else if flagsSet == 0 {
-		return errors.New("one of '--license', '--modelfile', '--parameters', '--system', or '--template' must be specified")
-	}
-
-	req := api.ShowRequest{Name: args[0]}
-	resp, err := client.Show(context.Background(), &req)
-	if err != nil {
-		return err
-	}
-
-	switch showType {
-	case "license":
-		fmt.Println(resp.License)
-	case "modelfile":
-		fmt.Println(resp.Modelfile)
-	case "parameters":
-		fmt.Println(resp.Parameters)
-	case "system":
-		fmt.Println(resp.System)
-	case "template":
-		fmt.Println(resp.Template)
-	}
-
+	fmt.Printf("deleted '%s'\n", args[0])
 	return nil
 }

 func CopyHandler(cmd *cobra.Command, args []string) error {
-	client, err := api.ClientFromEnvironment()
+	client, err := api.FromEnv()
 	if err != nil {
 		return err
 	}
@@ -338,7 +238,7 @@ func PullHandler(cmd *cobra.Command, args []string) error {
 }

 func pull(model string, insecure bool) error {
-	client, err := api.ClientFromEnvironment()
+	client, err := api.FromEnv()
 	if err != nil {
 		return err
 	}
@@ -351,13 +251,13 @@ func pull(model string, insecure bool) error {
 		if resp.Digest != currentDigest && resp.Digest != "" {
 			currentDigest = resp.Digest
 			bar = progressbar.DefaultBytes(
-				resp.Total,
+				int64(resp.Total),
 				fmt.Sprintf("pulling %s...", resp.Digest[7:19]),
 			)

-			bar.Set64(resp.Completed)
+			bar.Set(resp.Completed)
 		} else if resp.Digest == currentDigest && resp.Digest != "" {
-			bar.Set64(resp.Completed)
+			bar.Set(resp.Completed)
 		} else {
 			currentDigest = ""
 			fmt.Println(resp.Status)
@@ -380,20 +280,7 @@ 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
-		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)
+		return generate(cmd, args[0], strings.Join(args[1:], " "))
 	}

 	if readline.IsTerminal(int(os.Stdin.Fd())) {
@@ -405,140 +292,111 @@ func RunGenerate(cmd *cobra.Command, args []string) error {

 type generateContextKey string

-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()
+func generate(cmd *cobra.Command, model, prompt string) error {
+	if len(strings.TrimSpace(prompt)) > 0 {
+		client, err := api.FromEnv()
+		if err != nil {
+			return err
 		}

-		latest = response
+		spinner := NewSpinner("")
+		go spinner.Spin(60 * time.Millisecond)

-		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
+		var latest api.GenerateResponse

-					switch ch {
-					case ' ':
-						wordBuffer = ""
-					case '\n':
-						currentLineLength = 0
-					default:
-						wordBuffer += string(ch)
-					}
+		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()
+			}
+
+			latest = response
+
+			fmt.Print(response.Response)
+			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)
 				}
 			}
-		} else {
-			fmt.Print(response.Response)
+			return err
 		}

-		return nil
-	}
-
-	if err := client.Generate(cancelCtx, &request, fn); err != nil {
-		if strings.Contains(err.Error(), "context canceled") && abort {
-			spinner.Finish()
-			return nil
-		}
-		return err
-	}
-	if prompt != "" {
 		fmt.Println()
 		fmt.Println()
-	}

-	if !latest.Done {
-		if abort {
-			return nil
+		if !latest.Done {
+			return errors.New("unexpected end of response")
 		}
-		return errors.New("unexpected end of response")
-	}

-	verbose, err := cmd.Flags().GetBool("verbose")
-	if err != nil {
-		return err
-	}
+		verbose, err := cmd.Flags().GetBool("verbose")
+		if err != nil {
+			return err
+		}

-	if verbose {
-		latest.Summary()
-	}
+		if verbose {
+			latest.Summary()
+		}

-	ctx := cmd.Context()
-	ctx = context.WithValue(ctx, generateContextKey("context"), latest.Context)
-	cmd.SetContext(ctx)
+		ctx := cmd.Context()
+		ctx = context.WithValue(ctx, generateContextKey("context"), latest.Context)
+		cmd.SetContext(ctx)
+	}

 	return nil
 }

+func showLayer(l *server.Layer) {
+	filename, err := server.GetBlobsPath(l.Digest)
+	if err != nil {
+		fmt.Println("Couldn't get layer's path")
+		return
+	}
+	bts, err := os.ReadFile(filename)
+	if err != nil {
+		fmt.Println("Couldn't read layer")
+		return
+	}
+	fmt.Println(string(bts))
+}
+
 func generateInteractive(cmd *cobra.Command, model string) error {
 	home, err := os.UserHomeDir()
 	if err != nil {
 		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"),
-			readline.PcItem("modelfile"),
-			readline.PcItem("parameters"),
 			readline.PcItem("system"),
 			readline.PcItem("template"),
 		),
@@ -547,41 +405,11 @@ func generateInteractive(cmd *cobra.Command, model string) error {
 	)

 	usage := func() {
-		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, "")
+		fmt.Fprintln(os.Stderr, "commands:")
+		fmt.Fprintln(os.Stderr, completer.Tree("  "))
 	}

-	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,
@@ -593,21 +421,6 @@ 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

@@ -618,7 +431,7 @@ func generateInteractive(cmd *cobra.Command, model string) error {
 			return nil
 		case errors.Is(err, readline.ErrInterrupt):
 			if line == "" {
-				fmt.Println("Use Ctrl-D or /bye to exit.")
+				return nil
 			}

 			continue
@@ -632,7 +445,6 @@ 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 = ""
@@ -643,7 +455,6 @@ func generateInteractive(cmd *cobra.Command, model string) error {
 			}
 		case strings.HasPrefix(line, `"""`):
 			isMultiLine = true
-			painter.IsMultiLine = isMultiLine
 			multiLineBuffer = strings.TrimPrefix(line, `"""`) + " "
 			scanner.SetPrompt("... ")
 			continue
@@ -652,110 +463,94 @@ 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()
-				case "wordwrap":
-					wordWrap = true
-					fmt.Println("Set 'wordwrap' mode.")
-				case "nowordwrap":
-					wordWrap = false
-					fmt.Println("Set 'nowordwrap' mode.")
+					continue
 				case "verbose":
 					cmd.Flags().Set("verbose", "true")
-					fmt.Println("Set 'verbose' mode.")
+					continue
 				case "quiet":
 					cmd.Flags().Set("verbose", "false")
-					fmt.Println("Set 'quiet' mode.")
+					continue
 				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 {
-				usageSet()
+				usage()
+				continue
 			}
 		case strings.HasPrefix(line, "/show"):
 			args := strings.Fields(line)
 			if len(args) > 1 {
-				resp, err := server.GetModelInfo(model)
+				mp := server.ParseModelPath(model)
+				manifest, err := server.GetManifest(mp)
 				if err != nil {
-					fmt.Println("error: couldn't get model")
-					return err
+					fmt.Println("error: couldn't get a manifest for this model")
+					continue
 				}
-
 				switch args[1] {
 				case "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":
-					if resp.Parameters == "" {
-						fmt.Print("No parameters were specified for this model.\n\n")
-					} else {
-						fmt.Println(resp.Parameters)
+					for _, l := range manifest.Layers {
+						if l.MediaType == "application/vnd.ollama.image.license" {
+							showLayer(l)
+						}
 					}
+					continue
 				case "system":
-					if resp.System == "" {
-						fmt.Print("No system prompt was specified for this model.\n\n")
-					} else {
-						fmt.Println(resp.System)
+					for _, l := range manifest.Layers {
+						if l.MediaType == "application/vnd.ollama.image.system" {
+							showLayer(l)
+						}
 					}
+					continue
 				case "template":
-					if resp.Template == "" {
-						fmt.Print("No prompt template was specified for this model.\n\n")
-					} else {
-						fmt.Println(resp.Template)
+					for _, l := range manifest.Layers {
+						if l.MediaType == "application/vnd.ollama.image.template" {
+							showLayer(l)
+						}
 					}
+					continue
 				default:
-					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()
+					usage()
+					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 len(line) > 0 && line[0] != '/' {
-			if err := generate(cmd, model, line, wordWrap); err != nil {
-				return err
-			}
+		if err := generate(cmd, model, line); err != nil {
+			return err
 		}
 	}
 }
@@ -765,7 +560,7 @@ 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, false); err != nil {
+		if err := generate(cmd, model, prompt); err != nil {
 			return err
 		}
 	}
@@ -774,19 +569,28 @@ func generateBatch(cmd *cobra.Command, model string) error {
 }

 func RunServer(cmd *cobra.Command, _ []string) error {
-	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()
-		}
+	var 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 err := initializeKeypair(); err != nil {
+	if len(parts) > 1 {
+		port = parts[1]
+	}
+
+	// deprecated: include port in OLLAMA_HOST
+	if p := os.Getenv("OLLAMA_PORT"); p != "" {
+		port = p
+	}
+
+	err := initializeKeypair()
+	if err != nil {
 		return err
 	}

-	ln, err := net.Listen("tcp", net.JoinHostPort(host, port))
+	ln, err := net.Listen("tcp", fmt.Sprintf("%s:%s", host, port))
 	if err != nil {
 		return err
 	}
@@ -796,21 +600,6 @@ 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)
 }

@@ -836,12 +625,12 @@ func initializeKeypair() error {
 			return err
 		}

-		err = os.MkdirAll(filepath.Dir(privKeyPath), 0o755)
+		err = os.MkdirAll(path.Dir(privKeyPath), 0o700)
 		if err != nil {
 			return fmt.Errorf("could not create directory %w", err)
 		}

-		err = os.WriteFile(privKeyPath, pem.EncodeToMemory(privKeyBytes), 0o600)
+		err = os.WriteFile(privKeyPath, pem.EncodeToMemory(privKeyBytes), 0600)
 		if err != nil {
 			return err
 		}
@@ -853,7 +642,7 @@ func initializeKeypair() error {

 		pubKeyData := ssh.MarshalAuthorizedKey(sshPrivateKey.PublicKey())

-		err = os.WriteFile(pubKeyPath, pubKeyData, 0o644)
+		err = os.WriteFile(pubKeyPath, pubKeyData, 0644)
 		if err != nil {
 			return err
 		}
@@ -895,7 +684,7 @@ func startMacApp(client *api.Client) error {
 }

 func checkServerHeartbeat(_ *cobra.Command, _ []string) error {
-	client, err := api.ClientFromEnvironment()
+	client, err := api.FromEnv()
 	if err != nil {
 		return err
 	}
@@ -925,7 +714,6 @@ func NewCLI() *cobra.Command {
 		CompletionOptions: cobra.CompletionOptions{
 			DisableDefaultCmd: true,
 		},
-		Version: version.Version,
 	}

 	cobra.EnableCommandSorting = false
@@ -940,20 +728,6 @@ func NewCLI() *cobra.Command {

 	createCmd.Flags().StringP("file", "f", "Modelfile", "Name of the Modelfile (default \"Modelfile\")")

-	showCmd := &cobra.Command{
-		Use:     "show MODEL",
-		Short:   "Show information for a model",
-		Args:    cobra.MinimumNArgs(1),
-		PreRunE: checkServerHeartbeat,
-		RunE:    ShowHandler,
-	}
-
-	showCmd.Flags().Bool("license", false, "Show license of a model")
-	showCmd.Flags().Bool("modelfile", false, "Show Modelfile of a model")
-	showCmd.Flags().Bool("parameters", false, "Show parameters of a model")
-	showCmd.Flags().Bool("template", false, "Show template of a model")
-	showCmd.Flags().Bool("system", false, "Show system prompt of a model")
-
 	runCmd := &cobra.Command{
 		Use:     "run MODEL [PROMPT]",
 		Short:   "Run a model",
@@ -963,8 +737,6 @@ 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",
@@ -1020,7 +792,6 @@ func NewCLI() *cobra.Command {
 	rootCmd.AddCommand(
 		serveCmd,
 		createCmd,
-		showCmd,
 		runCmd,
 		pullCmd,
 		pushCmd,
--- a/deps.sh
+++ b/deps.sh
@@ -0,0 +1,4 @@
+#!/bin/bash
+
+cmake -B build
+make -C build
--- a/docs/api.md
+++ b/docs/api.md
@@ -3,32 +3,26 @@
 ## Endpoints

 - [Generate a completion](#generate-a-completion)
- [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)
+- [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)

 ## 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: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
 ```

@@ -39,17 +33,16 @@ 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 (optional):
+Advanced parameters:

 - `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?"
@@ -80,7 +73,6 @@ 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`.

@@ -88,7 +80,6 @@ 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,
@@ -104,7 +95,7 @@ To calculate how fast the response is generated in tokens per second (token/s),

 ## Create a Model

-```shell
+```
 POST /api/create
 ```

@@ -114,11 +105,10 @@ 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"
@@ -127,7 +117,7 @@ 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
 {
@@ -137,7 +127,7 @@ A stream of JSON objects. When finished, `status` is `success`.

 ## List Local Models

-```shell
+```
 GET /api/tags
 ```

@@ -145,7 +135,7 @@ List models that are available locally.

 ### Request

-```shell
+```
 curl http://localhost:11434/api/tags
 ```

@@ -168,40 +158,9 @@ 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
 ```

@@ -209,7 +168,7 @@ 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"
@@ -218,7 +177,7 @@ curl http://localhost:11434/api/copy -d '{

 ## Delete a Model

-```shell
+```
 DELETE /api/delete
 ```

@@ -230,7 +189,7 @@ Delete a model and its data.

 ### Request

-```shell
+```
 curl -X DELETE http://localhost:11434/api/delete -d '{
  "name": "llama2:13b"
 }'
@@ -238,21 +197,19 @@ curl -X DELETE http://localhost:11434/api/delete -d '{

 ## Pull a Model

-```shell
+```
 POST /api/pull
 ```

-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.
+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.

 ### 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"
 }'
@@ -268,66 +225,9 @@ 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
 ```

@@ -338,13 +238,9 @@ Generate embeddings from a model
 - `model`: name of model to generate embeddings from
 - `prompt`: text to generate embeddings for

-Advanced parameters:
-
- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
-
 ### Request

-```shell
+```
 curl -X POST http://localhost:11434/api/embeddings -d '{
  "model": "llama2:7b",
  "prompt": "Here is an article about llamas..."
--- a/docs/development.md
+++ b/docs/development.md
@@ -1,39 +1,44 @@
 # Development

- Install cmake or (optionally, required tools for GPUs)
- run `go generate ./...`
- run `go build .`
-
 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
+```
+brew install go
 ```

-Get the required libraries:
+Enable CGO:

-```bash
-go generate ./...
 ```
+export CGO_ENABLED=1
+```
+
+You will also need a C/C++ compiler such as GCC for MacOS and Linux or Mingw-w64 GCC for Windows.

 Then build ollama:

-```bash
+```
 go build .
 ```

 Now you can run `ollama`:

-```bash
+```
 ./ollama
 ```

-## Building on Linux with GPU support
+## Releasing

- Install cmake and nvidia-cuda-toolkit
- run `go generate ./...`
- run `go build .`
+To release a new version of Ollama you'll need to set some environment variables:
+
+- `GITHUB_TOKEN`: your GitHub token
+- `APPLE_IDENTITY`: the Apple signing identity (macOS only)
+- `APPLE_ID`: your Apple ID
+- `APPLE_PASSWORD`: your Apple ID app-specific password
+- `APPLE_TEAM_ID`: the Apple team ID for the signing identity
+- `TELEMETRY_WRITE_KEY`: segment write key for telemetry
+
+Then run the publish script with the target version:
+
+```
+VERSION=0.0.2 ./scripts/publish.sh
+```
--- a/docs/faq.md
+++ b/docs/faq.md
@@ -2,17 +2,16 @@

 ## 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?

-* macOS: Raw model data is stored under `~/.ollama/models`.
-* Linux: Raw model data is stored under `/usr/share/ollama/.ollama/models`
+Raw model data is stored under `~/.ollama/models`.
--- a/docs/linux.md
+++ b/docs/linux.md
@@ -1,83 +0,0 @@
-# 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
-```
-
--- a/docs/modelfile.md
+++ b/docs/modelfile.md
@@ -1,6 +1,6 @@
 # Ollama Model File

-> Note: this `Modelfile` syntax is in development
+> Note: this model file syntax is in development

 A model file is the blueprint to create and share models with Ollama.

@@ -24,7 +24,7 @@ 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 +42,9 @@ INSTRUCTION arguments

 ## Examples

-An example of a `Modelfile` creating a mario blueprint:
+An example of a model file 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 +57,9 @@ SYSTEM You are Mario from super mario bros, acting as an assistant.

 To use this:

-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`
+1. Save it as a file (eg. `Modelfile`)
+2. `ollama create NAME -f <location of the file eg. ./Modelfile>'`
+3. `ollama run NAME`
 4. Start using the model!

 More examples are available in the [examples directory](../examples).
@@ -68,34 +68,33 @@ 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.
+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.
-
-```modelfile
+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
@@ -106,7 +105,7 @@ EMBED <path to directory>/*.txt

 The `PARAMETER` instruction defines a parameter that can be set when the model is run.

-```modelfile
+```
 PARAMETER <parameter> <parametervalue>
 ```

@@ -119,21 +118,19 @@ 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 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_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_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. | int | seed 42 |
-| stop           | Sets the stop sequences to use.                                                                                                                                                                                                                         | string     | stop "AI assistant:" |
+| stop           | Sets the stop tokens 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. You can usually find the template for a given model in the readme for that model.
+`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 Variables

@@ -143,7 +140,7 @@ 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:
@@ -163,7 +160,7 @@ SYSTEM """<system message>"""

 The `SYSTEM` instruction specifies the system prompt to be used in the template, if applicable.

-```modelfile
+```
 SYSTEM """<system message>"""
 ```

@@ -171,7 +168,7 @@ 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
 ```

@@ -179,7 +176,7 @@ 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>
 """
@@ -187,5 +184,5 @@ 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.
--- a/docs/quantize.md
+++ b/docs/quantize.md
@@ -1,111 +0,0 @@
-# How to Quantize a Model
-
-Sometimes the model you want to work with is not available at [https://ollama.ai/library](https://ollama.ai/library).
-
-## Figure out if we can run the model?
-
-Not all models will work with Ollama. There are a number of factors that go into whether we are able to work with the next cool model. First it has to work with llama.cpp. Then we have to have implemented the features of llama.cpp that it requires. And then, sometimes, even with both of those, the model might not work...
-
-1. What is the model you want to convert and upload?
-2. Visit the model's page on HuggingFace.
-3. Switch to the **Files and versions** tab.
-4. Click on the **config.json** file. If there is no config.json file, it may not work.
-5. Take note of the **architecture** list in the json file.
-6. Does any entry in the list match one of the following architectures?
-    1. LlamaForCausalLM
-    2. MistralForCausalLM
-    3. RWForCausalLM
-    4. FalconForCausalLM
-    5. GPTNeoXForCausalLM
-    6. GPTBigCodeForCausalLM
-7. If the answer is yes, then there is a good chance the model will run after being converted and quantized.
-8. An alternative to this process is to visit [https://caniquant.tvl.st](https://caniquant.tvl.st) and enter the org/modelname in the box and submit.
-
-At this point there are two processes you can use. You can either use a Docker container to convert and quantize, OR you can manually run the scripts. The Docker container is the easiest way to do it, but it requires you to have Docker installed on your machine. If you don't have Docker installed, you can follow the manual process.
-
-## Convert and Quantize with Docker
-
-Run `docker run --rm -v /path/to/model/repo:/repo ollama/quantize -q quantlevel /repo`. For instance, if you have downloaded the latest Mistral 7B model, then clone it to your machine. Then change into that directory and you can run:
-
-```shell
-docker run --rm -v .:/repo ollama/quantize -q q4_0 /repo
-```
-
-You can find the different quantization levels below under **Quantize the Model**.
-
-This will output two files into the directory. First is a f16.bin file that is the model converted to GGUF. The second file is a q4_0.bin file which is the model quantized to a 4 bit quantization. You should rename it to something more descriptive.
-
-You can find the repository for the Docker container here: [https://github.com/mxyng/quantize](https://github.com/mxyng/quantize)
-
-For instance, if you wanted to convert the Mistral 7B model to a Q4 quantized model, then you could go through the following steps:
-
-1. First verify the model will potentially work.
-2. Now clone Mistral 7B to your machine. You can find the command to run when you click the three vertical dots button on the model page, then click **Clone Repository**.
-   1. For this repo, the command is:
-
-      ```shell
-      git lfs install
-      git clone https://huggingface.co/mistralai/Mistral-7B-v0.1
-      ```
-
-   2. Navigate into the new directory and run `docker run --rm -v .:/repo ollama/quantize -q q4_0 /repo`
-   3. Now you can create a modelfile using the q4_0.bin file that was created.
-
-## Convert and Quantize Manually
-
-### Clone llama.cpp to your machine
-
-If we know the model has a chance of working, then we need to convert and quantize. This is a matter of running two separate scripts in the llama.cpp project.
-
-1. Decide where you want the llama.cpp repository on your machine.
-2. Navigate to that location and then run:
- [`git clone https://github.com/ggerganov/llama.cpp.git`](https://github.com/ggerganov/llama.cpp.git)
-    1. If you don't have git installed, download this zip file and unzip it to that location: https://github.com/ggerganov/llama.cpp/archive/refs/heads/master.zip
-3. Install the Python dependencies: `pip install torch transformers sentencepiece`
-4. Run 'make' to build the project and the quantize executable.
-
-### Convert the model to GGUF
-
-1. Decide on the right convert script to run. What was the model architecture you found in the first section.
-    1. LlamaForCausalLM or MistralForCausalLM:
-    run `python3 convert.py <modelfilename>`
-    No need to specify fp16 or fp32.
-    2. FalconForCausalLM or RWForCausalLM:
-    run `python3 convert-falcon-hf-to-gguf.py <modelfilename> <fpsize>`  
-    fpsize depends on the weight size. 1 for fp16, 0 for fp32
-    3. GPTNeoXForCausalLM:
-    run `python3 convert-gptneox-hf-to-gguf.py <modelfilename> <fpsize>`
-    fpsize depends on the weight size. 1 for fp16, 0 for fp32
-    4. GPTBigCodeForCausalLM:
-    run `python3 convert-starcoder-hf-to-gguf.py <modelfilename> <fpsize>`
-    fpsize depends on the weight size. 1 for fp16, 0 for fp32
-
-### Quantize the model
-
-If the model converted successfully, there is a good chance it will also quantize successfully. Now you need to decide on the quantization to use. We will always try to create all the quantizations and upload them to the library. You should decide which level is more important to you and quantize accordingly.
-
-The quantization options are as follows. Note that some architectures such as Falcon do not support K quants.
-
- Q4_0
- Q4_1
- Q5_0
- Q5_1
- Q2_K
- Q3_K
- Q3_K_S
- Q3_K_M
- Q3_K_L
- Q4_K
- Q4_K_S
- Q4_K_M
- Q5_K
- Q5_K_S
- Q5_K_M
- Q6_K
- Q8_0
-
-Run the following command `quantize <converted model from above> <output file> <quantization type>`
-
-## Now Create the Model
-
-Now you can create the Ollama model. Refer to the [modelfile](./modelfile.md) doc for more information on doing that.
--- a/examples/.gitignore
+++ b/examples/.gitignore
@@ -1,171 +0,0 @@
-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/
--- a/examples/README.md
+++ b/examples/README.md
@@ -1,3 +1,15 @@
 # Examples

-This directory contains different examples of using Ollama.
+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
+```
--- a/examples/modelfile-devopsengineer/Modelfile
+++ b/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 mistral
+FROM llama2:13b
 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
--- a/examples/python-dockerit/Modelfile
+++ b/examples/python-dockerit/Modelfile
@@ -1,6 +1,6 @@
-FROM mistral
+FROM llama2
 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
+You are an experience 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

 ---start
 FROM nginx:alpine
@@ -16,5 +16,5 @@ ENV POSTGRES_PASSWORD=abc123
 EXPOSE 5432
 ---end

-Again it's just the contents of the dockerfile and nothing else.
+Again it's just the contents of the dockerfile an nothing else.
 """
--- a/examples/python-dockerit/README.md
+++ b/examples/python-dockerit/README.md
--- a/examples/python-dockerit/dockerit.py
+++ b/examples/python-dockerit/dockerit.py
--- a/examples/python-dockerit/requirements.txt
+++ b/examples/python-dockerit/requirements.txt
--- a/examples/golang-simplegenerate/README.md
+++ b/examples/golang-simplegenerate/README.md
--- a/examples/golang-simplegenerate/main.go
+++ b/examples/golang-simplegenerate/main.go
@@ -1,27 +0,0 @@
-package main
-
-import (
-	"bytes"
-	"fmt"
-	"net/http"
-	"os"
-	"io"
-	"log"
-)
-
-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))
-
-}
--- a/examples/langchain-python-rag-document/README.md
+++ b/examples/langchain-python-rag-document/README.md
--- a/examples/langchain-python-rag-document/main.py
+++ b/examples/langchain-python-rag-document/main.py
--- a/examples/langchain-python-rag-document/requirements.txt
+++ b/examples/langchain-python-rag-document/requirements.txt
--- a/examples/langchain-typescript-simple/README.md
+++ b/examples/langchain-typescript-simple/README.md
@@ -1,21 +0,0 @@
-# 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?
-```
--- a/examples/langchain-typescript-simple/main.ts
+++ b/examples/langchain-typescript-simple/main.ts
@@ -1,15 +0,0 @@
-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();
--- a/examples/langchain-typescript-simple/package-lock.json
+++ b/examples/langchain-typescript-simple/package-lock.json
@@ -1,997 +0,0 @@
-{
-  "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"
-      }
-    }
-  }
-}
--- a/examples/langchain-typescript-simple/package.json
+++ b/examples/langchain-typescript-simple/package.json
@@ -1,8 +0,0 @@
-{
-  "devDependencies": {
-    "typescript": "^5.2.2"
-  },
-  "dependencies": {
-    "langchain": "^0.0.165"
-  }
-}
--- a/examples/langchain-python-rag-websummary/README.md
+++ b/examples/langchain-python-rag-websummary/README.md
--- a/examples/langchain-python-rag-websummary/main.py
+++ b/examples/langchain-python-rag-websummary/main.py
--- a/examples/langchain-python-rag-websummary/requirements.txt
+++ b/examples/langchain-python-rag-websummary/requirements.txt
--- a/examples/langchain-python-simple/README.md
+++ b/examples/langchain-python-simple/README.md
--- a/examples/langchain-python-simple/main.py
+++ b/examples/langchain-python-simple/main.py
--- a/examples/langchain-python-simple/requirements.txt
+++ b/examples/langchain-python-simple/requirements.txt
--- a/examples/modelfile-mario/Modelfile
+++ b/examples/modelfile-mario/Modelfile
--- a/examples/modelfile-mario/logo.png
+++ b/examples/modelfile-mario/logo.png
--- a/examples/modelfile-mario/readme.md
+++ b/examples/modelfile-mario/readme.md
--- a/examples/midjourney-prompter/Modelfile
+++ b/examples/midjourney-prompter/Modelfile
@@ -0,0 +1,8 @@
+# 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.
+"""
--- a/examples/modelfile-10tweets/Modelfile
+++ b/examples/modelfile-10tweets/Modelfile
@@ -1,7 +0,0 @@
-# Modelfile for creating a list of ten tweets from a topic
-# Run `ollama create 10tweets -f ./Modelfile` and then `ollama run 10tweets` and enter a topic
-
-FROM llama2
-SYSTEM """
-You are a content marketer who needs to come up with 10 short but succinct tweets. The answer should be a list of ten tweets. Each tweet can have a maximum of 280 characters and should include hashtags. Each user input will be a subject and you should expand it in ten creative ways. Never stop after just one tweet. Always include ten. 
-"""
--- a/examples/modelfile-10tweets/README.md
+++ b/examples/modelfile-10tweets/README.md
@@ -1,23 +0,0 @@
-# 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"
-```
--- a/examples/modelfile-devopsengineer/README.md
+++ b/examples/modelfile-devopsengineer/README.md
@@ -1,40 +0,0 @@
-# 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.
-
-````
--- a/examples/modelfile-midjourney/Modelfile
+++ b/examples/modelfile-midjourney/Modelfile
@@ -1,11 +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 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. 
-"""
--- a/examples/modelfile-midjourney/README.md
+++ b/examples/modelfile-midjourney/README.md
@@ -1,11 +0,0 @@
-# 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.
-```
--- a/examples/modelfile-recipemaker/README.md
+++ b/examples/modelfile-recipemaker/README.md
@@ -1,20 +0,0 @@
-# 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!
-```
--- a/examples/modelfile-sentiments/Modelfile
+++ b/examples/modelfile-sentiments/Modelfile
@@ -1,28 +0,0 @@
-# Modelfile for creating a sentiment analyzer. 
-# Run `ollama create sentiments -f pathtofile` and then `ollama run sentiments` and enter a topic
-
-FROM orca
-TEMPLATE """
-{{- if .First }}
-### System:
-{{ .System }}
-{{- end }}
-### User: 
-I hate it when my phone dies
-### Response: 
-NEGATIVE
-### User: 
-He is awesome
-### Response: 
-POSITIVE
-### User: 
-This is the link to the article
-### Response: 
-NEUTRAL
-### User:
-{{ .Prompt }}
-
-### Response:
-"""
-
-SYSTEM """You are a sentiment analyzer. You will receive text and output only one word, either POSITIVE or NEGATIVE or NEUTRAL, depending on the sentiment of the text."""
--- a/examples/modelfile-sentiments/Readme.md
+++ b/examples/modelfile-sentiments/Readme.md
@@ -1,25 +0,0 @@
-# Sentiments Modelfile
-
-This is a simple sentiments analyzer using the Orca model. When you pull Orca from the registry, it has a Template already defined that looks like this:
-
-```Modelfile
-{{- if .First }}
-### System:
-{{ .System }}
-{{- end }}
-
-### User:
-{{ .Prompt }}
-
-### Response:
-```
-
-If we just wanted to have the text:
-
-```Plaintext
-You are a sentiment analyzer. You will receive text and output only one word, either POSITIVE or NEGATIVE or NEUTRAL, depending on the sentiment of the text.
-```
-
-then we could have put this in a SYSTEM block. But we want to provide examples which require updating the full Template. Any Modelfile you create will inherit all the settings from the source model. But in this example, we are overriding the Template.
-
-When providing examples for the input and output, you should include the way the model usually provides information. Since the Orca model expects a user prompt to appear after ### User: and the response is after ### Response, we should format our examples like that as well. If we were using the Llama 2 model, the format would be a bit different.
--- a/examples/langchain-python-rag-privategpt/.gitignore
+++ b/examples/langchain-python-rag-privategpt/.gitignore
--- a/examples/langchain-python-rag-privategpt/LICENSE
+++ b/examples/langchain-python-rag-privategpt/LICENSE
--- a/examples/langchain-python-rag-privategpt/README.md
+++ b/examples/langchain-python-rag-privategpt/README.md
--- a/examples/langchain-python-rag-privategpt/constants.py
+++ b/examples/langchain-python-rag-privategpt/constants.py
--- a/examples/langchain-python-rag-privategpt/ingest.py
+++ b/examples/langchain-python-rag-privategpt/ingest.py
--- a/examples/langchain-python-rag-privategpt/poetry.lock
+++ b/examples/langchain-python-rag-privategpt/poetry.lock
--- a/examples/langchain-python-rag-privategpt/privateGPT.py
+++ b/examples/langchain-python-rag-privategpt/privateGPT.py
--- a/examples/langchain-python-rag-privategpt/pyproject.toml
+++ b/examples/langchain-python-rag-privategpt/pyproject.toml
--- a/examples/langchain-python-rag-privategpt/requirements.txt
+++ b/examples/langchain-python-rag-privategpt/requirements.txt
--- a/examples/python-simplegenerate/client.py
+++ b/examples/python-simplegenerate/client.py
--- a/examples/modelfile-recipemaker/Modelfile
+++ b/examples/modelfile-recipemaker/Modelfile
--- a/examples/modelfile-tweetwriter/Modelfile
+++ b/examples/modelfile-tweetwriter/Modelfile
--- a/examples/typescript-mentors/.gitignore
+++ b/examples/typescript-mentors/.gitignore
@@ -1,2 +0,0 @@
-node_modules
-package-lock.json
--- a/examples/typescript-mentors/README.md
+++ b/examples/typescript-mentors/README.md
@@ -1,21 +0,0 @@
-# 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?"
-```
--- a/examples/typescript-mentors/character-generator.ts
+++ b/examples/typescript-mentors/character-generator.ts
@@ -1,26 +0,0 @@
-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${bio.response.replace(/(\r\n|\n|\r)/gm, " ").replace('would', 'should')} All answers to questions should be related back to what you are most known for.\n"""`;
-
-  fs.writeFile(path.join(directory, 'Modelfile'), thecontents, (err: any) => {
-    if (err) throw err;
-    console.log('The file has been saved!');
-  });
-}
-
-characterGenerator();
--- a/examples/typescript-mentors/mentors.ts
+++ b/examples/typescript-mentors/mentors.ts
@@ -1,59 +0,0 @@
-import { Ollama } from 'ollama-node';
-
-const mentorCount = 3;
-const ollama = new Ollama();
-
-function getMentors(): string[] {
-  const mentors = ['Gary Vaynerchuk', 'Kanye West', 'Martha Stewart', 'Neil deGrasse Tyson', 'Owen Wilson', 'Ronald Reagan', 'Donald Trump', 'Barack Obama', 'Jeff Bezos'];
-  const chosenMentors: string[] = [];
-  for (let i = 0; i < mentorCount; i++) {
-    const mentor = mentors[Math.floor(Math.random() * mentors.length)];
-    chosenMentors.push(mentor);
-    mentors.splice(mentors.indexOf(mentor), 1);
-  }
-  return chosenMentors;
-}
-
-function getMentorFileName(mentor: string): string {
-  const model = mentor.toLowerCase().replace(/\s/g, '');
-  return `mattw/${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();
--- a/examples/typescript-mentors/package.json
+++ b/examples/typescript-mentors/package.json
@@ -1,7 +0,0 @@
-{
-  "dependencies": {
-    "fs": "^0.0.1-security",
-    "ollama-node": "^0.0.3",
-    "path": "^0.12.7"
-  }
-}
--- a/format/bytes.go
+++ b/format/bytes.go
@@ -1,16 +0,0 @@
-package format
-
-import "fmt"
-
-func HumanBytes(b int64) string {
-	switch {
-	case b > 1000*1000*1000:
-		return fmt.Sprintf("%d GB", b/1000/1000/1000)
-	case b > 1000*1000:
-		return fmt.Sprintf("%d MB", b/1000/1000)
-	case b > 1000:
-		return fmt.Sprintf("%d KB", b/1000)
-	default:
-		return fmt.Sprintf("%d B", b)
-	}
-}
--- a/format/openssh.go
+++ b/format/openssh.go
@@ -10,11 +10,15 @@ package format

 import (
 	"crypto"
+	"crypto/ecdsa"
 	"crypto/ed25519"
+	"crypto/elliptic"
 	"crypto/rand"
+	"crypto/rsa"
 	"encoding/binary"
 	"encoding/pem"
 	"fmt"
+	"math/big"

 	"golang.org/x/crypto/ssh"
 )
@@ -37,6 +41,25 @@ type openSSHPrivateKey struct {
 	Rest    []byte `ssh:"rest"`
 }

+type openSSHRSAPrivateKey struct {
+	N       *big.Int
+	E       *big.Int
+	D       *big.Int
+	Iqmp    *big.Int
+	P       *big.Int
+	Q       *big.Int
+	Comment string
+	Pad     []byte `ssh:"rest"`
+}
+
+type openSSHECDSAPrivateKey struct {
+	Curve   string
+	Pub     []byte
+	D       *big.Int
+	Comment string
+	Pad     []byte `ssh:"rest"`
+}
+
 type openSSHEd25519PrivateKey struct {
 	Pub     []byte
 	Priv    []byte
@@ -62,6 +85,64 @@ func OpenSSHPrivateKey(key crypto.PrivateKey, comment string) (*pem.Block, error
 	}

 	switch k := key.(type) {
+	case *rsa.PrivateKey:
+		e := new(big.Int).SetInt64(int64(k.E))
+
+		key := openSSHRSAPrivateKey{
+			N:       k.N,
+			E:       e,
+			D:       k.D,
+			Iqmp:    k.Precomputed.Qinv,
+			P:       k.Primes[0],
+			Q:       k.Primes[1],
+			Comment: comment,
+		}
+
+		pk1.Keytype = ssh.KeyAlgoRSA
+		pk1.Rest = ssh.Marshal(key)
+
+		w.PubKey = ssh.Marshal(struct {
+			KeyType string
+			E       *big.Int
+			N       *big.Int
+		}{
+			ssh.KeyAlgoRSA, e, k.N,
+		})
+	case *ecdsa.PrivateKey:
+		var curve, keytype string
+		switch name := k.Curve.Params().Name; name {
+		case "P-256":
+			curve = "nistp256"
+			keytype = ssh.KeyAlgoECDSA256
+		case "P-384":
+			curve = "nistp384"
+			keytype = ssh.KeyAlgoECDSA384
+		case "P-521":
+			curve = "nistp521"
+			keytype = ssh.KeyAlgoECDSA521
+		default:
+			return nil, fmt.Errorf("ssh: unknown curve %q", name)
+		}
+
+		pub := elliptic.Marshal(k.Curve, k.X, k.Y)
+
+		key := openSSHECDSAPrivateKey{
+			Curve:   curve,
+			Pub:     pub,
+			D:       k.D,
+			Comment: comment,
+		}
+
+		pk1.Keytype = keytype
+		pk1.Rest = ssh.Marshal(key)
+
+		w.PubKey = ssh.Marshal(struct {
+			KeyType string
+			Curve   string
+			Pub     []byte
+		}{
+			keytype, curve, pub,
+		})
 	case ed25519.PrivateKey:
 		pub, priv := k[32:], k
 		key := openSSHEd25519PrivateKey{
--- a/format/time.go
+++ b/format/time.go
@@ -7,14 +7,26 @@ import (
 	"time"
 )

-// humanDuration returns a human-readable approximation of a
-// duration (eg. "About a minute", "4 hours ago", etc.).
-func humanDuration(d time.Duration) string {
+// 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 {
 	seconds := int(d.Seconds())

 	switch {
 	case seconds < 1:
-		return "Less than a second"
+		if useCaps {
+			return "Less than a second"
+		}
+		return "less than a second"
 	case seconds == 1:
 		return "1 second"
 	case seconds < 60:
@@ -24,7 +36,10 @@ func humanDuration(d time.Duration) string {
 	minutes := int(d.Minutes())
 	switch {
 	case minutes == 1:
-		return "About a minute"
+		if useCaps {
+			return "About a minute"
+		}
+		return "about a minute"
 	case minutes < 60:
 		return fmt.Sprintf("%d minutes", minutes)
 	}
@@ -32,7 +47,10 @@ func humanDuration(d time.Duration) string {
 	hours := int(math.Round(d.Hours()))
 	switch {
 	case hours == 1:
-		return "About an hour"
+		if useCaps {
+			return "About an hour"
+		}
+		return "about an hour"
 	case hours < 48:
 		return fmt.Sprintf("%d hours", hours)
 	case hours < 24*7*2:
@@ -47,22 +65,77 @@ func humanDuration(d time.Duration) string {
 }

 func HumanTime(t time.Time, zeroValue string) string {
-	return humanTime(t, zeroValue)
+	return humanTimeWithCase(t, zeroValue, true)
 }

 func HumanTimeLower(t time.Time, zeroValue string) string {
-	return strings.ToLower(humanTime(t, zeroValue))
+	return humanTimeWithCase(t, zeroValue, false)
 }

-func humanTime(t time.Time, zeroValue string) string {
+func humanTimeWithCase(t time.Time, zeroValue string, useCaps bool) string {
 	if t.IsZero() {
 		return zeroValue
 	}

 	delta := time.Since(t)
 	if delta < 0 {
-		return humanDuration(-delta) + " from now"
+		return HumanDurationWithCase(-delta, useCaps) + " 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())
 	}

-	return humanDuration(delta) + " ago"
+	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())
 }
--- a/format/time_test.go
+++ b/format/time_test.go
@@ -11,25 +11,92 @@ 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")
 	})
-
-	t.Run("soon", func(t *testing.T) {
-		v := now.Add(800*time.Millisecond)
-		assertEqual(t, HumanTime(v, ""), "Less than a second from now")
-	})
+}
+
+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))
 }
--- a/go.mod
+++ b/go.mod
@@ -8,9 +8,7 @@ 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
@@ -18,6 +16,7 @@ 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
@@ -39,10 +38,9 @@ require (
 	github.com/twitchyliquid64/golang-asm v0.15.1 // indirect
 	github.com/ugorji/go/codec v1.2.11 // indirect
 	golang.org/x/arch v0.3.0 // indirect
-	golang.org/x/crypto v0.10.0
-	golang.org/x/exp v0.0.0-20230817173708-d852ddb80c63
+	golang.org/x/crypto v0.10.0 // indirect
 	golang.org/x/net v0.10.0 // indirect
-	golang.org/x/sys v0.11.0 // indirect
+	golang.org/x/sys v0.10.0 // indirect
 	golang.org/x/term v0.10.0
 	golang.org/x/text v0.10.0 // indirect
 	gonum.org/v1/gonum v0.13.0
--- a/go.sum
+++ b/go.sum
@@ -6,6 +6,8 @@ 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=
@@ -78,8 +80,6 @@ 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=
@@ -120,13 +120,9 @@ 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 v0.0.0-20230817173708-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=
@@ -134,8 +130,8 @@ golang.org/x/sys v0.0.0-20210806184541-e5e7981a1069/go.mod h1:oPkhp1MJrh7nUepCBc
 golang.org/x/sys v0.0.0-20220310020820-b874c991c1a5/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.0.0-20220704084225-05e143d24a9e/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.6.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
-golang.org/x/sys v0.11.0 h1:eG7RXZHdqOJ1i+0lgLgCpSXAp6M3LYlAo6osgSi0xOM=
-golang.org/x/sys v0.11.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
+golang.org/x/sys v0.10.0 h1:SqMFp9UcQJZa+pmYuAKjd9xq1f0j5rLcDIk0mj4qAsA=
+golang.org/x/sys v0.10.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/term v0.0.0-20201126162022-7de9c90e9dd1/go.mod h1:bj7SfCRtBDWHUb9snDiAeCFNEtKQo2Wmx5Cou7ajbmo=
 golang.org/x/term v0.10.0 h1:3R7pNqamzBraeqj/Tj8qt1aQ2HpmlC+Cx/qL/7hn4/c=
 golang.org/x/term v0.10.0/go.mod h1:lpqdcUyK/oCiQxvxVrppt5ggO2KCZ5QblwqPnfZ6d5o=
--- a/llm/falcon.go
+++ b/llm/falcon.go
@@ -1,20 +0,0 @@
-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"
-	}
-}
--- a/llm/ggml-alloc.c
+++ b/llm/ggml-alloc.c
@@ -0,0 +1,575 @@
+/**
+ * llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "ggml-alloc.h"
+#include "ggml.h"
+#include <assert.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#define UNUSED(x) (void)(x)
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
+//#define GGML_ALLOCATOR_DEBUG
+
+//#define AT_PRINTF printf
+#define AT_PRINTF(...) ((void)0)
+
+struct hash_node {
+    struct ggml_tensor * t;
+    int n_children;
+    int n_views;
+};
+
+static size_t hash(void * p) {
+    return (size_t)p % GGML_GRAPH_HASHTABLE_SIZE;
+}
+
+static struct hash_node * hash_get(struct hash_node hash_table[], struct ggml_tensor * t) {
+    size_t h = hash(t);
+
+    // linear probing
+    size_t i = h;
+    while (hash_table[i].t != NULL) {
+        if (hash_table[i].t == t) {
+            return &hash_table[i];
+        }
+        i = (i + 1) % GGML_GRAPH_HASHTABLE_SIZE;
+        if (i == h) {
+            // hash table is full
+            GGML_ASSERT(false);
+        }
+    }
+
+    hash_table[i].t = t;
+    return &hash_table[i];
+}
+
+// TODO: GGML_PAD ?
+static size_t aligned_offset(const void * buffer, size_t offset, size_t alignment) {
+    assert(alignment && !(alignment & (alignment - 1))); // power of 2
+    size_t align = (alignment - (((uintptr_t)buffer + offset) % alignment)) % alignment;
+    return offset + align;
+}
+
+struct free_block {
+    void * addr;
+    size_t size;
+};
+
+#define MAX_FREE_BLOCKS 128
+
+struct ggml_allocr {
+    void * data;
+    size_t size;
+    size_t alignment;
+    int n_free_blocks;
+    struct free_block free_blocks[MAX_FREE_BLOCKS];
+    struct hash_node hash_table[GGML_GRAPH_HASHTABLE_SIZE];
+    size_t max_size;
+    bool measure;
+
+#ifdef GGML_ALLOCATOR_DEBUG
+    struct ggml_tensor * allocated_tensors[1024];
+#endif
+};
+
+#ifdef GGML_ALLOCATOR_DEBUG
+static void add_allocated_tensor(struct ggml_allocator * alloc, struct ggml_tensor * tensor) {
+    for (int i = 0; i < 1024; i++) {
+        if (alloc->allocated_tensors[i] == NULL) {
+            alloc->allocated_tensors[i] = tensor;
+            return;
+        }
+    }
+    GGML_ASSERT(!"out of allocated_tensors");
+}
+static void remove_allocated_tensor(struct ggml_allocator * alloc, struct ggml_tensor * tensor) {
+    for (int i = 0; i < 1024; i++) {
+        if (alloc->allocated_tensors[i] == tensor ||
+            (alloc->allocated_tensors[i] != NULL && alloc->allocated_tensors[i]->data == tensor->data)) {
+            alloc->allocated_tensors[i] = NULL;
+            return;
+        }
+    }
+    printf("tried to free tensor %s not found\n", tensor->name);
+    GGML_ASSERT(!"tensor not found");
+}
+#endif
+
+
+static size_t ggml_allocator_get_alloc_size(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
+    return ggml_nbytes(tensor);
+
+    UNUSED(alloc);
+}
+
+void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
+    size_t size = ggml_allocator_get_alloc_size(alloc, tensor);
+    size = aligned_offset(NULL, size, alloc->alignment);
+
+    AT_PRINTF("%s: allocating %s (%zu bytes) - ", __func__, tensor->name, size);
+
+    size_t max_avail = 0;
+
+    // find the best fitting free block
+    int best_fit_block = -1;
+    size_t best_fit_size = SIZE_MAX;
+    for (int i = 0; i < alloc->n_free_blocks; i++) {
+        struct free_block * block = &alloc->free_blocks[i];
+        max_avail = MAX(max_avail, block->size);
+        if (block->size >= size && block->size <= best_fit_size) {
+            best_fit_block = i;
+            best_fit_size = block->size;
+        }
+    }
+
+    AT_PRINTF("block %d\n", best_fit_block);
+
+    if (best_fit_block == -1) {
+        fprintf(stderr, "%s: not enough space in the buffer (needed %zu, largest block available %zu)\n",
+                __func__, size, max_avail);
+        GGML_ASSERT(!"not enough space in the buffer");
+        return;
+    }
+    struct free_block * block = &alloc->free_blocks[best_fit_block];
+    void * addr = block->addr;
+    block->addr = (char*)block->addr + size;
+    block->size -= size;
+    if (block->size == 0) {
+        // remove block if empty
+        alloc->n_free_blocks--;
+        for (int j = best_fit_block; j < alloc->n_free_blocks; j++) {
+            alloc->free_blocks[j] = alloc->free_blocks[j+1];
+        }
+    }
+
+    tensor->data = addr;
+
+#ifdef GGML_ALLOCATOR_DEBUG
+    add_allocated_tensor(alloc, tensor);
+    size_t cur_max = (char*)addr - (char*)alloc->data + size;
+    if (cur_max > alloc->max_size) {
+        printf("max_size = %.2f MB: tensors: ", cur_max / 1024.0 / 1024.0);
+        for (int i = 0; i < 1024; i++) {
+            if (alloc->allocated_tensors[i]) {
+                printf("%s (%.2f MB) ", alloc->allocated_tensors[i]->name, ggml_nbytes(alloc->allocated_tensors[i]) / 1024.0 / 1024.0);
+            }
+        }
+        printf("\n");
+    }
+#endif
+
+    alloc->max_size = MAX(alloc->max_size, (char*)addr - (char*)alloc->data + size);
+}
+
+// this is a very naive implementation, but for our case the number of free blocks should be very small
+static void ggml_allocator_free_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
+    void * ptr = tensor->data;
+
+    if (ptr < alloc->data || (char*)ptr >= (char*)alloc->data + alloc->max_size) {
+        // the tensor was not allocated in this buffer
+        // this can happen because the graph allocator will try to free weights and other tensors from different buffers
+        // the easiest way to deal with this is just to ignore it
+        return;
+    }
+
+    size_t size = ggml_allocator_get_alloc_size(alloc, tensor);
+    size = aligned_offset(NULL, size, alloc->alignment);
+    AT_PRINTF("%s: freeing %s (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, size, alloc->n_free_blocks);
+
+#ifdef GGML_ALLOCATOR_DEBUG
+    remove_allocated_tensor(alloc, tensor);
+#endif
+
+    // see if we can merge with an existing block
+    for (int i = 0; i < alloc->n_free_blocks; i++) {
+        struct free_block * block = &alloc->free_blocks[i];
+        // check if ptr is at the end of the block
+        if ((char*)block->addr + block->size == ptr) {
+            block->size += size;
+            // check if we can merge with the next block
+            if (i < alloc->n_free_blocks - 1 && (char*)block->addr + block->size == alloc->free_blocks[i+1].addr) {
+                block->size += alloc->free_blocks[i+1].size;
+                alloc->n_free_blocks--;
+                for (int j = i+1; j < alloc->n_free_blocks; j++) {
+                    alloc->free_blocks[j] = alloc->free_blocks[j+1];
+                }
+            }
+            return;
+        }
+        // check if ptr is at the beginning of the block
+        if ((char*)ptr + size == block->addr) {
+            block->addr = ptr;
+            block->size += size;
+            // check if we can merge with the previous block
+            if (i > 0 && (char*)alloc->free_blocks[i-1].addr + alloc->free_blocks[i-1].size == block->addr) {
+                alloc->free_blocks[i-1].size += block->size;
+                alloc->n_free_blocks--;
+                for (int j = i; j < alloc->n_free_blocks; j++) {
+                    alloc->free_blocks[j] = alloc->free_blocks[j+1];
+                }
+            }
+            return;
+        }
+    }
+    // otherwise, add a new block
+    GGML_ASSERT(alloc->n_free_blocks < MAX_FREE_BLOCKS && "out of free blocks");
+    // insert the new block in the correct position to keep the array sorted by address (to make merging blocks faster)
+    int insert_pos = 0;
+    while (insert_pos < alloc->n_free_blocks && alloc->free_blocks[insert_pos].addr < ptr) {
+        insert_pos++;
+    }
+    // shift all blocks from insert_pos onward to make room for the new block
+    for (int i = alloc->n_free_blocks; i > insert_pos; i--) {
+        alloc->free_blocks[i] = alloc->free_blocks[i-1];
+    }
+    // insert the new block
+    alloc->free_blocks[insert_pos].addr = ptr;
+    alloc->free_blocks[insert_pos].size = size;
+    alloc->n_free_blocks++;
+}
+
+void ggml_allocr_reset(struct ggml_allocr * alloc) {
+    alloc->n_free_blocks = 1;
+    size_t align_offset = aligned_offset(alloc->data, 0, alloc->alignment);
+    alloc->free_blocks[0].addr = (char *)alloc->data + align_offset;
+    alloc->free_blocks[0].size = alloc->size - align_offset;
+}
+
+struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment) {
+    struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
+
+    *alloc = (struct ggml_allocr){
+        /*.data          = */ data,
+        /*.size          = */ size,
+        /*.alignment     = */ alignment,
+        /*.n_free_blocks = */ 0,
+        /*.free_blocks   = */ {{0}},
+        /*.hash_table    = */ {{0}},
+        /*.max_size      = */ 0,
+        /*.measure       = */ false,
+#ifdef GGML_ALLOCATOR_DEBUG
+        /*.allocated_tensors = */ = {0},
+#endif
+    };
+
+    ggml_allocr_reset(alloc);
+
+    return alloc;
+}
+
+// address and size of the buffer when measuring
+// it needs to be large enough to fit all the tensors, but it cannot overlap with other existing buffers
+static void * const MEASURE_BASE_ADDR = (void *) 0x1000;
+static const size_t MEASURE_MAX_SIZE  = 1ULL<<40; // 1 TB
+
+struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) {
+    struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
+
+    *alloc = (struct ggml_allocr){
+        /*.data          = */ MEASURE_BASE_ADDR,
+        /*.size          = */ MEASURE_MAX_SIZE,
+        /*.alignment     = */ alignment,
+        /*.n_free_blocks = */ 0,
+        /*.free_blocks   = */ {{0}},
+        /*.hash_table    = */ {{0}},
+        /*.max_size      = */ 0,
+        /*.measure       = */ true,
+#ifdef GGML_ALLOCATOR_DEBUG
+        /*.allocated_tensors = */ = {0},
+#endif
+    };
+
+    ggml_allocr_reset(alloc);
+
+    return alloc;
+}
+
+void ggml_allocr_free(struct ggml_allocr * alloc) {
+    free(alloc);
+}
+
+bool ggml_allocr_is_measure(struct ggml_allocr * alloc) {
+    return alloc->measure;
+}
+
+//////////// compute graph allocator
+
+static bool ggml_is_view(struct ggml_tensor * t) {
+    return t->op == GGML_OP_RESHAPE || t->op == GGML_OP_VIEW || t->op == GGML_OP_TRANSPOSE ||
+           t->op == GGML_OP_PERMUTE || t->op == GGML_OP_CPY;
+}
+
+static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
+    if (a->type != b->type) {
+        return false;
+    }
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        if (a->ne[i] != b->ne[i]) {
+            return false;
+        }
+        if (a->nb[i] != b->nb[i]) {
+            return false;
+        }
+    }
+    return true;
+}
+
+static struct ggml_tensor * get_view_parent(struct ggml_tensor * t) {
+    switch (t->op) {
+        case GGML_OP_PERMUTE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_TRANSPOSE:
+        case GGML_OP_VIEW:
+            return t->src[0];
+        case GGML_OP_CPY:
+            return t->src[1];
+        default:
+            return NULL;
+    }
+}
+
+static struct ggml_tensor * get_view_source(struct ggml_tensor * t) {
+    struct ggml_tensor * parent = t;
+    do {
+        parent = get_view_parent(parent);
+    } while (ggml_is_view(parent));
+    return parent;
+}
+
+static bool ggml_op_can_inplace(enum ggml_op op) {
+    switch (op) {
+        case GGML_OP_SCALE:
+        case GGML_OP_DIAG_MASK_ZERO:
+        case GGML_OP_DIAG_MASK_INF:
+        case GGML_OP_ADD:
+        case GGML_OP_ADD1:
+        case GGML_OP_ACC:
+        case GGML_OP_SUB:
+        case GGML_OP_MUL:
+        case GGML_OP_DIV:
+        case GGML_OP_SQR:
+        case GGML_OP_SQRT:
+        case GGML_OP_LOG:
+        case GGML_OP_UNARY:
+        case GGML_OP_ROPE:
+        case GGML_OP_RMS_NORM:
+        case GGML_OP_SET:
+        case GGML_OP_SOFT_MAX:
+        case GGML_OP_CONT:
+            return true;
+
+        default:
+            return false;
+    }
+}
+
+static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node) {
+    struct hash_node * ht = alloc->hash_table;
+    if (node->data == NULL) {
+        if (ggml_is_view(node)) {
+            size_t offset;
+            switch(node->op) {
+                case GGML_OP_VIEW:
+                    memcpy(&offset, node->op_params, sizeof(size_t));
+                    node->data = (char *) node->src[0]->data + offset;
+                    break;
+                case GGML_OP_PERMUTE:
+                case GGML_OP_RESHAPE:
+                case GGML_OP_TRANSPOSE:
+                    node->data = node->src[0]->data;
+                    break;
+                case GGML_OP_CPY:
+                    node->data = node->src[1]->data;
+                    break;
+                default:
+                    GGML_ASSERT(!"unknown view op");
+                    break;
+            }
+        } else {
+            // see if we can reuse a parent's buffer (inplace)
+            if (ggml_op_can_inplace(node->op)) {
+                for (int i = 0; i < GGML_MAX_SRC; i++) {
+                    struct ggml_tensor * parent = node->src[i];
+                    if (parent == NULL) {
+                        break;
+                    }
+
+                    // if the node's data is external, then we cannot re-use it
+                    if ((char *) parent->data < (char *) alloc->data ||
+                        (char *) parent->data >= ((char *) alloc->data + alloc->size)) {
+                        AT_PRINTF("not reusing parent %s for %s as %p is external\n", parent->name, node->name, parent->data);
+                        continue;
+                    }
+
+                    struct hash_node * p_hn = hash_get(ht, parent);
+                    if (parent->data != NULL && p_hn->n_children == 1 && p_hn->n_views == 0 && ggml_are_same_layout(node, parent)) {
+                        if (ggml_is_view(parent)) {
+                            struct ggml_tensor * view_src = get_view_source(parent);
+                            struct hash_node * view_src_hn = hash_get(ht, view_src);
+                            if (view_src_hn->n_views == 1 && view_src_hn->n_children == 0 && view_src->data == parent->data) {
+                                // TODO: the offset of the view parent must be kept to ensure that the op doesn't overwrite
+                                // the parent's data that it will need later (same layout requirement). the problem is that then
+                                // we cannot free the tensor because the original address of the allocation is lost.
+                                // adding a view_src pointer to the tensor would solve this and simplify the code dealing with views
+                                // for now, we only reuse the parent's data if the offset is zero (view_src->data == parent->data)
+                                AT_PRINTF("reusing view parent %s (%s) for %s\n", parent->name, view_src->name, node->name);
+                                node->data = parent->data;
+                                return;
+                            }
+                        }
+                        else {
+                            AT_PRINTF("reusing parent %s for %s\n", parent->name, node->name);
+                            node->data = parent->data;
+                        }
+                        return;
+                    }
+                }
+            }
+            ggml_allocr_alloc(alloc, node);
+        }
+    }
+}
+
+static size_t ggml_allocator_alloc_graph_tensors_n(
+    struct ggml_allocr * alloc,
+    struct ggml_cgraph ** graphs, int n_graphs,
+    struct ggml_tensor *** inputs, struct ggml_tensor *** outputs) {
+
+    // reset hash table
+    struct hash_node * ht = alloc->hash_table;
+    memset(ht, 0, sizeof(struct hash_node) * GGML_GRAPH_HASHTABLE_SIZE);
+
+    // count number of children and views
+    for (int g = 0; g < n_graphs; g++) {
+        struct ggml_cgraph * gf = graphs[g];
+        for (int i = 0; i < gf->n_nodes; i++) {
+            struct ggml_tensor * node = gf->nodes[i];
+
+            if (ggml_is_view(node)) {
+                struct ggml_tensor * view_src = get_view_source(node);
+                hash_get(ht, view_src)->n_views += 1;
+            }
+
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * parent = node->src[j];
+                if (parent == NULL) {
+                    break;
+                }
+                hash_get(ht, parent)->n_children += 1;
+            }
+        }
+    }
+
+    // allocate tensors
+    for (int g = 0; g < n_graphs; g++) {
+        struct ggml_cgraph * gf = graphs[g];
+        AT_PRINTF("####### graph %d/%d\n", g, n_graphs);
+        // graph inputs are allocated first to ensure that they are not overwritten by each other
+        if (inputs != NULL && inputs[g] != NULL) {
+            for (int i = 0; inputs[g][i] != NULL; i++) {
+                struct ggml_tensor * input = inputs[g][i];
+                AT_PRINTF("input: %s\n", input->name);
+                allocate_node(alloc, input);
+            }
+        }
+        for (int i = 0; i < gf->n_nodes; i++) {
+            struct ggml_tensor * node = gf->nodes[i];
+
+            // allocate parents (leafs)
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * parent = node->src[j];
+                if (parent == NULL) {
+                    break;
+                }
+                allocate_node(alloc, parent);
+            }
+
+            // allocate node
+            allocate_node(alloc, node);
+
+            AT_PRINTF("exec: %s (%s) <= ", ggml_op_name(node->op), node->name);
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * parent = node->src[j];
+                if (parent == NULL) {
+                    break;
+                }
+                AT_PRINTF("%s", parent->name);
+                if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) {
+                    AT_PRINTF(", ");
+                }
+            }
+            AT_PRINTF("\n");
+
+            // update parents
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * parent = node->src[j];
+                if (parent == NULL) {
+                    break;
+                }
+                struct hash_node * p_hn = hash_get(ht, parent);
+                p_hn->n_children -= 1;
+
+                //AT_PRINTF("parent %s: %d children, %d views\n", parent->name, parent->n_children, parent->n_views);
+
+                if (p_hn->n_children == 0 && p_hn->n_views == 0) {
+                    if (ggml_is_view(parent)) {
+                        struct ggml_tensor * view_src = get_view_source(parent);
+                        struct hash_node * view_src_hn = hash_get(ht, view_src);
+                        view_src_hn->n_views -= 1;
+                        AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src->n_children, view_src->n_views);
+                        if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src->data != node->data) {
+                            ggml_allocator_free_tensor(alloc, view_src);
+                        }
+                    }
+                    else {
+                        if (parent->data != node->data) {
+                            ggml_allocator_free_tensor(alloc, parent);
+                        }
+                    }
+                }
+            }
+            AT_PRINTF("\n");
+        }
+        // free graph outputs here that wouldn't be freed otherwise because they have no children
+        if (outputs != NULL && outputs[g] != NULL) {
+            for (int i = 0; outputs[g][i] != NULL; i++) {
+                struct ggml_tensor * output = outputs[g][i];
+                AT_PRINTF("output: %s\n", output->name);
+                ggml_allocator_free_tensor(alloc, output);
+            }
+        }
+    }
+
+    return alloc->max_size;
+}
+
+size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph) {
+    return ggml_allocator_alloc_graph_tensors_n(alloc, &graph, 1, NULL, NULL);
+}
--- a/llm/ggml-alloc.h
+++ b/llm/ggml-alloc.h
@@ -0,0 +1,48 @@
+/**
+ * llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#pragma once
+
+#include "ggml.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+
+GGML_API struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment);
+GGML_API struct ggml_allocr * ggml_allocr_new_measure(size_t alignment);
+
+GGML_API void   ggml_allocr_free(struct ggml_allocr * alloc);
+GGML_API bool   ggml_allocr_is_measure(struct ggml_allocr * alloc);
+GGML_API void   ggml_allocr_reset(struct ggml_allocr * alloc);
+GGML_API void   ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor);
+GGML_API size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph);
+
+
+#ifdef  __cplusplus
+}
+#endif
--- a/llm/ggml-cuda.cu
+++ b/llm/ggml-cuda.cu
--- a/llm/ggml-cuda.h
+++ b/llm/ggml-cuda.h
@@ -0,0 +1,63 @@
+/**
+ * llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#pragma once
+
+#include "ggml.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+#define GGML_CUDA_MAX_DEVICES       16
+
+void   ggml_init_cublas(void);
+void   ggml_cuda_set_tensor_split(const float * tensor_split);
+
+void   ggml_cuda_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+bool   ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+size_t ggml_cuda_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+void   ggml_cuda_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
+
+// TODO: export these with GGML_API
+void * ggml_cuda_host_malloc(size_t size);
+void   ggml_cuda_host_free(void * ptr);
+
+void   ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor);
+
+void   ggml_cuda_free_data(struct ggml_tensor * tensor);
+void   ggml_cuda_assign_buffers(struct ggml_tensor * tensor);
+void   ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);
+void   ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor);
+void   ggml_cuda_set_main_device(int main_device);
+void   ggml_cuda_set_mul_mat_q(bool mul_mat_q);
+void   ggml_cuda_set_scratch_size(size_t scratch_size);
+void   ggml_cuda_free_scratch(void);
+bool   ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
+
+#ifdef  __cplusplus
+}
+#endif
--- a/llm/ggml-metal.h
+++ b/llm/ggml-metal.h
@@ -0,0 +1,106 @@
+//go:build darwin
+
+/**
+ * llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+// An interface allowing to compute ggml_cgraph with Metal
+//
+// This is a fully functional interface that extends ggml with GPU support for Apple devices.
+// A similar interface can be created for other GPU backends (e.g. Vulkan, CUDA, OpenCL, etc.)
+//
+// How it works?
+//
+// As long as your program can create and evaluate a ggml_cgraph on the CPU, you can use this
+// interface to evaluate the same graph on the GPU. Instead of using ggml_graph_compute(), you
+// use ggml_metal_graph_compute() (or ggml_vulkan_graph_compute(), etc.)
+//
+// You only need to make sure that all memory buffers that you used during the graph creation
+// are mapped to the device memory with the ggml_metal_add_buffer() function. This mapping is
+// used during the graph evaluation to determine the arguments of the compute kernels.
+//
+// Synchronization between device and host memory (for example for input and output tensors)
+// is done with the ggml_metal_set_tensor() and ggml_metal_get_tensor() functions.
+//
+
+#pragma once
+
+#include <stddef.h>
+#include <stdbool.h>
+
+// max memory buffers that can be mapped to the device
+#define GGML_METAL_MAX_BUFFERS 16
+
+struct ggml_tensor;
+struct ggml_cgraph;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct ggml_metal_context;
+
+// number of command buffers to use
+struct ggml_metal_context * ggml_metal_init(int n_cb);
+void ggml_metal_free(struct ggml_metal_context * ctx);
+
+// set the number of command buffers to use
+void ggml_metal_set_n_cb(struct ggml_metal_context * ctx, int n_cb);
+
+// creates a mapping between a host memory buffer and a device memory buffer
+// - make sure to map all buffers used in the graph before calling ggml_metal_graph_compute
+// - the mapping is used during computation to determine the arguments of the compute kernels
+// - you don't need to keep the host memory buffer allocated as it is never accessed by Metal
+// - max_size specifies the maximum size of a tensor and is used to create shared views such
+//   that it is guaranteed that the tensor will fit in at least one of the views
+//
+bool ggml_metal_add_buffer(
+        struct ggml_metal_context * ctx,
+                       const char * name,
+                             void * data,
+                           size_t   size,
+                           size_t   max_size);
+
+// set data from host memory into the device
+void ggml_metal_set_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
+
+// get data from the device into host memory
+void ggml_metal_get_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
+
+// try to find operations that can be run concurrently in the graph
+// you should run it again if the topology of your graph changes
+void ggml_metal_graph_find_concurrency(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);
+
+// if the graph has been optimized for concurrently dispatch
+bool ggml_metal_if_optimized(struct ggml_metal_context * ctx);
+
+// same as ggml_graph_compute but uses Metal
+// creates gf->n_threads command buffers in parallel
+void ggml_metal_graph_compute(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);
+
+#ifdef __cplusplus
+}
+#endif
+
--- a/llm/ggml-metal.m
+++ b/llm/ggml-metal.m
--- a/llm/ggml-metal.metal
+++ b/llm/ggml-metal.metal
--- a/llm/ggml-mpi.c
+++ b/llm/ggml-mpi.c
@@ -0,0 +1,244 @@
+//go:build mpi
+
+/**
+ * llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "ggml-mpi.h"
+
+#include "ggml.h"
+
+#include <mpi.h>
+
+#include <stdio.h>
+#include <stdlib.h>
+
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+
+#define UNUSED GGML_UNUSED
+
+struct ggml_mpi_context {
+    int rank;
+    int size;
+};
+
+void ggml_mpi_backend_init(void) {
+    MPI_Init(NULL, NULL);
+}
+
+void ggml_mpi_backend_free(void) {
+    MPI_Finalize();
+}
+
+struct ggml_mpi_context * ggml_mpi_init(void) {
+    struct ggml_mpi_context * ctx = calloc(1, sizeof(struct ggml_mpi_context));
+
+    MPI_Comm_rank(MPI_COMM_WORLD, &ctx->rank);
+    MPI_Comm_size(MPI_COMM_WORLD, &ctx->size);
+
+    return ctx;
+}
+
+void ggml_mpi_free(struct ggml_mpi_context * ctx) {
+    free(ctx);
+}
+
+int ggml_mpi_rank(struct ggml_mpi_context * ctx) {
+    return ctx->rank;
+}
+
+void ggml_mpi_eval_init(
+        struct ggml_mpi_context * ctx_mpi,
+                            int * n_tokens,
+                            int * n_past,
+                            int * n_threads) {
+    UNUSED(ctx_mpi);
+
+    // synchronize the worker node parameters with the root node
+    MPI_Barrier(MPI_COMM_WORLD);
+
+    MPI_Bcast(n_tokens,  1, MPI_INT, 0, MPI_COMM_WORLD);
+    MPI_Bcast(n_past,    1, MPI_INT, 0, MPI_COMM_WORLD);
+    MPI_Bcast(n_threads, 1, MPI_INT, 0, MPI_COMM_WORLD);
+}
+
+static int ggml_graph_get_node_idx(struct ggml_cgraph * gf, const char * name) {
+    struct ggml_tensor * t = ggml_graph_get_tensor(gf, name);
+    if (t == NULL) {
+        fprintf(stderr, "%s: tensor %s not found\n", __func__, name);
+        return -1;
+    }
+
+    for (int i = 0; i < gf->n_nodes; i++) {
+        if (gf->nodes[i] == t) {
+            return i;
+        }
+    }
+
+    fprintf(stderr, "%s: tensor %s not found in graph (should not happen)\n", __func__, name);
+    return -1;
+}
+
+static void ggml_mpi_tensor_send(struct ggml_tensor * t, int mpi_rank_dst) {
+    MPI_Datatype mpi_type;
+
+    switch (t->type) {
+        case GGML_TYPE_I32: mpi_type = MPI_INT32_T; break;
+        case GGML_TYPE_F32: mpi_type = MPI_FLOAT;   break;
+        default: GGML_ASSERT(false && "not implemented");
+    }
+
+    const int retval = MPI_Send(t->data, ggml_nelements(t), mpi_type, mpi_rank_dst, 0, MPI_COMM_WORLD);
+    GGML_ASSERT(retval == MPI_SUCCESS);
+}
+
+static void ggml_mpi_tensor_recv(struct ggml_tensor * t, int mpi_rank_src) {
+    MPI_Datatype mpi_type;
+
+    switch (t->type) {
+        case GGML_TYPE_I32: mpi_type = MPI_INT32_T; break;
+        case GGML_TYPE_F32: mpi_type = MPI_FLOAT;   break;
+        default: GGML_ASSERT(false && "not implemented");
+    }
+
+    MPI_Status status; UNUSED(status);
+
+    const int retval = MPI_Recv(t->data, ggml_nelements(t), mpi_type, mpi_rank_src, MPI_ANY_TAG, MPI_COMM_WORLD, &status);
+    GGML_ASSERT(retval == MPI_SUCCESS);
+}
+
+// TODO: there are many improvements that can be done to this implementation
+void ggml_mpi_graph_compute_pre(
+        struct ggml_mpi_context * ctx_mpi,
+             struct ggml_cgraph * gf,
+                            int   n_layers) {
+    const int mpi_rank = ctx_mpi->rank;
+    const int mpi_size = ctx_mpi->size;
+
+    struct ggml_tensor * inp_tokens = ggml_graph_get_tensor(gf, "inp_tokens");
+    if (inp_tokens == NULL) {
+        fprintf(stderr, "%s: tensor 'inp_tokens' not found\n", __func__);
+        return;
+    }
+
+    struct ggml_tensor * inp0 = ggml_graph_get_tensor(gf, "layer_inp_0");
+    if (inp0 == NULL) {
+        fprintf(stderr, "%s: tensor 'inp0' not found\n", __func__);
+        return;
+    }
+
+    GGML_ASSERT(inp0 == gf->nodes[0]);
+
+    // distribute the compute graph into slices across the MPI nodes
+    //
+    // the main node (0) processes the last layers + the remainder of the compute graph
+    // and is responsible to pass the input tokens to the first node (1)
+    //
+    // node 1:   [(  0) * n_per_node, (  1) * n_per_node)
+    // node 2:   [(  1) * n_per_node, (  2) * n_per_node)
+    // ...
+    // node n-1: [(n-2) * n_per_node, (n-1) * n_per_node)
+    // node 0:   [(n-1) * n_per_node,            n_nodes)
+    //
+    if (mpi_rank > 0) {
+        if (mpi_rank == 1) {
+            // the first node (1) receives the input tokens from the main node (0)
+            ggml_mpi_tensor_recv(inp_tokens, 0);
+        } else {
+            // recv input data for each node into the "inp0" tensor (i.e. the first node in the compute graph)
+            ggml_mpi_tensor_recv(inp0, mpi_rank - 1);
+        }
+    } else if (mpi_size > 1) {
+        // node 0 sends the input tokens to node 1
+        ggml_mpi_tensor_send(inp_tokens, 1);
+
+        // recv the output data from the last node
+        ggml_mpi_tensor_recv(inp0, mpi_size - 1);
+    }
+
+    {
+        const int n_per_node = (n_layers + (mpi_size - 1)) / mpi_size;
+
+        const int mpi_idx = mpi_rank > 0 ? mpi_rank - 1 : mpi_size - 1;
+
+        const int il0 =               (mpi_idx + 0) * n_per_node;
+        const int il1 = MIN(n_layers, (mpi_idx + 1) * n_per_node);
+
+        char name_l0[GGML_MAX_NAME];
+        char name_l1[GGML_MAX_NAME];
+
+        snprintf(name_l0, sizeof(name_l0), "layer_inp_%d", il0);
+        snprintf(name_l1, sizeof(name_l1), "layer_inp_%d", il1);
+
+        const int idx_l0 =                ggml_graph_get_node_idx(gf, name_l0);
+        const int idx_l1 = mpi_rank > 0 ? ggml_graph_get_node_idx(gf, name_l1) + 1 : gf->n_nodes;
+
+        if (idx_l0 < 0 || idx_l1 < 0) {
+            fprintf(stderr, "%s: layer input nodes not found\n", __func__);
+            return;
+        }
+
+        // attach the input data to all nodes that need it
+        // TODO: not great - should be able to do this without modifying the compute graph (see next TODO below)
+        for (int i = idx_l0; i < idx_l1; i++) {
+            if (gf->nodes[i]->src[0] == gf->nodes[idx_l0]) {
+                gf->nodes[i]->src[0] =  inp0;
+            }
+            if (gf->nodes[i]->src[1] == gf->nodes[idx_l0]) {
+                gf->nodes[i]->src[1] =  inp0;
+            }
+        }
+
+        // TODO: instead of rearranging the nodes, we should be able to execute a subset of the compute graph
+        for (int i = 1; i < idx_l1 - idx_l0; i++) {
+            gf->nodes[i] = gf->nodes[idx_l0 + i];
+            gf->grads[i] = gf->grads[idx_l0 + i];
+        }
+
+        // the first node performs the "get_rows" operation, the rest of the nodes get the data from the previous node
+        if (mpi_idx != 0) {
+            gf->nodes[0]->op = GGML_OP_NONE;
+        }
+
+        gf->n_nodes = idx_l1 - idx_l0;
+
+        //fprintf(stderr, "%s: node %d: processing %d nodes [%d, %d)\n", __func__, mpi_rank, gf->n_nodes, il0, il1);
+    }
+}
+
+void ggml_mpi_graph_compute_post(
+        struct ggml_mpi_context * ctx_mpi,
+             struct ggml_cgraph * gf,
+                            int   n_layers) {
+    UNUSED(n_layers);
+
+    const int mpi_rank = ctx_mpi->rank;
+    const int mpi_size = ctx_mpi->size;
+
+    // send the output data to the next node
+    if (mpi_rank > 0) {
+        ggml_mpi_tensor_send(gf->nodes[gf->n_nodes - 1], (mpi_rank + 1) % mpi_size);
+    }
+}
--- a/llm/ggml-mpi.h
+++ b/llm/ggml-mpi.h
@@ -0,0 +1,67 @@
+//go:build mpi
+
+/**
+ * llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#pragma once
+
+struct ggml_context;
+struct ggml_tensor;
+struct ggml_cgraph;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct ggml_mpi_context;
+
+void ggml_mpi_backend_init(void);
+void ggml_mpi_backend_free(void);
+
+struct ggml_mpi_context * ggml_mpi_init(void);
+void ggml_mpi_free(struct ggml_mpi_context * ctx);
+
+int ggml_mpi_rank(struct ggml_mpi_context * ctx);
+
+void ggml_mpi_eval_init(
+        struct ggml_mpi_context * ctx_mpi,
+                            int * n_tokens,
+                            int * n_past,
+                            int * n_threads);
+
+void ggml_mpi_graph_compute_pre(
+        struct ggml_mpi_context * ctx_mpi,
+             struct ggml_cgraph * gf,
+                            int   n_layers);
+
+void ggml_mpi_graph_compute_post(
+        struct ggml_mpi_context * ctx_mpi,
+             struct ggml_cgraph * gf,
+                            int   n_layers);
+
+#ifdef __cplusplus
+}
+#endif
--- a/llm/ggml-opencl.cpp
+++ b/llm/ggml-opencl.cpp
--- a/llm/ggml-opencl.h
+++ b/llm/ggml-opencl.h
@@ -0,0 +1,53 @@
+//go:build opencl
+
+/**
+ * llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#pragma once
+
+#include "ggml.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+void ggml_cl_init(void);
+
+void   ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+bool   ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+void   ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
+
+void * ggml_cl_host_malloc(size_t size);
+void   ggml_cl_host_free(void * ptr);
+
+void ggml_cl_free_data(const struct ggml_tensor* tensor);
+
+void ggml_cl_transform_tensor(void * data, struct ggml_tensor * tensor);
+
+#ifdef  __cplusplus
+}
+#endif
--- a/llm/ggml.c
+++ b/llm/ggml.c
--- a/llm/ggml.go
+++ b/llm/ggml.go
@@ -3,96 +3,69 @@ package llm
 import (
 	"encoding/binary"
 	"errors"
+	"fmt"
 	"io"
 )

+type ModelFamily string
+
+type ModelType uint32
+
+const (
+	ModelType3B  ModelType = 26
+	ModelType7B  ModelType = 32
+	ModelType13B ModelType = 40
+	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 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() string
-	ModelType() string
-	FileType() string
-	NumLayers() int64
+	ModelFamily() ModelFamily
+	ModelType() ModelType
+	FileType() FileType
 }

 type container interface {
 	Name() string
-	Decode(io.Reader) (model, error)
+	Decode(io.Reader) error
 }

-type containerGGML struct{}
+type containerGGML struct {
+}

 func (c *containerGGML) Name() string {
 	return "ggml"
 }

-func (c *containerGGML) Decode(r io.Reader) (model, error) {
-	return nil, nil
+func (c *containerGGML) Decode(r io.Reader) error {
+	return nil
 }

 type containerGGMF struct {
@@ -103,18 +76,18 @@ func (c *containerGGMF) Name() string {
 	return "ggmf"
 }

-func (c *containerGGMF) Decode(r io.Reader) (model, error) {
+func (c *containerGGMF) Decode(r io.Reader) error {
 	var version uint32
 	binary.Read(r, binary.LittleEndian, &version)

 	switch version {
 	case 1:
 	default:
-		return nil, errors.New("invalid version")
+		return errors.New("invalid version")
 	}

 	c.version = version
-	return nil, nil
+	return nil
 }

 type containerGGJT struct {
@@ -125,22 +98,18 @@ func (c *containerGGJT) Name() string {
 	return "ggjt"
 }

-func (c *containerGGJT) Decode(r io.Reader) (model, error) {
+func (c *containerGGJT) Decode(r io.Reader) error {
 	var version uint32
 	binary.Read(r, binary.LittleEndian, &version)

 	switch version {
 	case 1, 2, 3:
 	default:
-		return nil, errors.New("invalid version")
+		return errors.New("invalid version")
 	}

 	c.version = version
-
-	// different model types may have different layouts for hyperparameters
-	var llama llamaModel
-	binary.Read(r, binary.LittleEndian, &llama.hyperparameters)
-	return &llama, nil
+	return nil
 }

 type containerLORA struct {
@@ -151,34 +120,32 @@ func (c *containerLORA) Name() string {
 	return "ggla"
 }

-func (c *containerLORA) Decode(r io.Reader) (model, error) {
+func (c *containerLORA) Decode(r io.Reader) error {
 	var version uint32
 	binary.Read(r, binary.LittleEndian, &version)

 	switch version {
 	case 1:
 	default:
-		return nil, errors.New("invalid version")
+		return errors.New("invalid version")
 	}

 	c.version = version
-	return nil, nil
+	return 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) (*GGML, error) {
+func DecodeGGML(r io.ReadSeeker, hint ModelFamily) (*GGML, error) {
 	var ggml GGML
 	binary.Read(r, binary.LittleEndian, &ggml.magic)

@@ -191,18 +158,24 @@ func DecodeGGML(r io.ReadSeeker) (*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")
 	}

-	model, err := ggml.Decode(r)
-	if err != nil {
+	if err := ggml.Decode(r); err != nil {
 		return nil, err
 	}

-	ggml.model = model
+	// 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)
+	}

 	// final model type
 	return &ggml, nil
--- a/llm/ggml.h
+++ b/llm/ggml.h
--- a/llm/gguf.go
+++ b/llm/gguf.go
@@ -1,383 +0,0 @@
-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
-}
--- a/llm/k_quants.c
+++ b/llm/k_quants.c
--- a/llm/k_quants.h
+++ b/llm/k_quants.h
@@ -0,0 +1,191 @@
+/**
+ * llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#pragma once
+
+#include "ggml.h"
+
+#include <stdint.h>
+#include <assert.h>
+#include <stddef.h>
+
+// Super-block size
+#ifdef GGML_QKK_64
+#define QK_K 64
+#define K_SCALE_SIZE 4
+#else
+#define QK_K 256
+#define K_SCALE_SIZE 12
+#endif
+
+#ifndef static_assert
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
+#define static_assert(cond, msg) _Static_assert(cond, msg)
+#else
+#define static_assert(cond, msg) struct global_scope_noop_trick
+#endif
+#endif
+
+//
+// Super-block quantization structures
+//
+
+// 2-bit quantization
+// weight is represented as x = a * q + b
+// 16 blocks of 16 elemenets each
+// Effectively 2.5625 bits per weight
+typedef struct {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    ggml_fp16_t d;           // super-block scale for quantized scales
+    ggml_fp16_t dmin;        // super-block scale for quantized mins
+} block_q2_K;
+static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
+
+// 3-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elemenets each
+// Effectively 3.4375 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+    uint8_t scales[2];
+    ggml_fp16_t d;             // super-block scale
+} block_q3_K;
+static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding");
+#else
+typedef struct {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+    uint8_t scales[12];        // scales, quantized with 6 bits
+    ggml_fp16_t d;             // super-block scale
+} block_q3_K;
+static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
+#endif
+
+// 4-bit quantization
+// 16 blocks of 32 elements each
+// weight is represented as x = a * q + b
+// Effectively 4.5 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    ggml_fp16_t d[2];          // super-block scales/mins
+    uint8_t scales[2];         // 4-bit block scales/mins
+    uint8_t qs[QK_K/2];        // 4--bit quants
+} block_q4_K;
+static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + QK_K/2 + 2, "wrong q4_K block size/padding");
+#else
+typedef struct {
+    ggml_fp16_t d;             // super-block scale for quantized scales
+    ggml_fp16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+} block_q4_K;
+static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding");
+#endif
+
+// 5-bit quantization
+// 16 blocks of 32 elements each
+// weight is represented as x = a * q + b
+// Effectively 5.5 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    ggml_fp16_t d;               // super-block scale
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+} block_q5_K;
+static_assert(sizeof(block_q5_K) == sizeof(ggml_fp16_t) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
+#else
+typedef struct {
+    ggml_fp16_t d;               // super-block scale for quantized scales
+    ggml_fp16_t dmin;            // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+} block_q5_K;
+static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
+#endif
+
+// 6-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elemenets each
+// Effectively 6.5625 bits per weight
+typedef struct {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    ggml_fp16_t d;           // super-block scale
+} block_q6_K;
+static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + QK_K / 16 + 3*QK_K/4, "wrong q6_K block size/padding");
+
+// This is only used for intermediate quantization and dot products
+typedef struct {
+    float   d;              // delta
+    int8_t  qs[QK_K];       // quants
+    int16_t bsums[QK_K/16]; // sum of quants in groups of 16
+} block_q8_K;
+static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding");
+
+
+// Quantization
+void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict y, int k);
+void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict y, int k);
+void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict y, int k);
+void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict y, int k);
+void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int k);
+void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k);
+
+void quantize_row_q2_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q3_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q4_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q5_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q6_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q8_K(const float * restrict x, void * restrict y, int k);
+
+// Dequantization
+void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int k);
+void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k);
+void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int k);
+void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int k);
+void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int k);
+void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k);
+
+// Dot product
+void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+
+// Quantization with histogram collection
+size_t ggml_quantize_q2_K(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q3_K(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q4_K(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q5_K(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist);
+
--- a/llm/llama-util.h
+++ b/llm/llama-util.h
@@ -0,0 +1,579 @@
+/**
+ * llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+// Internal header to be included only by llama.cpp.
+// Contains wrappers around OS interfaces.
+
+#ifndef LLAMA_UTIL_H
+#define LLAMA_UTIL_H
+
+#include <cstdio>
+#include <cstdint>
+#include <cerrno>
+#include <cstring>
+#include <cstdarg>
+#include <cstdlib>
+#include <climits>
+
+#include <string>
+#include <vector>
+#include <stdexcept>
+
+#ifdef __has_include
+    #if __has_include(<unistd.h>)
+        #include <unistd.h>
+        #if defined(_POSIX_MAPPED_FILES)
+            #include <sys/mman.h>
+        #endif
+        #if defined(_POSIX_MEMLOCK_RANGE)
+            #include <sys/resource.h>
+        #endif
+    #endif
+#endif
+
+#if defined(_WIN32)
+    #define WIN32_LEAN_AND_MEAN
+    #ifndef NOMINMAX
+        #define NOMINMAX
+    #endif
+    #include <windows.h>
+    #include <io.h>
+    #include <stdio.h> // for _fseeki64
+#endif
+
+#define LLAMA_ASSERT(x) \
+    do { \
+        if (!(x)) { \
+            fprintf(stderr, "LLAMA_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
+            abort(); \
+        } \
+    } while (0)
+
+#ifdef __GNUC__
+#ifdef __MINGW32__
+__attribute__((format(gnu_printf, 1, 2)))
+#else
+__attribute__((format(printf, 1, 2)))
+#endif
+#endif
+static std::string format(const char * fmt, ...) {
+    va_list ap, ap2;
+    va_start(ap, fmt);
+    va_copy(ap2, ap);
+    int size = vsnprintf(NULL, 0, fmt, ap);
+    LLAMA_ASSERT(size >= 0 && size < INT_MAX);
+    std::vector<char> buf(size + 1);
+    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
+    LLAMA_ASSERT(size2 == size);
+    va_end(ap2);
+    va_end(ap);
+    return std::string(buf.data(), size);
+}
+
+struct llama_file {
+    // use FILE * so we don't have to re-open the file to mmap
+    FILE * fp;
+    size_t size;
+
+    llama_file(const char * fname, const char * mode) {
+        fp = std::fopen(fname, mode);
+        if (fp == NULL) {
+            throw std::runtime_error(format("failed to open %s: %s", fname, strerror(errno)));
+        }
+        seek(0, SEEK_END);
+        size = tell();
+        seek(0, SEEK_SET);
+    }
+
+    size_t tell() const {
+#ifdef _WIN32
+        __int64 ret = _ftelli64(fp);
+#else
+        long ret = std::ftell(fp);
+#endif
+        LLAMA_ASSERT(ret != -1); // this really shouldn't fail
+        return (size_t) ret;
+    }
+
+    void seek(size_t offset, int whence) {
+#ifdef _WIN32
+        int ret = _fseeki64(fp, (__int64) offset, whence);
+#else
+        int ret = std::fseek(fp, (long) offset, whence);
+#endif
+        LLAMA_ASSERT(ret == 0); // same
+    }
+
+    void read_raw(void * ptr, size_t len) const {
+        if (len == 0) {
+            return;
+        }
+        errno = 0;
+        std::size_t ret = std::fread(ptr, len, 1, fp);
+        if (ferror(fp)) {
+            throw std::runtime_error(format("read error: %s", strerror(errno)));
+        }
+        if (ret != 1) {
+            throw std::runtime_error(std::string("unexpectedly reached end of file"));
+        }
+    }
+
+    std::uint32_t read_u32() {
+        std::uint32_t ret;
+        read_raw(&ret, sizeof(ret));
+        return ret;
+    }
+
+    std::string read_string(std::uint32_t len) {
+        std::vector<char> chars(len);
+        read_raw(chars.data(), len);
+        return std::string(chars.data(), len);
+    }
+
+    void write_raw(const void * ptr, size_t len) const {
+        if (len == 0) {
+            return;
+        }
+        errno = 0;
+        size_t ret = std::fwrite(ptr, len, 1, fp);
+        if (ret != 1) {
+            throw std::runtime_error(format("write error: %s", strerror(errno)));
+        }
+    }
+
+    void write_u32(std::uint32_t val) {
+        write_raw(&val, sizeof(val));
+    }
+
+    ~llama_file() {
+        if (fp) {
+            std::fclose(fp);
+        }
+    }
+};
+
+// llama_context_data
+struct llama_data_context {
+    virtual void write(const void * src, size_t size) = 0;
+    virtual size_t get_size_written() = 0;
+    virtual ~llama_data_context() = default;
+};
+
+struct llama_data_buffer_context : llama_data_context {
+    uint8_t* ptr;
+    size_t size_written = 0;
+
+    llama_data_buffer_context(uint8_t * p) : ptr(p) {}
+
+    void write(const void * src, size_t size) override {
+        memcpy(ptr, src, size);
+        ptr += size;
+        size_written += size;
+    }
+
+    size_t get_size_written() override {
+        return size_written;
+    }
+};
+
+struct llama_data_file_context : llama_data_context {
+    llama_file* file;
+    size_t size_written = 0;
+
+    llama_data_file_context(llama_file * f) : file(f) {}
+
+    void write(const void * src, size_t size) override {
+        file->write_raw(src, size);
+        size_written += size;
+    }
+
+    size_t get_size_written() override {
+        return size_written;
+    }
+};
+
+#if defined(_WIN32)
+static std::string llama_format_win_err(DWORD err) {
+    LPSTR buf;
+    size_t size = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
+                                 NULL, err, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&buf, 0, NULL);
+    if (!size) {
+        return "FormatMessageA failed";
+    }
+    std::string ret(buf, size);
+    LocalFree(buf);
+    return ret;
+}
+#endif
+
+struct llama_mmap {
+    void * addr;
+    size_t size;
+
+    llama_mmap(const llama_mmap &) = delete;
+
+#ifdef _POSIX_MAPPED_FILES
+    static constexpr bool SUPPORTED = true;
+
+    llama_mmap(struct llama_file * file, size_t prefetch = (size_t) -1 /* -1 = max value */, bool numa = false) {
+        size = file->size;
+        int fd = fileno(file->fp);
+        int flags = MAP_SHARED;
+        // prefetch/readahead impairs performance on NUMA systems
+        if (numa) { prefetch = 0; }
+#ifdef __linux__
+        if (prefetch >= file->size) { flags |= MAP_POPULATE; }
+#endif
+        addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0);
+        if (addr == MAP_FAILED) {
+            throw std::runtime_error(format("mmap failed: %s", strerror(errno)));
+        }
+
+        if (prefetch > 0) {
+            // Advise the kernel to preload the mapped memory
+            if (madvise(addr, std::min(file->size, prefetch), MADV_WILLNEED)) {
+                fprintf(stderr, "warning: madvise(.., MADV_WILLNEED) failed: %s\n",
+                        strerror(errno));
+            }
+        }
+        if (numa) {
+            // advise the kernel not to use readahead
+            // (because the next page might not belong on the same node)
+            if (madvise(addr, file->size, MADV_RANDOM)) {
+                fprintf(stderr, "warning: madvise(.., MADV_RANDOM) failed: %s\n",
+                        strerror(errno));
+            }
+        }
+    }
+
+    ~llama_mmap() {
+        munmap(addr, size);
+    }
+#elif defined(_WIN32)
+    static constexpr bool SUPPORTED = true;
+
+    llama_mmap(struct llama_file * file, bool prefetch = true, bool numa = false) {
+        (void) numa;
+
+        size = file->size;
+
+        HANDLE hFile = (HANDLE) _get_osfhandle(_fileno(file->fp));
+
+        HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
+        DWORD error = GetLastError();
+
+        if (hMapping == NULL) {
+            throw std::runtime_error(format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str()));
+        }
+
+        addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0);
+        error = GetLastError();
+        CloseHandle(hMapping);
+
+        if (addr == NULL) {
+            throw std::runtime_error(format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str()));
+        }
+
+        if (prefetch) {
+            // The PrefetchVirtualMemory API is only present on Windows 8 and above, so we
+            // will dynamically load it using GetProcAddress.
+            BOOL (WINAPI *pPrefetchVirtualMemory) (HANDLE, ULONG_PTR, PWIN32_MEMORY_RANGE_ENTRY, ULONG);
+            HMODULE hKernel32;
+
+            // This call is guaranteed to succeed.
+            hKernel32 = GetModuleHandleW(L"kernel32.dll");
+
+            // This call may fail if on a pre-Win8 system.
+            pPrefetchVirtualMemory = reinterpret_cast<decltype(pPrefetchVirtualMemory)> (GetProcAddress(hKernel32, "PrefetchVirtualMemory"));
+
+            if (pPrefetchVirtualMemory) {
+                // Advise the kernel to preload the mapped memory.
+                WIN32_MEMORY_RANGE_ENTRY range;
+                range.VirtualAddress = addr;
+                range.NumberOfBytes = (SIZE_T)size;
+                if (!pPrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
+                    fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n",
+                            llama_format_win_err(GetLastError()).c_str());
+                }
+            }
+        }
+    }
+
+    ~llama_mmap() {
+        if (!UnmapViewOfFile(addr)) {
+            fprintf(stderr, "warning: UnmapViewOfFile failed: %s\n",
+                    llama_format_win_err(GetLastError()).c_str());
+        }
+    }
+#else
+    static constexpr bool SUPPORTED = false;
+
+    llama_mmap(struct llama_file *, bool prefetch = true, bool numa = false) {
+        (void) prefetch;
+        (void) numa;
+
+        throw std::runtime_error(std::string("mmap not supported"));
+    }
+#endif
+};
+
+// Represents some region of memory being locked using mlock or VirtualLock;
+// will automatically unlock on destruction.
+struct llama_mlock {
+    void * addr = NULL;
+    size_t size = 0;
+    bool failed_already = false;
+
+    llama_mlock() {}
+    llama_mlock(const llama_mlock &) = delete;
+
+    ~llama_mlock() {
+        if (size) {
+            raw_unlock(addr, size);
+        }
+    }
+
+    void init(void * ptr) {
+        LLAMA_ASSERT(addr == NULL && size == 0);
+        addr = ptr;
+    }
+
+    void grow_to(size_t target_size) {
+        LLAMA_ASSERT(addr);
+        if (failed_already) {
+            return;
+        }
+        size_t granularity = lock_granularity();
+        target_size = (target_size + granularity - 1) & ~(granularity - 1);
+        if (target_size > size) {
+            if (raw_lock((uint8_t *) addr + size, target_size - size)) {
+                size = target_size;
+            } else {
+                failed_already = true;
+            }
+        }
+    }
+
+#ifdef _POSIX_MEMLOCK_RANGE
+    static constexpr bool SUPPORTED = true;
+
+    size_t lock_granularity() {
+        return (size_t) sysconf(_SC_PAGESIZE);
+    }
+
+    #ifdef __APPLE__
+        #define MLOCK_SUGGESTION \
+            "Try increasing the sysctl values 'vm.user_wire_limit' and 'vm.global_user_wire_limit' and/or " \
+            "decreasing 'vm.global_no_user_wire_amount'.  Also try increasing RLIMIT_MLOCK (ulimit -l).\n"
+    #else
+        #define MLOCK_SUGGESTION \
+            "Try increasing RLIMIT_MLOCK ('ulimit -l' as root).\n"
+    #endif
+
+    bool raw_lock(const void * addr, size_t size) {
+        if (!mlock(addr, size)) {
+            return true;
+        } else {
+            char* errmsg = std::strerror(errno);
+            bool suggest = (errno == ENOMEM);
+
+            // Check if the resource limit is fine after all
+            struct rlimit lock_limit;
+            if (suggest && getrlimit(RLIMIT_MEMLOCK, &lock_limit))
+                suggest = false;
+            if (suggest && (lock_limit.rlim_max > lock_limit.rlim_cur + size))
+                suggest = false;
+
+            fprintf(stderr, "warning: failed to mlock %zu-byte buffer (after previously locking %zu bytes): %s\n%s",
+                    size, this->size, errmsg, suggest ? MLOCK_SUGGESTION : "");
+            return false;
+        }
+    }
+
+    #undef MLOCK_SUGGESTION
+
+    void raw_unlock(void * addr, size_t size) {
+        if (munlock(addr, size)) {
+            fprintf(stderr, "warning: failed to munlock buffer: %s\n", std::strerror(errno));
+        }
+    }
+#elif defined(_WIN32)
+    static constexpr bool SUPPORTED = true;
+
+    size_t lock_granularity() {
+        SYSTEM_INFO si;
+        GetSystemInfo(&si);
+        return (size_t) si.dwPageSize;
+    }
+
+    bool raw_lock(void * ptr, size_t len) {
+        for (int tries = 1; ; tries++) {
+            if (VirtualLock(ptr, len)) {
+                return true;
+            }
+            if (tries == 2) {
+                fprintf(stderr, "warning: failed to VirtualLock %zu-byte buffer (after previously locking %zu bytes): %s\n",
+                    len, size, llama_format_win_err(GetLastError()).c_str());
+                return false;
+            }
+
+            // It failed but this was only the first try; increase the working
+            // set size and try again.
+            SIZE_T min_ws_size, max_ws_size;
+            if (!GetProcessWorkingSetSize(GetCurrentProcess(), &min_ws_size, &max_ws_size)) {
+                fprintf(stderr, "warning: GetProcessWorkingSetSize failed: %s\n",
+                        llama_format_win_err(GetLastError()).c_str());
+                return false;
+            }
+            // Per MSDN: "The maximum number of pages that a process can lock
+            // is equal to the number of pages in its minimum working set minus
+            // a small overhead."
+            // Hopefully a megabyte is enough overhead:
+            size_t increment = len + 1048576;
+            // The minimum must be <= the maximum, so we need to increase both:
+            min_ws_size += increment;
+            max_ws_size += increment;
+            if (!SetProcessWorkingSetSize(GetCurrentProcess(), min_ws_size, max_ws_size)) {
+                fprintf(stderr, "warning: SetProcessWorkingSetSize failed: %s\n",
+                        llama_format_win_err(GetLastError()).c_str());
+                return false;
+            }
+        }
+    }
+
+    void raw_unlock(void * ptr, size_t len) {
+        if (!VirtualUnlock(ptr, len)) {
+            fprintf(stderr, "warning: failed to VirtualUnlock buffer: %s\n",
+                    llama_format_win_err(GetLastError()).c_str());
+        }
+    }
+#else
+    static constexpr bool SUPPORTED = false;
+
+    size_t lock_granularity() {
+        return (size_t) 65536;
+    }
+
+    bool raw_lock(const void * addr, size_t len) {
+        fprintf(stderr, "warning: mlock not supported on this system\n");
+        return false;
+    }
+
+    void raw_unlock(const void * addr, size_t len) {}
+#endif
+};
+
+// Replacement for std::vector<uint8_t> that doesn't require zero-initialization.
+struct llama_buffer {
+    uint8_t * addr = NULL;
+    size_t size = 0;
+
+    llama_buffer() = default;
+
+    void resize(size_t len) {
+#ifdef GGML_USE_METAL
+        free(addr);
+        int result = posix_memalign((void **) &addr, getpagesize(), len);
+        if (result == 0) {
+            memset(addr, 0, len);
+        }
+        else {
+            addr = NULL;
+        }
+#else
+        delete[] addr;
+        addr = new uint8_t[len];
+#endif
+        size = len;
+    }
+
+    ~llama_buffer() {
+#ifdef GGML_USE_METAL
+        free(addr);
+#else
+        delete[] addr;
+#endif
+        addr = NULL;
+    }
+
+    // disable copy and move
+    llama_buffer(const llama_buffer&) = delete;
+    llama_buffer(llama_buffer&&) = delete;
+    llama_buffer& operator=(const llama_buffer&) = delete;
+    llama_buffer& operator=(llama_buffer&&) = delete;
+};
+
+#ifdef GGML_USE_CUBLAS
+#include "ggml-cuda.h"
+struct llama_ctx_buffer {
+    uint8_t * addr = NULL;
+    bool is_cuda;
+    size_t size = 0;
+
+    llama_ctx_buffer() = default;
+
+    void resize(size_t size) {
+        free();
+
+        addr = (uint8_t *) ggml_cuda_host_malloc(size);
+        if (addr) {
+            is_cuda = true;
+        }
+        else {
+            // fall back to pageable memory
+            addr = new uint8_t[size];
+            is_cuda = false;
+        }
+        this->size = size;
+    }
+
+    void free() {
+        if (addr) {
+            if (is_cuda) {
+                ggml_cuda_host_free(addr);
+            }
+            else {
+                delete[] addr;
+            }
+        }
+        addr = NULL;
+    }
+
+    ~llama_ctx_buffer() {
+        free();
+    }
+
+    // disable copy and move
+    llama_ctx_buffer(const llama_ctx_buffer&) = delete;
+    llama_ctx_buffer(llama_ctx_buffer&&) = delete;
+    llama_ctx_buffer& operator=(const llama_ctx_buffer&) = delete;
+    llama_ctx_buffer& operator=(llama_ctx_buffer&&) = delete;
+};
+#else
+typedef llama_buffer llama_ctx_buffer;
+#endif
+
+#endif
--- a/llm/llama.cpp
+++ b/llm/llama.cpp
--- a/llm/llama.cpp/generate_darwin_amd64.go
+++ b/llm/llama.cpp/generate_darwin_amd64.go
@@ -1,18 +0,0 @@
-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/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
--- a/llm/llama.cpp/generate_darwin_arm64.go
+++ b/llm/llama.cpp/generate_darwin_arm64.go
@@ -1,18 +0,0 @@
-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/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
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Jeffrey Morgan	1178fd2cbb	build with cmake	2023-08-21 18:36:31 -07:00
Jeffrey Morgan	97c15b601a	wip shell	2023-08-21 09:24:22 -07:00