Restore system prompt on requests

update downloads to use retry wrapper

README.md

@@ -29,7 +29,8 @@ curl https://ollama.ai/install.sh | sh
 
 ### Docker
 
-See the official [Docker image](https://hub.docker.com/r/ollama/ollama).
+The official [Ollama Docker image `ollama/ollama`](https://hub.docker.com/r/ollama/ollama)
+is available on Docker Hub.
 
 ## Quickstart
 
@@ -88,7 +89,7 @@ See the [guide](docs/import.md) on importing models for more information.
 
 ### Customize a prompt
 
-Models from the Ollama library can be customized with a prompt. The example
+Models from the Ollama library can be customized with a prompt. For example, to customize the `llama2` model:
 
 ```
 ollama pull llama2
@@ -178,8 +179,7 @@ ollama list
 Install `cmake` and `go`:
 
 ```
-brew install cmake
-brew install go
+brew install cmake go
 ```
 
 Then generate dependencies and build:
@@ -203,9 +203,8 @@ Finally, in a separate shell, run a model:
 
 ## REST API
 
-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:
+Ollama has a REST API for running and managing models.
+For example, to generate text from a model:
 
 ```
 curl -X POST http://localhost:11434/api/generate -d '{
@@ -214,6 +213,8 @@ curl -X POST http://localhost:11434/api/generate -d '{
 }'
 ```
 
+See the [API documentation](./docs/api.md) for all endpoints.
+
 ## Community Integrations
 
 - [LangChain](https://python.langchain.com/docs/integrations/llms/ollama) and [LangChain.js](https://js.langchain.com/docs/modules/model_io/models/llms/integrations/ollama) with [example](https://js.langchain.com/docs/use_cases/question_answering/local_retrieval_qa)
@@ -233,3 +234,4 @@ curl -X POST http://localhost:11434/api/generate -d '{
 - [oterm](https://github.com/ggozad/oterm)
 - [Ellama Emacs client](https://github.com/s-kostyaev/ellama)
 - [OllamaSharp for .NET](https://github.com/awaescher/OllamaSharp)
+- [Minimalistic React UI for Ollama Models](https://github.com/richawo/minimal-llm-ui)

api/client.go

@@ -18,10 +18,6 @@ import (
 	"github.com/jmorganca/ollama/version"
 )
 
-const DefaultHost = "127.0.0.1:11434"
-
-var envHost = os.Getenv("OLLAMA_HOST")
-
 type Client struct {
 	base *url.URL
 	http http.Client
@@ -44,14 +40,24 @@ func checkError(resp *http.Response, body []byte) error {
 }
 
 func ClientFromEnvironment() (*Client, error) {
+	defaultPort := "11434"
+
 	scheme, hostport, ok := strings.Cut(os.Getenv("OLLAMA_HOST"), "://")
-	if !ok {
+	switch {
+	case !ok:
 		scheme, hostport = "http", os.Getenv("OLLAMA_HOST")
+	case scheme == "http":
+		defaultPort = "80"
+	case scheme == "https":
+		defaultPort = "443"
 	}
 
+	// trim trailing slashes
+	hostport = strings.TrimRight(hostport, "/")
+
 	host, port, err := net.SplitHostPort(hostport)
 	if err != nil {
-		host, port = "127.0.0.1", "11434"
+		host, port = "127.0.0.1", defaultPort
 		if ip := net.ParseIP(strings.Trim(hostport, "[]")); ip != nil {
 			host = ip.String()
 		} else if hostport != "" {
@@ -66,7 +72,7 @@ func ClientFromEnvironment() (*Client, error) {
 		},
 	}
 
-	mockRequest, err := http.NewRequest("HEAD", client.base.String(), nil)
+	mockRequest, err := http.NewRequest(http.MethodHead, client.base.String(), nil)
 	if err != nil {
 		return nil, err
 	}

api/client_test.go

@@ -0,0 +1,43 @@
+package api
+
+import "testing"
+
+func TestClientFromEnvironment(t *testing.T) {
+	type testCase struct {
+		value  string
+		expect string
+		err    error
+	}
+
+	testCases := map[string]*testCase{
+		"empty":                      {value: "", expect: "http://127.0.0.1:11434"},
+		"only address":               {value: "1.2.3.4", expect: "http://1.2.3.4:11434"},
+		"only port":                  {value: ":1234", expect: "http://:1234"},
+		"address and port":           {value: "1.2.3.4:1234", expect: "http://1.2.3.4:1234"},
+		"scheme http and address":    {value: "http://1.2.3.4", expect: "http://1.2.3.4:80"},
+		"scheme https and address":   {value: "https://1.2.3.4", expect: "https://1.2.3.4:443"},
+		"scheme, address, and port":  {value: "https://1.2.3.4:1234", expect: "https://1.2.3.4:1234"},
+		"hostname":                   {value: "example.com", expect: "http://example.com:11434"},
+		"hostname and port":          {value: "example.com:1234", expect: "http://example.com:1234"},
+		"scheme http and hostname":   {value: "http://example.com", expect: "http://example.com:80"},
+		"scheme https and hostname":  {value: "https://example.com", expect: "https://example.com:443"},
+		"scheme, hostname, and port": {value: "https://example.com:1234", expect: "https://example.com:1234"},
+		"trailing slash":             {value: "example.com/", expect: "http://example.com:11434"},
+		"trailing slash port":        {value: "example.com:1234/", expect: "http://example.com:1234"},
+	}
+
+	for k, v := range testCases {
+		t.Run(k, func(t *testing.T) {
+			t.Setenv("OLLAMA_HOST", v.value)
+
+			client, err := ClientFromEnvironment()
+			if err != v.err {
+				t.Fatalf("expected %s, got %s", v.err, err)
+			}
+
+			if client.base.String() != v.expect {
+				t.Fatalf("expected %s, got %s", v.expect, client.base.String())
+			}
+		})
+	}
+}

api/types.go

@@ -293,7 +293,7 @@ func DefaultOptions() Options {
 	return Options{
 		// options set on request to runner
 		NumPredict:       -1,
-		NumKeep:          -1,
+		NumKeep:          0,
 		Temperature:      0.8,
 		TopK:             40,
 		TopP:             0.9,

cmd/cmd.go

@@ -11,6 +11,7 @@ import (
 	"io"
 	"log"
 	"net"
+	"net/http"
 	"os"
 	"os/exec"
 	"os/signal"
@@ -22,7 +23,6 @@ import (
 
 	"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"
@@ -30,30 +30,11 @@ import (
 	"github.com/jmorganca/ollama/api"
 	"github.com/jmorganca/ollama/format"
 	"github.com/jmorganca/ollama/progressbar"
+	"github.com/jmorganca/ollama/readline"
 	"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)
@@ -118,19 +99,16 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}
 
-	models, err := client.List(context.Background())
-	if err != nil {
-		return err
-	}
-
-	canonicalModelPath := server.ParseModelPath(args[0])
-	for _, model := range models.Models {
-		if model.Name == canonicalModelPath.GetShortTagname() {
-			return RunGenerate(cmd, args)
+	name := args[0]
+	// check if the model exists on the server
+	_, err = client.Show(context.Background(), &api.ShowRequest{Name: name})
+	var statusError api.StatusError
+	switch {
+	case errors.As(err, &statusError) && statusError.StatusCode == http.StatusNotFound:
+		if err := PullHandler(cmd, args); err != nil {
+			return err
 		}
-	}
-
-	if err := PullHandler(cmd, args); err != nil {
+	case err != nil:
 		return err
 	}
 
@@ -508,38 +486,11 @@ func generate(cmd *cobra.Command, model, prompt string, wordWrap bool) error {
 }
 
 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("/show",
-			readline.PcItem("license"),
-			readline.PcItem("modelfile"),
-			readline.PcItem("parameters"),
-			readline.PcItem("system"),
-			readline.PcItem("template"),
-		),
-		readline.PcItem("/exit"),
-		readline.PcItem("/bye"),
-	)
-
 	usage := func() {
 		fmt.Fprintln(os.Stderr, "Available Commands:")
 		fmt.Fprintln(os.Stderr, "  /set         Set session variables")
@@ -572,20 +523,17 @@ func generateInteractive(cmd *cobra.Command, model string) error {
 		fmt.Fprintln(os.Stderr, "")
 	}
 
-	var painter Painter
-
-	config := readline.Config{
-		Painter:      &painter,
-		Prompt:       ">>> ",
-		HistoryFile:  filepath.Join(home, ".ollama", "history"),
-		AutoComplete: completer,
+	prompt := readline.Prompt{
+		Prompt:         ">>> ",
+		AltPrompt:      "... ",
+		Placeholder:    "Send a message (/? for help)",
+		AltPlaceholder: `Use """ to end multi-line input`,
 	}
 
-	scanner, err := readline.NewEx(&config)
+	scanner, err := readline.New(prompt)
 	if err != nil {
 		return err
 	}
-	defer scanner.Close()
 
 	var wordWrap bool
 	termType := os.Getenv("TERM")
@@ -602,17 +550,20 @@ func generateInteractive(cmd *cobra.Command, model string) error {
 		wordWrap = false
 	}
 
+	fmt.Print(readline.StartBracketedPaste)
+	defer fmt.Printf(readline.EndBracketedPaste)
+
 	var multiLineBuffer string
-	var isMultiLine bool
 
 	for {
 		line, err := scanner.Readline()
 		switch {
 		case errors.Is(err, io.EOF):
+			fmt.Println()
 			return nil
 		case errors.Is(err, readline.ErrInterrupt):
 			if line == "" {
-				fmt.Println("Use Ctrl-D or /bye to exit.")
+				fmt.Println("\nUse Ctrl-D or /bye to exit.")
 			}
 
 			continue
@@ -623,23 +574,19 @@ func generateInteractive(cmd *cobra.Command, model string) error {
 		line = strings.TrimSpace(line)
 
 		switch {
-		case isMultiLine:
+		case scanner.Prompt.UseAlt:
 			if strings.HasSuffix(line, `"""`) {
-				isMultiLine = false
-				painter.IsMultiLine = isMultiLine
+				scanner.Prompt.UseAlt = false
 				multiLineBuffer += strings.TrimSuffix(line, `"""`)
 				line = multiLineBuffer
 				multiLineBuffer = ""
-				scanner.SetPrompt(">>> ")
 			} else {
 				multiLineBuffer += line + " "
 				continue
 			}
 		case strings.HasPrefix(line, `"""`):
-			isMultiLine = true
-			painter.IsMultiLine = isMultiLine
+			scanner.Prompt.UseAlt = true
 			multiLineBuffer = strings.TrimPrefix(line, `"""`) + " "
-			scanner.SetPrompt("... ")
 			continue
 		case strings.HasPrefix(line, "/list"):
 			args := strings.Fields(line)
@@ -666,19 +613,6 @@ func generateInteractive(cmd *cobra.Command, model string) error {
 				case "quiet":
 					cmd.Flags().Set("verbose", "false")
 					fmt.Println("Set 'quiet' mode.")
-				case "mode":
-					if len(args) > 2 {
-						switch args[2] {
-						case "vim":
-							scanner.SetVimMode(true)
-						case "emacs", "default":
-							scanner.SetVimMode(false)
-						default:
-							usage()
-						}
-					} else {
-						usage()
-					}
 				default:
 					fmt.Printf("Unknown command '/set %s'. Type /? for help\n", args[1])
 				}
@@ -795,21 +729,6 @@ 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)
 }
 

docs/api.md

@@ -45,9 +45,11 @@ Advanced parameters (optional):
 - `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
+- `stream`: if `false` the response will be returned as a single response object, rather than a stream of objects
 
-### Request
+### Examples
+
+#### Request
 
 ```shell
 curl -X POST http://localhost:11434/api/generate -d '{
@@ -56,9 +58,9 @@ curl -X POST http://localhost:11434/api/generate -d '{
 }'
 ```
 
-### Response
+#### Response
 
-A stream of JSON objects:
+A stream of JSON objects is returned:
 
 ```json
 {
@@ -102,6 +104,38 @@ To calculate how fast the response is generated in tokens per second (token/s),
 }
 ```
 
+#### Request
+
+```shell
+curl -X POST http://localhost:11434/api/generate -d '{
+  "model": "llama2:7b",
+  "prompt": "Why is the sky blue?",
+  "stream": false
+}'
+```
+
+#### Response
+
+If `stream` is set to `false`, the response will be a single JSON object:
+
+```json
+{
+  "model": "llama2:7b",
+  "created_at": "2023-08-04T19:22:45.499127Z",
+  "response": "The sky is blue because it is the color of the sky.",
+  "context": [1, 2, 3],
+  "done": true,
+  "total_duration": 5589157167,
+  "load_duration": 3013701500,
+  "sample_count": 114,
+  "sample_duration": 81442000,
+  "prompt_eval_count": 46,
+  "prompt_eval_duration": 1160282000,
+  "eval_count": 13,
+  "eval_duration": 1325948000
+}
+```
+
 ## Create a Model
 
 ```shell
@@ -114,9 +148,11 @@ Create a model from a [`Modelfile`](./modelfile.md)
 
 - `name`: name of the model to create
 - `path`: path to the Modelfile
-- `stream`: (optional) if `false` the response will be be returned as a single response object, rather than a stream of objects
+- `stream`: (optional) if `false` the response will be returned as a single response object, rather than a stream of objects
 
-### Request
+### Examples
+
+#### Request
 
 ```shell
 curl -X POST http://localhost:11434/api/create -d '{
@@ -125,7 +161,7 @@ curl -X POST http://localhost:11434/api/create -d '{
 }'
 ```
 
-### Response
+#### Response
 
 A stream of JSON objects. When finished, `status` is `success`.
 
@@ -143,13 +179,17 @@ GET /api/tags
 
 List models that are available locally.
 
-### Request
+### Examples
+
+#### Request
 
 ```shell
 curl http://localhost:11434/api/tags
 ```
 
-### Response
+#### Response
+
+A single JSON object will be returned.
 
 ```json
 {
@@ -180,7 +220,9 @@ Show details about a model including modelfile, template, parameters, license, a
 
 - `name`: name of the model to show
 
-### Request
+### Examples
+
+#### Request
 
 ```shell
 curl http://localhost:11434/api/show -d '{
@@ -188,7 +230,7 @@ curl http://localhost:11434/api/show -d '{
 }'
 ```
 
-### Response
+#### Response
 
 ```json
 {
@@ -207,7 +249,9 @@ POST /api/copy
 
 Copy a model. Creates a model with another name from an existing model.
 
-### Request
+### Examples
+
+#### Request
 
 ```shell
 curl http://localhost:11434/api/copy -d '{
@@ -216,6 +260,10 @@ curl http://localhost:11434/api/copy -d '{
 }'
 ```
 
+#### Response
+
+The only response is a 200 OK if successful.
+
 ## Delete a Model
 
 ```shell
@@ -226,9 +274,11 @@ Delete a model and its data.
 
 ### Parameters
 
-- `model`: model name to delete
+- `name`: model name to delete
 
-### Request
+### Examples
+
+#### Request
 
 ```shell
 curl -X DELETE http://localhost:11434/api/delete -d '{
@@ -236,6 +286,10 @@ curl -X DELETE http://localhost:11434/api/delete -d '{
 }'
 ```
 
+#### Response
+
+If successful, the only response is a 200 OK.
+
 ## Pull a Model
 
 ```shell
@@ -248,9 +302,11 @@ Download a model from the ollama library. Cancelled pulls are resumed from where
 
 - `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
+- `stream`: (optional) if `false` the response will be returned as a single response object, rather than a stream of objects
 
-### Request
+### Examples
+
+#### Request
 
 ```shell
 curl -X POST http://localhost:11434/api/pull -d '{
@@ -258,13 +314,51 @@ curl -X POST http://localhost:11434/api/pull -d '{
 }'
 ```
 
-### Response
+#### Response
+
+If `stream` is not specified, or set to `true`, a stream of JSON objects is returned:
+
+The first object is the manifest:
+
+```json
+{
+  "status": "pulling manifest"
+}
+```
+
+Then there is a series of downloading responses. Until any of the download is completed, the `completed` key may not be included. The number of files to be downloaded depends on the number of layers specified in the manifest.
 
 ```json
 {
   "status": "downloading digestname",
   "digest": "digestname",
-  "total": 2142590208
+  "total": 2142590208,
+  "completed": 241970
+}
+```
+
+After all the files are downloaded, the final responses are:
+
+```json
+{
+    "status": "verifying sha256 digest"
+}
+{
+    "status": "writing manifest"
+}
+{
+    "status": "removing any unused layers"
+}
+{
+    "status": "success"
+}
+```
+
+if `stream` is set to false, then the response is a single JSON object:
+
+```json
+{
+  "status": "success"
 }
 ```
 
@@ -280,9 +374,11 @@ Upload a model to a model library. Requires registering for ollama.ai and adding
 
 - `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
+- `stream`: (optional) if `false` the response will be returned as a single response object, rather than a stream of objects
 
-### Request
+### Examples
+
+#### Request
 
 ```shell
 curl -X POST http://localhost:11434/api/push -d '{
@@ -290,9 +386,9 @@ curl -X POST http://localhost:11434/api/push -d '{
 }'
 ```
 
-### Response
+#### Response
 
-Streaming response that starts with:
+If `stream` is not specified, or set to `true`, a stream of JSON objects is returned:
 
 ```json
 { "status": "retrieving manifest" }
@@ -325,6 +421,12 @@ Finally, when the upload is complete:
 {"status":"success"}
 ```
 
+If `stream` is set to `false`, then the response is a single JSON object:
+
+```json
+{ "status": "success" }
+```
+
 ## Generate Embeddings
 
 ```shell
@@ -342,7 +444,9 @@ Advanced parameters:
 
 - `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
 
-### Request
+### Examples
+
+#### Request
 
 ```shell
 curl -X POST http://localhost:11434/api/embeddings -d '{
@@ -351,7 +455,7 @@ curl -X POST http://localhost:11434/api/embeddings -d '{
 }'
 ```
 
-### Response
+#### Response
 
 ```json
 {

docs/faq.md

@@ -16,19 +16,64 @@ journalctl -u ollama
 
 If you're running `ollama serve` directly, the logs will be printed to the console.
 
-## How can I expose the Ollama server?
+## How can I expose Ollama on my network?
+
+Ollama binds to 127.0.0.1 port 11434 by default. Change the bind address with the `OLLAMA_HOST` environment variable.
+
+On macOS:
 
 ```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:
+On Linux:
+
+Create a `systemd` drop-in directory and set `Environment=OLLAMA_HOST`
+
+```bash
+mkdir -p /etc/systemd/system/ollama.service.d
+echo "[Service]" >>/etc/systemd/system/ollama.service.d/environment.conf
+```
+
+```bash
+echo "Environment=OLLAMA_HOST=0.0.0.0:11434" >>/etc/systemd/system/ollama.service.d/environment.conf
+```
+
+Reload `systemd` and restart Ollama:
+
+```bash
+systemctl daemon-reload
+systemctl restart ollama
+```
+
+## How can I allow additional web origins to access Ollama?
+
+Ollama allows cross origin requests from `127.0.0.1` and `0.0.0.0` by default. Add additional origins with the `OLLAMA_ORIGINS` environment variable:
+
+On macOS:
 
 ```bash
 OLLAMA_ORIGINS=http://192.168.1.1:*,https://example.com ollama serve
 ```
 
+On Linux:
+
+```bash
+echo "Environment=OLLAMA_ORIGINS=http://129.168.1.1:*,https://example.com" >>/etc/systemd/system/ollama.service.d/environment.conf
+```
+
+Reload `systemd` and restart Ollama:
+
+```bash
+systemctl daemon-reload
+systemctl restart ollama
+```
+
 ## 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`
+
+### How can I change where Ollama stores models?
+
+To modify where models are stored, you can use the `OLLAMA_MODELS` environment variable. Note that on Linux this means defining `OLLAMA_MODELS` in a drop-in `/etc/systemd/system/ollama.service.d` service file, reloading systemd, and restarting the ollama service.

docs/import.md

@@ -1,8 +1,43 @@
 # Import a model
 
-This guide walks through importing a PyTorch, Safetensors or GGUF model.
+This guide walks through importing a GGUF, PyTorch or Safetensors model.
 
-## Supported models
+## Importing (GGUF)
+
+### Step 1: Write a `Modelfile`
+
+Start by creating a `Modelfile`. This file is the blueprint for your model, specifying weights, parameters, prompt templates and more.
+
+```
+FROM ./mistral-7b-v0.1.Q4_0.gguf
+```
+
+(Optional) many chat models require a prompt template in order to answer correctly. A default prompt template can be specified with the `TEMPLATE` instruction in the `Modelfile`:
+
+```
+FROM ./q4_0.bin
+TEMPLATE "[INST] {{ .Prompt }} [/INST]"
+```
+
+### Step 2: Create the Ollama model
+
+Finally, create a model from your `Modelfile`:
+
+```
+ollama create example -f Modelfile
+```
+
+### Step 3: Run your model
+
+Next, test the model with `ollama run`:
+
+```
+ollama run example "What is your favourite condiment?"
+```
+
+## Importing (PyTorch & Safetensors)
+
+### Supported models
 
 Ollama supports a set of model architectures, with support for more coming soon:
 
@@ -13,8 +48,6 @@ Ollama supports a set of model architectures, with support for more coming soon:
 
 To view a model's architecture, check the `config.json` file in its HuggingFace repo. You should see an entry under `architectures` (e.g. `LlamaForCausalLM`).
 
-## Importing
-
 ### Step 1: Clone the HuggingFace repository (optional)
 
 If the model is currently hosted in a HuggingFace repository, first clone that repository to download the raw model.
@@ -44,7 +77,7 @@ This will output two files into the directory:
 
 ### Step 3: Write a `Modelfile`
 
-Next, create a `Modelfile` for your model. This file is the blueprint for your model, specifying weights, parameters, prompt templates and more.
+Next, create a `Modelfile` for your model:
 
 ```
 FROM ./q4_0.bin
@@ -65,13 +98,15 @@ Finally, create a model from your `Modelfile`:
 ollama create example -f Modelfile
 ```
 
+### Step 5: Run your model
+
 Next, test the model with `ollama run`:
 
 ```
 ollama run example "What is your favourite condiment?"
 ```
 
-### Step 5: Publish your model (optional – early alpha)
+## Publishing your model (optional – early alpha)
 
 Publishing models is in early alpha. If you'd like to publish your model to share with others, follow these steps:
 
@@ -150,7 +185,7 @@ python convert.py <path to model directory>
 python convert-falcon-hf-to-gguf.py <path to model directory>
 
 # GPTNeoXForCausalLM
-python convert-falcon-hf-to-gguf.py <path to model directory>
+python convert-gptneox-hf-to-gguf.py <path to model directory>
 
 # GPTBigCodeForCausalLM
 python convert-starcoder-hf-to-gguf.py <path to model directory>

docs/linux.md

@@ -1,12 +1,16 @@
-# Installing Ollama on Linux
+# Ollama on Linux
 
-> Note: A one line installer for Ollama is available by running:
+## Install
+
+Install Ollama running this one-liner:
 >
-> ```bash
-> curl https://ollama.ai/install.sh | sh
-> ```
+```bash
+curl https://ollama.ai/install.sh | sh
+```
 
-## Download the `ollama` binary
+## Manual install
+
+### Download the `ollama` binary
 
 Ollama is distributed as a self-contained binary. Download it to a directory in your PATH:
 
@@ -15,31 +19,7 @@ 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)
+### Adding Ollama as a startup service (recommended)
 
 Create a user for Ollama:
 
@@ -60,7 +40,6 @@ User=ollama
 Group=ollama
 Restart=always
 RestartSec=3
-Environment="HOME=/usr/share/ollama"
 
 [Install]
 WantedBy=default.target
@@ -73,10 +52,65 @@ sudo systemctl daemon-reload
 sudo systemctl enable ollama
 ```
 
-### Viewing logs
+### 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
+```
+
+### Start Ollama
+
+Start Ollama using `systemd`:
+
+```bash
+sudo systemctl start ollama
+```
+
+## Update
+
+Update ollama by running the install script again:
+
+```bash
+curl https://ollama.ai/install.sh | sh
+```
+
+Or by downloading the ollama binary:
+
+```bash
+sudo curl -L https://ollama.ai/download/ollama-linux-amd64 -o /usr/bin/ollama
+sudo chmod +x /usr/bin/ollama
+```
+
+## Viewing logs
 
 To view logs of Ollama running as a startup service, run:
 
 ```bash
 journalctl -u ollama
 ```
+
+## Uninstall
+
+Remove the ollama service:
+
+```bash
+sudo systemctl stop ollama
+sudo systemctl disable ollama
+sudo rm /etc/systemd/system/ollama.service
+```
+
+Remove the ollama binary from your bin directory (either `/usr/local/bin`, `/usr/bin`, or `/bin`):
+
+```bash
+sudo rm $(which ollama)
+```
+
+Remove the downloaded models and Ollama service user:
+```bash
+sudo rm -r /usr/share/ollama
+sudo userdel ollama
+```

examples/langchain-python-rag-privategpt/constants.py

@@ -6,7 +6,6 @@ PERSIST_DIRECTORY = os.environ.get('PERSIST_DIRECTORY', 'db')
 
 # Define the Chroma settings
 CHROMA_SETTINGS = Settings(
-        chroma_db_impl='duckdb+parquet',
         persist_directory=PERSIST_DIRECTORY,
         anonymized_telemetry=False
 )

examples/langchain-python-rag-privategpt/ingest.py

@@ -150,7 +150,7 @@ def main():
         print("Creating new vectorstore")
         texts = process_documents()
         print(f"Creating embeddings. May take some minutes...")
-        db = Chroma.from_documents(texts, embeddings, persist_directory=persist_directory, client_settings=CHROMA_SETTINGS)
+        db = Chroma.from_documents(texts, embeddings, persist_directory=persist_directory)
     db.persist()
     db = None
 

examples/langchain-python-rag-privategpt/privateGPT.py

@@ -4,6 +4,7 @@ from langchain.embeddings import HuggingFaceEmbeddings
 from langchain.callbacks.streaming_stdout import StreamingStdOutCallbackHandler
 from langchain.vectorstores import Chroma
 from langchain.llms import Ollama
+import chromadb
 import os
 import argparse
 import time
@@ -22,7 +23,9 @@ def main():
     # Parse the command line arguments
     args = parse_arguments()
     embeddings = HuggingFaceEmbeddings(model_name=embeddings_model_name)
-    db = Chroma(persist_directory=persist_directory, embedding_function=embeddings, client_settings=CHROMA_SETTINGS)
+
+    db = Chroma(persist_directory=persist_directory, embedding_function=embeddings)
+
     retriever = db.as_retriever(search_kwargs={"k": target_source_chunks})
     # activate/deactivate the streaming StdOut callback for LLMs
     callbacks = [] if args.mute_stream else [StreamingStdOutCallbackHandler()]

go.mod

@@ -4,11 +4,11 @@ go 1.20
 
 require (
 	github.com/dustin/go-humanize v1.0.1
+	github.com/emirpasic/gods v1.18.1
 	github.com/gin-gonic/gin v1.9.1
 	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
 )
@@ -39,12 +39,12 @@ 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/crypto v0.14.0
 	golang.org/x/exp v0.0.0-20230817173708-d852ddb80c63
-	golang.org/x/net v0.10.0 // indirect
-	golang.org/x/sys v0.11.0 // indirect
-	golang.org/x/term v0.10.0
-	golang.org/x/text v0.10.0 // indirect
+	golang.org/x/net v0.17.0 // indirect
+	golang.org/x/sys v0.13.0 // indirect
+	golang.org/x/term v0.13.0
+	golang.org/x/text v0.13.0 // indirect
 	google.golang.org/protobuf v1.30.0 // indirect
 	gopkg.in/yaml.v3 v3.0.1 // indirect
 )

go.sum

@@ -4,10 +4,6 @@ github.com/bytedance/sonic v1.9.1/go.mod h1:i736AoUSYt75HyZLoJW9ERYxcy6eaN6h4BZX
 github.com/chenzhuoyu/base64x v0.0.0-20211019084208-fb5309c8db06/go.mod h1:DH46F32mSOjUmXrMHnKwZdA8wcEefY7UVqBKYGjpdQY=
 github.com/chenzhuoyu/base64x v0.0.0-20221115062448-fe3a3abad311 h1:qSGYFH7+jGhDF8vLC+iwCD4WpbV1EBDSzWkJODFLams=
 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/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=
 github.com/creack/pty v1.1.9/go.mod h1:oKZEueFk5CKHvIhNR5MUki03XCEU+Q6VDXinZuGJ33E=
 github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
@@ -15,6 +11,8 @@ github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c
 github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
 github.com/dustin/go-humanize v1.0.1 h1:GzkhY7T5VNhEkwH0PVJgjz+fX1rhBrR7pRT3mDkpeCY=
 github.com/dustin/go-humanize v1.0.1/go.mod h1:Mu1zIs6XwVuF/gI1OepvI0qD18qycQx+mFykh5fBlto=
+github.com/emirpasic/gods v1.18.1 h1:FXtiHYKDGKCW2KzwZKx0iC0PQmdlorYgdFG9jPXJ1Bc=
+github.com/emirpasic/gods v1.18.1/go.mod h1:8tpGGwCnJ5H4r6BWwaV6OrWmMoPhUl5jm/FMNAnJvWQ=
 github.com/gabriel-vasile/mimetype v1.4.2 h1:w5qFW6JKBz9Y393Y4q372O9A7cUSequkh1Q7OhCmWKU=
 github.com/gabriel-vasile/mimetype v1.4.2/go.mod h1:zApsH/mKG4w07erKIaJPFiX0Tsq9BFQgN3qGY5GnNgA=
 github.com/gin-contrib/cors v1.4.0 h1:oJ6gwtUl3lqV0WEIwM/LxPF1QZ5qe2lGWdY2+bz7y0g=
@@ -78,8 +76,6 @@ 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=
@@ -118,31 +114,30 @@ golang.org/x/arch v0.0.0-20210923205945-b76863e36670/go.mod h1:5om86z9Hs0C8fWVUu
 golang.org/x/arch v0.3.0 h1:02VY4/ZcO/gBOH6PUaoiptASxtXU10jazRCP865E97k=
 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/crypto v0.14.0 h1:wBqGXzWJW6m1XrIKlAH0Hs1JJ7+9KBwnIO8v66Q9cHc=
+golang.org/x/crypto v0.14.0/go.mod h1:MVFd36DqK4CsrnJYDkBA3VC4m2GkXAM0PvzMCn4JQf4=
 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/net v0.17.0 h1:pVaXccu2ozPjCXewfr1S7xza/zcXTity9cCdXQYSjIM=
+golang.org/x/net v0.17.0/go.mod h1:NxSsAGuq816PNPmqtQdLE42eU2Fs7NoRIZrHJAlaCOE=
 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=
 golang.org/x/sys v0.0.0-20210806184541-e5e7981a1069/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
-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 v0.13.0 h1:Af8nKPmuFypiUBjVoU9V20FiaFXOcuZI21p0ycVYYGE=
+golang.org/x/sys v0.13.0/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=
+golang.org/x/term v0.13.0 h1:bb+I9cTfFazGW51MZqBVmZy7+JEJMouUHTUSKVQLBek=
+golang.org/x/term v0.13.0/go.mod h1:LTmsnFJwVN6bCy1rVCoS+qHT1HhALEFxKncY3WNNh4U=
 golang.org/x/text v0.3.3/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
 golang.org/x/text v0.3.6/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
-golang.org/x/text v0.10.0 h1:UpjohKhiEgNc0CSauXmwYftY1+LlaC75SJwh0SgCX58=
-golang.org/x/text v0.10.0/go.mod h1:TvPlkZtksWOMsz7fbANvkp4WM8x/WCo/om8BMLbz+aE=
+golang.org/x/text v0.13.0 h1:ablQoSUd0tRdKxZewP80B+BaqeKJuVhuRxj/dkrun3k=
+golang.org/x/text v0.13.0/go.mod h1:TvPlkZtksWOMsz7fbANvkp4WM8x/WCo/om8BMLbz+aE=
 golang.org/x/tools v0.0.0-20180917221912-90fa682c2a6e/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
 golang.org/x/xerrors v0.0.0-20191204190536-9bdfabe68543/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 google.golang.org/protobuf v1.26.0-rc.1/go.mod h1:jlhhOSvTdKEhbULTjvd4ARK9grFBp09yW+WbY/TyQbw=

llm/ggml.go

@@ -175,7 +175,8 @@ const (
 	// Magic constant for `ggla` files (LoRA adapter).
 	FILE_MAGIC_GGLA = 0x67676C61
 	// Magic constant for `gguf` files (versioned, gguf)
-	FILE_MAGIC_GGUF = 0x46554747
+	FILE_MAGIC_GGUF_LE = 0x46554747
+	FILE_MAGIC_GGUF_BE = 0x47475546
 )
 
 func DecodeGGML(r io.ReadSeeker) (*GGML, error) {
@@ -191,8 +192,10 @@ func DecodeGGML(r io.ReadSeeker) (*GGML, error) {
 		ggml.container = &containerGGJT{}
 	case FILE_MAGIC_GGLA:
 		ggml.container = &containerLORA{}
-	case FILE_MAGIC_GGUF:
-		ggml.container = &containerGGUF{}
+	case FILE_MAGIC_GGUF_LE:
+		ggml.container = &containerGGUF{bo: binary.LittleEndian}
+	case FILE_MAGIC_GGUF_BE:
+		ggml.container = &containerGGUF{bo: binary.BigEndian}
 	default:
 		return nil, errors.New("invalid file magic")
 	}

llm/gguf.go

@@ -3,12 +3,13 @@ package llm
 import (
 	"bytes"
 	"encoding/binary"
-	"errors"
 	"fmt"
 	"io"
 )
 
 type containerGGUF struct {
+	bo binary.ByteOrder
+
 	Version uint32
 
 	V1 struct {
@@ -27,15 +28,13 @@ func (c *containerGGUF) Name() string {
 }
 
 func (c *containerGGUF) Decode(r io.Reader) (model, error) {
-	binary.Read(r, binary.LittleEndian, &c.Version)
+	binary.Read(r, c.bo, &c.Version)
 
 	switch c.Version {
 	case 1:
-		binary.Read(r, binary.LittleEndian, &c.V1)
-	case 2:
-		binary.Read(r, binary.LittleEndian, &c.V2)
+		binary.Read(r, c.bo, &c.V1)
 	default:
-		return nil, errors.New("invalid version")
+		binary.Read(r, c.bo, &c.V2)
 	}
 
 	model := newGGUFModel(c)
@@ -209,75 +208,75 @@ func (llm *ggufModel) NumLayers() int64 {
 	return int64(v)
 }
 
-func (ggufModel) readU8(r io.Reader) uint8 {
+func (llm ggufModel) readU8(r io.Reader) uint8 {
 	var u8 uint8
-	binary.Read(r, binary.LittleEndian, &u8)
+	binary.Read(r, llm.bo, &u8)
 	return u8
 }
 
-func (ggufModel) readI8(r io.Reader) int8 {
+func (llm ggufModel) readI8(r io.Reader) int8 {
 	var i8 int8
-	binary.Read(r, binary.LittleEndian, &i8)
+	binary.Read(r, llm.bo, &i8)
 	return i8
 }
 
-func (ggufModel) readU16(r io.Reader) uint16 {
+func (llm ggufModel) readU16(r io.Reader) uint16 {
 	var u16 uint16
-	binary.Read(r, binary.LittleEndian, &u16)
+	binary.Read(r, llm.bo, &u16)
 	return u16
 }
 
-func (ggufModel) readI16(r io.Reader) int16 {
+func (llm ggufModel) readI16(r io.Reader) int16 {
 	var i16 int16
-	binary.Read(r, binary.LittleEndian, &i16)
+	binary.Read(r, llm.bo, &i16)
 	return i16
 }
 
-func (ggufModel) readU32(r io.Reader) uint32 {
+func (llm ggufModel) readU32(r io.Reader) uint32 {
 	var u32 uint32
-	binary.Read(r, binary.LittleEndian, &u32)
+	binary.Read(r, llm.bo, &u32)
 	return u32
 }
 
-func (ggufModel) readI32(r io.Reader) int32 {
+func (llm ggufModel) readI32(r io.Reader) int32 {
 	var i32 int32
-	binary.Read(r, binary.LittleEndian, &i32)
+	binary.Read(r, llm.bo, &i32)
 	return i32
 }
 
-func (ggufModel) readU64(r io.Reader) uint64 {
+func (llm ggufModel) readU64(r io.Reader) uint64 {
 	var u64 uint64
-	binary.Read(r, binary.LittleEndian, &u64)
+	binary.Read(r, llm.bo, &u64)
 	return u64
 }
 
-func (ggufModel) readI64(r io.Reader) int64 {
+func (llm ggufModel) readI64(r io.Reader) int64 {
 	var i64 int64
-	binary.Read(r, binary.LittleEndian, &i64)
+	binary.Read(r, llm.bo, &i64)
 	return i64
 }
 
-func (ggufModel) readF32(r io.Reader) float32 {
+func (llm ggufModel) readF32(r io.Reader) float32 {
 	var f32 float32
-	binary.Read(r, binary.LittleEndian, &f32)
+	binary.Read(r, llm.bo, &f32)
 	return f32
 }
 
-func (ggufModel) readF64(r io.Reader) float64 {
+func (llm ggufModel) readF64(r io.Reader) float64 {
 	var f64 float64
-	binary.Read(r, binary.LittleEndian, &f64)
+	binary.Read(r, llm.bo, &f64)
 	return f64
 }
 
-func (ggufModel) readBool(r io.Reader) bool {
+func (llm ggufModel) readBool(r io.Reader) bool {
 	var b bool
-	binary.Read(r, binary.LittleEndian, &b)
+	binary.Read(r, llm.bo, &b)
 	return b
 }
 
-func (ggufModel) readStringV1(r io.Reader) (string, error) {
+func (llm ggufModel) readStringV1(r io.Reader) (string, error) {
 	var nameLength uint32
-	binary.Read(r, binary.LittleEndian, &nameLength)
+	binary.Read(r, llm.bo, &nameLength)
 
 	var b bytes.Buffer
 	if _, err := io.CopyN(&b, r, int64(nameLength)); err != nil {
@@ -292,7 +291,7 @@ func (ggufModel) readStringV1(r io.Reader) (string, error) {
 
 func (llm ggufModel) readString(r io.Reader) (string, error) {
 	var nameLength uint64
-	binary.Read(r, binary.LittleEndian, &nameLength)
+	binary.Read(r, llm.bo, &nameLength)
 
 	var b bytes.Buffer
 	if _, err := io.CopyN(&b, r, int64(nameLength)); err != nil {

llm/llama.cpp/generate_darwin_amd64.go

@@ -12,7 +12,8 @@ package llm
 //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 git -C gguf apply ../patches/0001-update-default-log-target.patch
+//go:generate git -C gguf apply ../patches/0001-metal-handle-ggml_scale-for-n-4-0-close-3754.patch
 //go:generate cmake -S gguf -B gguf/build/cpu -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=x86_64 -DCMAKE_OSX_ARCHITECTURES=x86_64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
 //go:generate cmake --build gguf/build/cpu --target server --config Release
 //go:generate mv gguf/build/cpu/bin/server gguf/build/cpu/bin/ollama-runner

llm/llama.cpp/generate_darwin_arm64.go

@@ -12,7 +12,8 @@ package llm
 //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 git -C gguf apply ../patches/0001-update-default-log-target.patch
+//go:generate git -C gguf apply ../patches/0001-metal-handle-ggml_scale-for-n-4-0-close-3754.patch
 //go:generate cmake -S gguf -B gguf/build/metal -DLLAMA_METAL=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=arm64 -DCMAKE_OSX_ARCHITECTURES=arm64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
 //go:generate cmake --build gguf/build/metal --target server --config Release
 //go:generate mv gguf/build/metal/bin/server gguf/build/metal/bin/ollama-runner

llm/llama.cpp/generate_linux.go

@@ -13,14 +13,14 @@ package llm
 
 //go:generate git submodule update --force gguf
 //go:generate git -C gguf apply ../patches/0001-copy-cuda-runtime-libraries.patch
-//go:generate git -C gguf apply ../patches/0001-remove-warm-up-logging.patch
-//go:generate cmake -S gguf -B gguf/build/cpu -DLLAMA_K_QUANTS=on
+//go:generate git -C gguf apply ../patches/0001-update-default-log-target.patch
+//go:generate cmake -S gguf -B gguf/build/cpu -DLLAMA_K_QUANTS=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off
 //go:generate cmake --build gguf/build/cpu --target server --config Release
 //go:generate mv gguf/build/cpu/bin/server gguf/build/cpu/bin/ollama-runner
 
 //go:generate cmake -S ggml -B ggml/build/cuda -DLLAMA_CUBLAS=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on
 //go:generate cmake --build ggml/build/cuda --target server --config Release
 //go:generate mv ggml/build/cuda/bin/server ggml/build/cuda/bin/ollama-runner
-//go:generate cmake -S gguf -B gguf/build/cuda -DLLAMA_CUBLAS=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on
+//go:generate cmake -S gguf -B gguf/build/cuda -DLLAMA_CUBLAS=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off
 //go:generate cmake --build gguf/build/cuda --target server --config Release
 //go:generate mv gguf/build/cuda/bin/server gguf/build/cuda/bin/ollama-runner

llm/llama.cpp/generate_windows.go

@@ -10,7 +10,7 @@ package llm
 //go:generate cmd /c move ggml\build\cpu\bin\Release\server.exe ggml\build\cpu\bin\Release\ollama-runner.exe
 
 //go:generate git submodule update --force gguf
-//go:generate git -C gguf apply ../patches/0001-remove-warm-up-logging.patch
-//go:generate cmake -S gguf -B gguf/build/cpu -DLLAMA_K_QUANTS=on
+//go:generate git -C gguf apply ../patches/0001-update-default-log-target.patch
+//go:generate cmake -S gguf -B gguf/build/cpu -DLLAMA_K_QUANTS=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off
 //go:generate cmake --build gguf/build/cpu --target server --config Release
 //go:generate cmd /c move gguf\build\cpu\bin\Release\server.exe gguf\build\cpu\bin\Release\ollama-runner.exe

llm/llama.cpp/patches/0001-metal-handle-ggml_scale-for-n-4-0-close-3754.patch

@@ -0,0 +1,91 @@
+From 469c9addef75893e6be12edda852d12e840bf064 Mon Sep 17 00:00:00 2001
+From: Georgi Gerganov <ggerganov@gmail.com>
+Date: Tue, 24 Oct 2023 09:46:50 +0300
+Subject: [PATCH 1/2] metal : handle ggml_scale for n%4 != 0 (close #3754)
+
+ggml-ci
+---
+ ggml-metal.m     | 18 +++++++++++++-----
+ ggml-metal.metal | 10 +++++++++-
+ 2 files changed, 22 insertions(+), 6 deletions(-)
+
+diff --git a/ggml-metal.m b/ggml-metal.m
+index c908106..c1901dc 100644
+--- a/ggml-metal.m
++++ b/ggml-metal.m
+@@ -62,6 +62,7 @@
+     GGML_METAL_DECL_KERNEL(mul);
+     GGML_METAL_DECL_KERNEL(mul_row); // TODO: avoid this extra kernel, instead extend the "mul" kernel to support broadcast
+     GGML_METAL_DECL_KERNEL(scale);
++    GGML_METAL_DECL_KERNEL(scale_4);
+     GGML_METAL_DECL_KERNEL(silu);
+     GGML_METAL_DECL_KERNEL(relu);
+     GGML_METAL_DECL_KERNEL(gelu);
+@@ -249,6 +250,7 @@ static void ggml_metal_log(enum ggml_log_level level, const char* format, ...){
+         GGML_METAL_ADD_KERNEL(mul);
+         GGML_METAL_ADD_KERNEL(mul_row);
+         GGML_METAL_ADD_KERNEL(scale);
++        GGML_METAL_ADD_KERNEL(scale_4);
+         GGML_METAL_ADD_KERNEL(silu);
+         GGML_METAL_ADD_KERNEL(relu);
+         GGML_METAL_ADD_KERNEL(gelu);
+@@ -347,6 +349,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
+     GGML_METAL_DEL_KERNEL(mul);
+     GGML_METAL_DEL_KERNEL(mul_row);
+     GGML_METAL_DEL_KERNEL(scale);
++    GGML_METAL_DEL_KERNEL(scale_4);
+     GGML_METAL_DEL_KERNEL(silu);
+     GGML_METAL_DEL_KERNEL(relu);
+     GGML_METAL_DEL_KERNEL(gelu);
+@@ -923,15 +926,20 @@ void ggml_metal_graph_compute(
+ 
+                             const float scale = *(const float *) src1->data;
+ 
+-                            [encoder setComputePipelineState:ctx->pipeline_scale];
++                            int64_t n = ggml_nelements(dst);
++
++                            if (n % 4 == 0) {
++                                n /= 4;
++                                [encoder setComputePipelineState:ctx->pipeline_scale_4];
++                            } else {
++                                [encoder setComputePipelineState:ctx->pipeline_scale];
++                            }
++
+                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                             [encoder setBytes:&scale length:sizeof(scale) atIndex:2];
+ 
+-                            const int64_t n = ggml_nelements(dst);
+-                            GGML_ASSERT(n % 4 == 0);
+-
+-                            [encoder dispatchThreadgroups:MTLSizeMake(n/4, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
++                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                         } break;
+                     case GGML_OP_UNARY:
+                         switch (ggml_get_unary_op(gf->nodes[i])) {
+diff --git a/ggml-metal.metal b/ggml-metal.metal
+index 69fc713..f4b4605 100644
+--- a/ggml-metal.metal
++++ b/ggml-metal.metal
+@@ -125,9 +125,17 @@ kernel void kernel_mul_row(
+ }
+ 
+ kernel void kernel_scale(
++        device const float * src0,
++        device       float * dst,
++        constant     float & scale,
++        uint tpig[[thread_position_in_grid]]) {
++    dst[tpig] = src0[tpig] * scale;
++}
++
++kernel void kernel_scale_4(
+         device const float4 * src0,
+         device       float4 * dst,
+-        constant     float & scale,
++        constant     float  & scale,
+         uint tpig[[thread_position_in_grid]]) {
+     dst[tpig] = src0[tpig] * scale;
+ }
+-- 
+2.39.3 (Apple Git-145)
+

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

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

llm/llama.cpp/patches/0001-update-default-log-target.patch

@@ -0,0 +1,25 @@
+From 6465fec6290f0a7f5d4d0fbe6bcf634e4810dde6 Mon Sep 17 00:00:00 2001
+From: Michael Yang <mxyng@pm.me>
+Date: Mon, 23 Oct 2023 10:39:34 -0700
+Subject: [PATCH] default log stderr
+
+---
+ common/log.h | 2 +-
+ 1 file changed, 1 insertion(+), 1 deletion(-)
+
+diff --git a/common/log.h b/common/log.h
+index b8953fd..25522cd 100644
+--- a/common/log.h
++++ b/common/log.h
+@@ -90,7 +90,7 @@
+ //  }
+ //
+ #ifndef LOG_TARGET
+-    #define LOG_TARGET log_handler()
++    #define LOG_TARGET nullptr
+ #endif
+ 
+ #ifndef LOG_TEE_TARGET
+-- 
+2.42.0
+

llm/llama.go

@@ -212,6 +212,10 @@ func CheckVRAM() (int64, error) {
 	scanner := bufio.NewScanner(&stdout)
 	for scanner.Scan() {
 		line := scanner.Text()
+		if strings.Contains(line, "[Insufficient Permissions]") {
+			return 0, fmt.Errorf("GPU support may not enabled, check you have installed GPU drivers and have the necessary permissions to run nvidia-smi")
+		}
+
 		vram, err := strconv.ParseInt(strings.TrimSpace(line), 10, 64)
 		if err != nil {
 			return 0, fmt.Errorf("failed to parse available VRAM: %v", err)
@@ -243,12 +247,15 @@ func NumGPU(numLayer, fileSizeBytes int64, opts api.Options) int {
 			return 0
 		}
 
-		// Calculate bytes per layer
-		// TODO: this is a rough heuristic, better would be to calculate this based on number of layers and context size
+		/*
+		 Calculate bytes per layer, this will roughly be the size of the model file divided by the number of layers.
+		 We can store the model weights and the kv cache in vram,
+		 to enable kv chache vram storage add two additional layers to the number of layers retrieved from the model file.
+		*/
 		bytesPerLayer := fileSizeBytes / numLayer
 
-		// max number of layers we can fit in VRAM, subtract 8% to prevent consuming all available VRAM and running out of memory
-		layers := int(freeBytes/bytesPerLayer) * 92 / 100
+		// 75% of the absolute max number of layers we can fit in available VRAM, off-loading too many layers to the GPU can cause OOM errors
+		layers := int(freeBytes/bytesPerLayer) * 3 / 4
 		log.Printf("%d MB VRAM available, loading up to %d GPU layers", freeBytes/(1024*1024), layers)
 
 		return layers
@@ -299,13 +306,19 @@ func newLlama(model string, adapters []string, runners []ModelRunner, numLayers
 	params := []string{
 		"--model", model,
 		"--ctx-size", fmt.Sprintf("%d", opts.NumCtx),
-		"--rope-freq-base", fmt.Sprintf("%f", opts.RopeFrequencyBase),
-		"--rope-freq-scale", fmt.Sprintf("%f", opts.RopeFrequencyScale),
 		"--batch-size", fmt.Sprintf("%d", opts.NumBatch),
 		"--n-gpu-layers", fmt.Sprintf("%d", numGPU),
 		"--embedding",
 	}
 
+	if opts.RopeFrequencyBase > 0 {
+		params = append(params, "--rope-freq-base", fmt.Sprintf("%f", opts.RopeFrequencyBase))
+	}
+
+	if opts.RopeFrequencyScale > 0 {
+		params = append(params, "--rope-freq-scale", fmt.Sprintf("%f", opts.RopeFrequencyScale))
+	}
+
 	if opts.NumGQA > 0 {
 		params = append(params, "--gqa", fmt.Sprintf("%d", opts.NumGQA))
 	}
@@ -353,7 +366,15 @@ func newLlama(model string, adapters []string, runners []ModelRunner, numLayers
 			runner.Path,
 			append(params, "--port", strconv.Itoa(port))...,
 		)
-		cmd.Env = append(os.Environ(), fmt.Sprintf("LD_LIBRARY_PATH=%s", filepath.Dir(runner.Path)))
+
+		var libraryPaths []string
+		if libraryPath, ok := os.LookupEnv("LD_LIBRARY_PATH"); ok {
+			libraryPaths = append(libraryPaths, libraryPath)
+		}
+
+		libraryPaths = append(libraryPaths, filepath.Dir(runner.Path))
+
+		cmd.Env = append(os.Environ(), fmt.Sprintf("LD_LIBRARY_PATH=%s", strings.Join(libraryPaths, ":")))
 		cmd.Stdout = os.Stderr
 		statusWriter := NewStatusWriter()
 		cmd.Stderr = statusWriter

llm/llm.go

@@ -85,7 +85,10 @@ func New(workDir, model string, adapters []string, opts api.Options) (LLM, error
 
 	switch ggml.Name() {
 	case "gguf":
-		opts.NumGQA = 0 // TODO: remove this when llama.cpp runners differ enough to need separate newLlama functions
+		// TODO: gguf will load these options automatically from the model binary
+		opts.NumGQA = 0
+		opts.RopeFrequencyBase = 0.0
+		opts.RopeFrequencyScale = 0.0
 		return newLlama(model, adapters, chooseRunners(workDir, "gguf"), ggml.NumLayers(), opts)
 	case "ggml", "ggmf", "ggjt", "ggla":
 		return newLlama(model, adapters, chooseRunners(workDir, "ggml"), ggml.NumLayers(), opts)

readline/buffer.go

@@ -0,0 +1,372 @@
+package readline
+
+import (
+	"fmt"
+	"os"
+
+	"github.com/emirpasic/gods/lists/arraylist"
+	"golang.org/x/term"
+)
+
+type Buffer struct {
+	Pos       int
+	Buf       *arraylist.List
+	Prompt    *Prompt
+	LineWidth int
+	Width     int
+	Height    int
+}
+
+func NewBuffer(prompt *Prompt) (*Buffer, error) {
+	fd := int(os.Stdout.Fd())
+	width, height, err := term.GetSize(fd)
+	if err != nil {
+		fmt.Println("Error getting size:", err)
+		return nil, err
+	}
+
+	lwidth := width - len(prompt.Prompt)
+	if prompt.UseAlt {
+		lwidth = width - len(prompt.AltPrompt)
+	}
+
+	b := &Buffer{
+		Pos:       0,
+		Buf:       arraylist.New(),
+		Prompt:    prompt,
+		Width:     width,
+		Height:    height,
+		LineWidth: lwidth,
+	}
+
+	return b, nil
+}
+
+func (b *Buffer) MoveLeft() {
+	if b.Pos > 0 {
+		if b.Pos%b.LineWidth == 0 {
+			fmt.Printf(CursorUp + CursorBOL + cursorRightN(b.Width))
+		} else {
+			fmt.Print(CursorLeft)
+		}
+		b.Pos -= 1
+	}
+}
+
+func (b *Buffer) MoveLeftWord() {
+	if b.Pos > 0 {
+		var foundNonspace bool
+		for {
+			v, _ := b.Buf.Get(b.Pos - 1)
+			if v == ' ' {
+				if foundNonspace {
+					break
+				}
+			} else {
+				foundNonspace = true
+			}
+			b.MoveLeft()
+
+			if b.Pos == 0 {
+				break
+			}
+		}
+	}
+}
+
+func (b *Buffer) MoveRight() {
+	if b.Pos < b.Size() {
+		b.Pos += 1
+		if b.Pos%b.LineWidth == 0 {
+			fmt.Printf(CursorDown + CursorBOL + cursorRightN(b.PromptSize()))
+		} else {
+			fmt.Print(CursorRight)
+		}
+	}
+}
+
+func (b *Buffer) MoveRightWord() {
+	if b.Pos < b.Size() {
+		for {
+			b.MoveRight()
+			v, _ := b.Buf.Get(b.Pos)
+			if v == ' ' {
+				break
+			}
+
+			if b.Pos == b.Size() {
+				break
+			}
+		}
+	}
+}
+
+func (b *Buffer) MoveToStart() {
+	if b.Pos > 0 {
+		currLine := b.Pos / b.LineWidth
+		if currLine > 0 {
+			for cnt := 0; cnt < currLine; cnt++ {
+				fmt.Print(CursorUp)
+			}
+		}
+		fmt.Printf(CursorBOL + cursorRightN(b.PromptSize()))
+		b.Pos = 0
+	}
+}
+
+func (b *Buffer) MoveToEnd() {
+	if b.Pos < b.Size() {
+		currLine := b.Pos / b.LineWidth
+		totalLines := b.Size() / b.LineWidth
+		if currLine < totalLines {
+			for cnt := 0; cnt < totalLines-currLine; cnt++ {
+				fmt.Print(CursorDown)
+			}
+			remainder := b.Size() % b.LineWidth
+			fmt.Printf(CursorBOL + cursorRightN(b.PromptSize()+remainder))
+		} else {
+			fmt.Print(cursorRightN(b.Size() - b.Pos))
+		}
+
+		b.Pos = b.Size()
+	}
+}
+
+func (b *Buffer) Size() int {
+	return b.Buf.Size()
+}
+
+func min(n, m int) int {
+	if n > m {
+		return m
+	}
+	return n
+}
+
+func (b *Buffer) PromptSize() int {
+	if b.Prompt.UseAlt {
+		return len(b.Prompt.AltPrompt)
+	}
+	return len(b.Prompt.Prompt)
+}
+
+func (b *Buffer) Add(r rune) {
+	if b.Pos == b.Buf.Size() {
+		fmt.Printf("%c", r)
+		b.Buf.Add(r)
+		b.Pos += 1
+		if b.Pos > 0 && b.Pos%b.LineWidth == 0 {
+			fmt.Printf("\n%s", b.Prompt.AltPrompt)
+		}
+	} else {
+		fmt.Printf("%c", r)
+		b.Buf.Insert(b.Pos, r)
+		b.Pos += 1
+		if b.Pos > 0 && b.Pos%b.LineWidth == 0 {
+			fmt.Printf("\n%s", b.Prompt.AltPrompt)
+		}
+		b.drawRemaining()
+	}
+}
+
+func (b *Buffer) drawRemaining() {
+	var place int
+	remainingText := b.StringN(b.Pos)
+	if b.Pos > 0 {
+		place = b.Pos % b.LineWidth
+	}
+	fmt.Print(CursorHide)
+
+	// render the rest of the current line
+	currLine := remainingText[:min(b.LineWidth-place, len(remainingText))]
+	if len(currLine) > 0 {
+		fmt.Printf(ClearToEOL + currLine)
+		fmt.Print(cursorLeftN(len(currLine)))
+	} else {
+		fmt.Print(ClearToEOL)
+	}
+
+	// render the other lines
+	if len(remainingText) > len(currLine) {
+		remaining := []rune(remainingText[len(currLine):])
+		var totalLines int
+		for i, c := range remaining {
+			if i%b.LineWidth == 0 {
+				fmt.Printf("\n%s", b.Prompt.AltPrompt)
+				totalLines += 1
+			}
+			fmt.Printf("%c", c)
+		}
+		fmt.Print(ClearToEOL)
+		fmt.Print(cursorUpN(totalLines))
+		fmt.Printf(CursorBOL + cursorRightN(b.Width-len(currLine)))
+	}
+
+	fmt.Print(CursorShow)
+}
+
+func (b *Buffer) Remove() {
+	if b.Buf.Size() > 0 && b.Pos > 0 {
+		if b.Pos%b.LineWidth == 0 {
+			// if the user backspaces over the word boundary, do this magic to clear the line
+			// and move to the end of the previous line
+			fmt.Printf(CursorBOL + ClearToEOL)
+			fmt.Printf(CursorUp + CursorBOL + cursorRightN(b.Width) + " " + CursorLeft)
+		} else {
+			fmt.Printf(CursorLeft + " " + CursorLeft)
+		}
+
+		var eraseExtraLine bool
+		if (b.Size()-1)%b.LineWidth == 0 {
+			eraseExtraLine = true
+		}
+
+		b.Pos -= 1
+		b.Buf.Remove(b.Pos)
+
+		if b.Pos < b.Size() {
+			b.drawRemaining()
+			// this erases a line which is left over when backspacing in the middle of a line and there
+			// are trailing characters which go over the line width boundary
+			if eraseExtraLine {
+				remainingLines := (b.Size() - b.Pos) / b.LineWidth
+				fmt.Printf(cursorDownN(remainingLines+1) + CursorBOL + ClearToEOL)
+				place := b.Pos % b.LineWidth
+				fmt.Printf(cursorUpN(remainingLines+1) + cursorRightN(place+len(b.Prompt.Prompt)))
+			}
+		}
+	}
+}
+
+func (b *Buffer) Delete() {
+	if b.Size() > 0 && b.Pos < b.Size() {
+		b.Buf.Remove(b.Pos)
+		b.drawRemaining()
+		if b.Size()%b.LineWidth == 0 {
+			if b.Pos != b.Size() {
+				remainingLines := (b.Size() - b.Pos) / b.LineWidth
+				fmt.Printf(cursorDownN(remainingLines) + CursorBOL + ClearToEOL)
+				place := b.Pos % b.LineWidth
+				fmt.Printf(cursorUpN(remainingLines) + cursorRightN(place+len(b.Prompt.Prompt)))
+			}
+		}
+	}
+}
+
+func (b *Buffer) DeleteBefore() {
+	if b.Pos > 0 {
+		for cnt := b.Pos - 1; cnt >= 0; cnt-- {
+			b.Remove()
+		}
+	}
+}
+
+func (b *Buffer) DeleteRemaining() {
+	if b.Size() > 0 && b.Pos < b.Size() {
+		charsToDel := b.Size() - b.Pos
+		for cnt := 0; cnt < charsToDel; cnt++ {
+			b.Delete()
+		}
+	}
+}
+
+func (b *Buffer) DeleteWord() {
+	if b.Buf.Size() > 0 && b.Pos > 0 {
+		var foundNonspace bool
+		for {
+			v, _ := b.Buf.Get(b.Pos - 1)
+			if v == ' ' {
+				if !foundNonspace {
+					b.Remove()
+				} else {
+					break
+				}
+			} else {
+				foundNonspace = true
+				b.Remove()
+			}
+
+			if b.Pos == 0 {
+				break
+			}
+		}
+	}
+}
+
+func (b *Buffer) ClearScreen() {
+	fmt.Printf(ClearScreen + CursorReset + b.Prompt.Prompt)
+	if b.IsEmpty() {
+		ph := b.Prompt.Placeholder
+		fmt.Printf(ColorGrey + ph + cursorLeftN(len(ph)) + ColorDefault)
+	} else {
+		currPos := b.Pos
+		b.Pos = 0
+		b.drawRemaining()
+		fmt.Printf(CursorReset + cursorRightN(len(b.Prompt.Prompt)))
+		if currPos > 0 {
+			targetLine := currPos / b.LineWidth
+			if targetLine > 0 {
+				for cnt := 0; cnt < targetLine; cnt++ {
+					fmt.Print(CursorDown)
+				}
+			}
+			remainder := currPos % b.LineWidth
+			if remainder > 0 {
+				fmt.Print(cursorRightN(remainder))
+			}
+			if currPos%b.LineWidth == 0 {
+				fmt.Printf(CursorBOL + b.Prompt.AltPrompt)
+			}
+		}
+		b.Pos = currPos
+	}
+}
+
+func (b *Buffer) IsEmpty() bool {
+	return b.Buf.Empty()
+}
+
+func (b *Buffer) Replace(r []rune) {
+	b.Pos = 0
+	b.Buf.Clear()
+	fmt.Printf(ClearLine + CursorBOL + b.Prompt.Prompt)
+	for _, c := range r {
+		b.Add(c)
+	}
+}
+
+func (b *Buffer) String() string {
+	return b.StringN(0)
+}
+
+func (b *Buffer) StringN(n int) string {
+	return b.StringNM(n, 0)
+}
+
+func (b *Buffer) StringNM(n, m int) string {
+	var s string
+	if m == 0 {
+		m = b.Size()
+	}
+	for cnt := n; cnt < m; cnt++ {
+		c, _ := b.Buf.Get(cnt)
+		s += string(c.(rune))
+	}
+	return s
+}
+
+func cursorLeftN(n int) string {
+	return fmt.Sprintf(CursorLeftN, n)
+}
+
+func cursorRightN(n int) string {
+	return fmt.Sprintf(CursorRightN, n)
+}
+
+func cursorUpN(n int) string {
+	return fmt.Sprintf(CursorUpN, n)
+}
+
+func cursorDownN(n int) string {
+	return fmt.Sprintf(CursorDownN, n)
+}

readline/errors.go

@@ -0,0 +1,17 @@
+package readline
+
+import (
+	"errors"
+)
+
+var (
+	ErrInterrupt = errors.New("Interrupt")
+)
+
+type InterruptError struct {
+	Line []rune
+}
+
+func (*InterruptError) Error() string {
+	return "Interrupted"
+}

readline/history.go

@@ -0,0 +1,152 @@
+package readline
+
+import (
+	"bufio"
+	"errors"
+	"io"
+	"os"
+	"path/filepath"
+	"strings"
+
+	"github.com/emirpasic/gods/lists/arraylist"
+)
+
+type History struct {
+	Buf      *arraylist.List
+	Autosave bool
+	Pos      int
+	Limit    int
+	Filename string
+	Enabled  bool
+}
+
+func NewHistory() (*History, error) {
+	h := &History{
+		Buf:      arraylist.New(),
+		Limit:    100, //resizeme
+		Autosave: true,
+		Enabled:  true,
+	}
+
+	err := h.Init()
+	if err != nil {
+		return nil, err
+	}
+
+	return h, nil
+}
+
+func (h *History) Init() error {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return err
+	}
+
+	path := filepath.Join(home, ".ollama", "history")
+	h.Filename = path
+
+	//todo check if the file exists
+	f, err := os.OpenFile(path, os.O_CREATE|os.O_RDONLY, 0600)
+	if err != nil {
+		if errors.Is(err, os.ErrNotExist) {
+			return nil
+		}
+		return err
+	}
+	defer f.Close()
+
+	r := bufio.NewReader(f)
+	for {
+		line, err := r.ReadString('\n')
+		if err != nil {
+			if err == io.EOF {
+				break
+			}
+			return err
+		}
+
+		line = strings.TrimSpace(line)
+		if len(line) == 0 {
+			continue
+		}
+
+		h.Add([]rune(line))
+	}
+
+	return nil
+}
+
+func (h *History) Add(l []rune) {
+	h.Buf.Add(l)
+	h.Compact()
+	h.Pos = h.Size()
+	if h.Autosave {
+		h.Save()
+	}
+}
+
+func (h *History) Compact() {
+	s := h.Buf.Size()
+	if s > h.Limit {
+		for cnt := 0; cnt < s-h.Limit; cnt++ {
+			h.Buf.Remove(0)
+		}
+	}
+}
+
+func (h *History) Clear() {
+	h.Buf.Clear()
+}
+
+func (h *History) Prev() []rune {
+	var line []rune
+	if h.Pos > 0 {
+		h.Pos -= 1
+	}
+	v, _ := h.Buf.Get(h.Pos)
+	line, _ = v.([]rune)
+	return line
+}
+
+func (h *History) Next() []rune {
+	var line []rune
+	if h.Pos < h.Buf.Size() {
+		h.Pos += 1
+		v, _ := h.Buf.Get(h.Pos)
+		line, _ = v.([]rune)
+	}
+	return line
+}
+
+func (h *History) Size() int {
+	return h.Buf.Size()
+}
+
+func (h *History) Save() error {
+	if !h.Enabled {
+		return nil
+	}
+
+	tmpFile := h.Filename + ".tmp"
+
+	f, err := os.OpenFile(tmpFile, os.O_CREATE|os.O_WRONLY|os.O_TRUNC|os.O_APPEND, 0666)
+	if err != nil {
+		return err
+	}
+	defer f.Close()
+
+	buf := bufio.NewWriter(f)
+	for cnt := 0; cnt < h.Size(); cnt++ {
+		v, _ := h.Buf.Get(cnt)
+		line, _ := v.([]rune)
+		buf.WriteString(string(line) + "\n")
+	}
+	buf.Flush()
+	f.Close()
+
+	if err = os.Rename(tmpFile, h.Filename); err != nil {
+		return err
+	}
+
+	return nil
+}

readline/readline.go

@@ -0,0 +1,254 @@
+package readline
+
+import (
+	"bufio"
+	"fmt"
+	"io"
+	"os"
+	"syscall"
+)
+
+type Prompt struct {
+	Prompt         string
+	AltPrompt      string
+	Placeholder    string
+	AltPlaceholder string
+	UseAlt         bool
+}
+
+type Terminal struct {
+	outchan chan rune
+}
+
+type Instance struct {
+	Prompt   *Prompt
+	Terminal *Terminal
+	History  *History
+}
+
+func New(prompt Prompt) (*Instance, error) {
+	term, err := NewTerminal()
+	if err != nil {
+		return nil, err
+	}
+
+	history, err := NewHistory()
+	if err != nil {
+		return nil, err
+	}
+
+	return &Instance{
+		Prompt:   &prompt,
+		Terminal: term,
+		History:  history,
+	}, nil
+}
+
+func (i *Instance) Readline() (string, error) {
+	prompt := i.Prompt.Prompt
+	if i.Prompt.UseAlt {
+		prompt = i.Prompt.AltPrompt
+	}
+	fmt.Print(prompt)
+
+	fd := int(syscall.Stdin)
+	termios, err := SetRawMode(fd)
+	if err != nil {
+		return "", err
+	}
+	defer UnsetRawMode(fd, termios)
+
+	buf, _ := NewBuffer(i.Prompt)
+
+	var esc bool
+	var escex bool
+	var metaDel bool
+	var pasteMode PasteMode
+
+	var currentLineBuf []rune
+
+	for {
+		if buf.IsEmpty() {
+			ph := i.Prompt.Placeholder
+			if i.Prompt.UseAlt {
+				ph = i.Prompt.AltPlaceholder
+			}
+			fmt.Printf(ColorGrey + ph + fmt.Sprintf(CursorLeftN, len(ph)) + ColorDefault)
+		}
+
+		r, err := i.Terminal.Read()
+
+		if buf.IsEmpty() {
+			fmt.Print(ClearToEOL)
+		}
+
+		if err != nil {
+			return "", io.EOF
+		}
+
+		if escex {
+			escex = false
+
+			switch r {
+			case KeyUp:
+				if i.History.Pos > 0 {
+					if i.History.Pos == i.History.Size() {
+						currentLineBuf = []rune(buf.String())
+					}
+					buf.Replace(i.History.Prev())
+				}
+			case KeyDown:
+				if i.History.Pos < i.History.Size() {
+					buf.Replace(i.History.Next())
+					if i.History.Pos == i.History.Size() {
+						buf.Replace(currentLineBuf)
+					}
+				}
+			case KeyLeft:
+				buf.MoveLeft()
+			case KeyRight:
+				buf.MoveRight()
+			case CharBracketedPaste:
+				var code string
+				for cnt := 0; cnt < 3; cnt++ {
+					r, err = i.Terminal.Read()
+					if err != nil {
+						return "", io.EOF
+					}
+
+					code += string(r)
+				}
+				if code == CharBracketedPasteStart {
+					pasteMode = PasteModeStart
+				} else if code == CharBracketedPasteEnd {
+					pasteMode = PasteModeEnd
+				}
+			case KeyDel:
+				if buf.Size() > 0 {
+					buf.Delete()
+				}
+				metaDel = true
+			case MetaStart:
+				buf.MoveToStart()
+			case MetaEnd:
+				buf.MoveToEnd()
+			default:
+				// skip any keys we don't know about
+				continue
+			}
+			continue
+		} else if esc {
+			esc = false
+
+			switch r {
+			case 'b':
+				buf.MoveLeftWord()
+			case 'f':
+				buf.MoveRightWord()
+			case CharEscapeEx:
+				escex = true
+			}
+			continue
+		}
+
+		switch r {
+		case CharNull:
+			continue
+		case CharEsc:
+			esc = true
+		case CharInterrupt:
+			return "", ErrInterrupt
+		case CharLineStart:
+			buf.MoveToStart()
+		case CharLineEnd:
+			buf.MoveToEnd()
+		case CharBackward:
+			buf.MoveLeft()
+		case CharForward:
+			buf.MoveRight()
+		case CharBackspace, CharCtrlH:
+			buf.Remove()
+		case CharTab:
+			// todo: convert back to real tabs
+			for cnt := 0; cnt < 8; cnt++ {
+				buf.Add(' ')
+			}
+		case CharDelete:
+			if buf.Size() > 0 {
+				buf.Delete()
+			} else {
+				return "", io.EOF
+			}
+		case CharKill:
+			buf.DeleteRemaining()
+		case CharCtrlU:
+			buf.DeleteBefore()
+		case CharCtrlL:
+			buf.ClearScreen()
+		case CharCtrlW:
+			buf.DeleteWord()
+		case CharEnter:
+			output := buf.String()
+			if output != "" {
+				i.History.Add([]rune(output))
+			}
+			buf.MoveToEnd()
+			fmt.Println()
+			switch pasteMode {
+			case PasteModeStart:
+				output = `"""` + output
+			case PasteModeEnd:
+				output = output + `"""`
+			}
+			return output, nil
+		default:
+			if metaDel {
+				metaDel = false
+				continue
+			}
+			if r >= CharSpace || r == CharEnter {
+				buf.Add(r)
+			}
+		}
+	}
+}
+
+func (i *Instance) HistoryEnable() {
+	i.History.Enabled = true
+}
+
+func (i *Instance) HistoryDisable() {
+	i.History.Enabled = false
+}
+
+func NewTerminal() (*Terminal, error) {
+	t := &Terminal{
+		outchan: make(chan rune),
+	}
+
+	go t.ioloop()
+
+	return t, nil
+}
+
+func (t *Terminal) ioloop() {
+	buf := bufio.NewReader(os.Stdin)
+
+	for {
+		r, _, err := buf.ReadRune()
+		if err != nil {
+			close(t.outchan)
+			break
+		}
+		t.outchan <- r
+	}
+}
+
+func (t *Terminal) Read() (rune, error) {
+	r, ok := <-t.outchan
+	if !ok {
+		return 0, io.EOF
+	}
+
+	return r, nil
+}

readline/term.go

@@ -0,0 +1,36 @@
+//go:build aix || darwin || dragonfly || freebsd || (linux && !appengine) || netbsd || openbsd || os400 || solaris
+
+package readline
+
+import (
+	"syscall"
+)
+
+type Termios syscall.Termios
+
+func SetRawMode(fd int) (*Termios, error) {
+	termios, err := getTermios(fd)
+	if err != nil {
+		return nil, err
+	}
+
+	newTermios := *termios
+	newTermios.Iflag &^= syscall.IGNBRK | syscall.BRKINT | syscall.PARMRK | syscall.ISTRIP | syscall.INLCR | syscall.IGNCR | syscall.ICRNL | syscall.IXON
+	newTermios.Lflag &^= syscall.ECHO | syscall.ECHONL | syscall.ICANON | syscall.ISIG | syscall.IEXTEN
+	newTermios.Cflag &^= syscall.CSIZE | syscall.PARENB
+	newTermios.Cflag |= syscall.CS8
+	newTermios.Cc[syscall.VMIN] = 1
+	newTermios.Cc[syscall.VTIME] = 0
+
+	return termios, setTermios(fd, &newTermios)
+}
+
+func UnsetRawMode(fd int, termios *Termios) error {
+	return setTermios(fd, termios)
+}
+
+// IsTerminal returns true if the given file descriptor is a terminal.
+func IsTerminal(fd int) bool {
+	_, err := getTermios(fd)
+	return err == nil
+}

readline/term_bsd.go

@@ -0,0 +1,25 @@
+//go:build darwin || freebsd || netbsd || openbsd
+
+package readline
+
+import (
+	"syscall"
+	"unsafe"
+)
+
+func getTermios(fd int) (*Termios, error) {
+	termios := new(Termios)
+	_, _, err := syscall.Syscall6(syscall.SYS_IOCTL, uintptr(fd), syscall.TIOCGETA, uintptr(unsafe.Pointer(termios)), 0, 0, 0)
+	if err != 0 {
+		return nil, err
+	}
+	return termios, nil
+}
+
+func setTermios(fd int, termios *Termios) error {
+	_, _, err := syscall.Syscall6(syscall.SYS_IOCTL, uintptr(fd), syscall.TIOCSETA, uintptr(unsafe.Pointer(termios)), 0, 0, 0)
+	if err != 0 {
+		return err
+	}
+	return nil
+}

readline/term_linux.go

@@ -0,0 +1,28 @@
+//go:build linux || solaris
+
+package readline
+
+import (
+	"syscall"
+	"unsafe"
+)
+
+const tcgets = 0x5401
+const tcsets = 0x5402
+
+func getTermios(fd int) (*Termios, error) {
+	termios := new(Termios)
+	_, _, err := syscall.Syscall6(syscall.SYS_IOCTL, uintptr(fd), tcgets, uintptr(unsafe.Pointer(termios)), 0, 0, 0)
+	if err != 0 {
+		return nil, err
+	}
+	return termios, nil
+}
+
+func setTermios(fd int, termios *Termios) error {
+	_, _, err := syscall.Syscall6(syscall.SYS_IOCTL, uintptr(fd), tcsets, uintptr(unsafe.Pointer(termios)), 0, 0, 0)
+	if err != 0 {
+		return err
+	}
+	return nil
+}

readline/term_windows.go

@@ -0,0 +1,62 @@
+package readline
+
+import (
+	"syscall"
+	"unsafe"
+)
+
+const (
+	enableLineInput       = 2
+	enableWindowInput     = 8
+	enableMouseInput      = 16
+	enableInsertMode      = 32
+	enableQuickEditMode   = 64
+	enableExtendedFlags   = 128
+	enableProcessedOutput = 1
+	enableWrapAtEolOutput = 2
+	enableAutoPosition    = 256 // Cursor position is not affected by writing data to the console.
+	enableEchoInput       = 4   // Characters are written to the console as they're read.
+	enableProcessedInput  = 1   // Enables input processing (like recognizing Ctrl+C).
+)
+
+var kernel32 = syscall.NewLazyDLL("kernel32.dll")
+
+var (
+	procGetConsoleMode = kernel32.NewProc("GetConsoleMode")
+	procSetConsoleMode = kernel32.NewProc("SetConsoleMode")
+)
+
+type State struct {
+	mode uint32
+}
+
+// IsTerminal checks if the given file descriptor is associated with a terminal
+func IsTerminal(fd int) bool {
+	var st uint32
+	r, _, e := syscall.SyscallN(procGetConsoleMode.Addr(), uintptr(fd), uintptr(unsafe.Pointer(&st)), 0)
+	// if the call succeeds and doesn't produce an error, it's a terminal
+	return r != 0 && e == 0
+}
+
+func SetRawMode(fd int) (*State, error) {
+	var st uint32
+	// retrieve the current mode of the terminal
+	_, _, e := syscall.SyscallN(procGetConsoleMode.Addr(), uintptr(fd), uintptr(unsafe.Pointer(&st)), 0)
+	if e != 0 {
+		return nil, error(e)
+	}
+	// modify the mode to set it to raw
+	raw := st &^ (enableEchoInput | enableProcessedInput | enableLineInput | enableProcessedOutput)
+	// apply the new mode to the terminal
+	_, _, e = syscall.SyscallN(procSetConsoleMode.Addr(), uintptr(fd), uintptr(raw), 0)
+	if e != 0 {
+		return nil, error(e)
+	}
+	// return the original state so that it can be restored later
+	return &State{st}, nil
+}
+
+func UnsetRawMode(fd int, state *State) error {
+	_, _, err := syscall.SyscallN(procSetConsoleMode.Addr(), uintptr(fd), uintptr(state.mode), 0)
+	return err
+}

readline/types.go

@@ -0,0 +1,86 @@
+package readline
+
+const (
+	CharNull      = 0
+	CharLineStart = 1
+	CharBackward  = 2
+	CharInterrupt = 3
+	CharDelete    = 4
+	CharLineEnd   = 5
+	CharForward   = 6
+	CharBell      = 7
+	CharCtrlH     = 8
+	CharTab       = 9
+	CharCtrlJ     = 10
+	CharKill      = 11
+	CharCtrlL     = 12
+	CharEnter     = 13
+	CharNext      = 14
+	CharPrev      = 16
+	CharBckSearch = 18
+	CharFwdSearch = 19
+	CharTranspose = 20
+	CharCtrlU     = 21
+	CharCtrlW     = 23
+	CharCtrlY     = 25
+	CharCtrlZ     = 26
+	CharEsc       = 27
+	CharSpace     = 32
+	CharEscapeEx  = 91
+	CharBackspace = 127
+)
+
+const (
+	KeyDel    = 51
+	KeyUp     = 65
+	KeyDown   = 66
+	KeyRight  = 67
+	KeyLeft   = 68
+	MetaEnd   = 70
+	MetaStart = 72
+)
+
+const (
+	CursorUp    = "\033[1A"
+	CursorDown  = "\033[1B"
+	CursorRight = "\033[1C"
+	CursorLeft  = "\033[1D"
+
+	CursorSave    = "\033[s"
+	CursorRestore = "\033[u"
+
+	CursorUpN    = "\033[%dA"
+	CursorDownN  = "\033[%dB"
+	CursorRightN = "\033[%dC"
+	CursorLeftN  = "\033[%dD"
+
+	CursorEOL  = "\033[E"
+	CursorBOL  = "\033[1G"
+	CursorHide = "\033[?25l"
+	CursorShow = "\033[?25h"
+
+	ClearToEOL  = "\033[K"
+	ClearLine   = "\033[2K"
+	ClearScreen = "\033[2J"
+	CursorReset = "\033[0;0f"
+
+	ColorGrey    = "\033[38;5;245m"
+	ColorDefault = "\033[0m"
+
+	StartBracketedPaste = "\033[?2004h"
+	EndBracketedPaste   = "\033[?2004l"
+)
+
+const (
+	CharBracketedPaste      = 50
+	CharBracketedPasteStart = "00~"
+	CharBracketedPasteEnd   = "01~"
+)
+
+type PasteMode int
+
+const (
+	PastModeOff = iota
+	PasteModeStart
+	PasteModeEnd
+)

scripts/install.sh

@@ -63,7 +63,10 @@ status "Installing ollama to $BINDIR..."
 $SUDO install -o0 -g0 -m755 -d $BINDIR
 $SUDO install -o0 -g0 -m755 $TEMP_DIR/ollama $BINDIR/ollama
 
-install_success() { status 'Install complete. Run "ollama" from the command line.'; }
+install_success() { 
+    status 'The Ollama API is now available at 0.0.0.0:11434.'
+    status 'Install complete. Run "ollama" from the command line.'
+}
 trap install_success EXIT
 
 # Everything from this point onwards is optional.
@@ -74,6 +77,9 @@ configure_systemd() {
         $SUDO useradd -r -s /bin/false -m -d /usr/share/ollama ollama
     fi
 
+    status "Adding current user to ollama group..."
+    $SUDO usermod -a -G ollama $(whoami)
+
     status "Creating ollama systemd service..."
     cat <<EOF | $SUDO tee /etc/systemd/system/ollama.service >/dev/null
 [Unit]
@@ -86,7 +92,6 @@ User=ollama
 Group=ollama
 Restart=always
 RestartSec=3
-Environment="HOME=/usr/share/ollama"
 Environment="PATH=$PATH"
 
 [Install]
@@ -128,6 +133,7 @@ if check_gpu nvidia-smi; then
 fi
 
 if ! check_gpu lspci && ! check_gpu lshw; then
+    install_success
     warning "No NVIDIA GPU detected. Ollama will run in CPU-only mode."
     exit 0
 fi

server/auth.go

@@ -91,7 +91,7 @@ func getAuthToken(ctx context.Context, redirData AuthRedirect) (string, error) {
 	}
 
 	s := SignatureData{
-		Method: "GET",
+		Method: http.MethodGet,
 		Path:   redirectURL.String(),
 		Data:   nil,
 	}
@@ -103,7 +103,7 @@ func getAuthToken(ctx context.Context, redirData AuthRedirect) (string, error) {
 
 	headers := make(http.Header)
 	headers.Set("Authorization", sig)
-	resp, err := makeRequest(ctx, "GET", redirectURL, headers, nil, nil)
+	resp, err := makeRequest(ctx, http.MethodGet, redirectURL, headers, nil, nil)
 	if err != nil {
 		log.Printf("couldn't get token: %q", err)
 		return "", err

server/download.go

@@ -15,6 +15,7 @@ import (
 	"strings"
 	"sync"
 	"sync/atomic"
+	"syscall"
 	"time"
 
 	"golang.org/x/sync/errgroup"
@@ -88,17 +89,12 @@ func (b *blobDownload) Prepare(ctx context.Context, requestURL *url.URL, opts *R
 	}
 
 	if len(b.Parts) == 0 {
-		resp, err := makeRequest(ctx, "HEAD", requestURL, nil, nil, opts)
+		resp, err := makeRequestWithRetry(ctx, http.MethodHead, requestURL, nil, nil, opts)
 		if err != nil {
 			return err
 		}
 		defer resp.Body.Close()
 
-		if resp.StatusCode >= http.StatusBadRequest {
-			body, _ := io.ReadAll(resp.Body)
-			return fmt.Errorf("registry responded with code %d: %v", resp.StatusCode, string(body))
-		}
-
 		b.Total, _ = strconv.ParseInt(resp.Header.Get("Content-Length"), 10, 64)
 
 		var size = b.Total / numDownloadParts
@@ -133,7 +129,6 @@ func (b *blobDownload) Run(ctx context.Context, requestURL *url.URL, opts *Regis
 
 func (b *blobDownload) run(ctx context.Context, requestURL *url.URL, opts *RegistryOptions) error {
 	defer blobDownloadManager.Delete(b.Digest)
-
 	ctx, b.CancelFunc = context.WithCancel(ctx)
 
 	file, err := os.OpenFile(b.Name+"-partial", os.O_CREATE|os.O_RDWR, 0644)
@@ -158,7 +153,8 @@ func (b *blobDownload) run(ctx context.Context, requestURL *url.URL, opts *Regis
 				w := io.NewOffsetWriter(file, part.StartsAt())
 				err := b.downloadChunk(inner, requestURL, w, part, opts)
 				switch {
-				case errors.Is(err, context.Canceled):
+				case errors.Is(err, context.Canceled), errors.Is(err, syscall.ENOSPC):
+					// return immediately if the context is canceled or the device is out of space
 					return err
 				case err != nil:
 					log.Printf("%s part %d attempt %d failed: %v, retrying", b.Digest[7:19], i, try, err)
@@ -168,7 +164,7 @@ func (b *blobDownload) run(ctx context.Context, requestURL *url.URL, opts *Regis
 				}
 			}
 
-			return errors.New("max retries exceeded")
+			return errMaxRetriesExceeded
 		})
 	}
 
@@ -198,7 +194,7 @@ func (b *blobDownload) run(ctx context.Context, requestURL *url.URL, opts *Regis
 func (b *blobDownload) downloadChunk(ctx context.Context, requestURL *url.URL, w io.Writer, part *blobDownloadPart, opts *RegistryOptions) error {
 	headers := make(http.Header)
 	headers.Set("Range", fmt.Sprintf("bytes=%d-%d", part.StartsAt(), part.StopsAt()-1))
-	resp, err := makeRequest(ctx, "GET", requestURL, headers, nil, opts)
+	resp, err := makeRequestWithRetry(ctx, http.MethodGet, requestURL, headers, nil, opts)
 	if err != nil {
 		return err
 	}
@@ -306,6 +302,8 @@ type downloadOpts struct {
 
 const maxRetries = 3
 
+var errMaxRetriesExceeded = errors.New("max retries exceeded")
+
 // downloadBlob downloads a blob from the registry and stores it in the blobs directory
 func downloadBlob(ctx context.Context, opts downloadOpts) error {
 	fp, err := GetBlobsPath(opts.digest)

server/images.go

@@ -63,15 +63,11 @@ func (m *Model) Prompt(request api.GenerateRequest) (string, error) {
 		First  bool
 		System string
 		Prompt string
-
-		// deprecated: versions <= 0.0.7 used this to omit the system prompt
-		Context []int
 	}
 
 	vars.First = len(request.Context) == 0
 	vars.System = m.System
 	vars.Prompt = request.Prompt
-	vars.Context = request.Context
 
 	if request.System != "" {
 		vars.System = request.System
@@ -131,7 +127,7 @@ func (m *ManifestV2) GetTotalSize() (total int64) {
 }
 
 func GetManifest(mp ModelPath) (*ManifestV2, string, error) {
-	fp, err := mp.GetManifestPath(false)
+	fp, err := mp.GetManifestPath()
 	if err != nil {
 		return nil, "", err
 	}
@@ -595,10 +591,13 @@ func CreateManifest(name string, cfg *LayerReader, layers []*Layer) error {
 		return err
 	}
 
-	fp, err := mp.GetManifestPath(true)
+	fp, err := mp.GetManifestPath()
 	if err != nil {
 		return err
 	}
+	if err := os.MkdirAll(filepath.Dir(fp), 0o755); err != nil {
+		return err
+	}
 	return os.WriteFile(fp, manifestJSON, 0o644)
 }
 
@@ -710,16 +709,19 @@ func CreateLayer(f io.ReadSeeker) (*LayerReader, error) {
 
 func CopyModel(src, dest string) error {
 	srcModelPath := ParseModelPath(src)
-	srcPath, err := srcModelPath.GetManifestPath(false)
+	srcPath, err := srcModelPath.GetManifestPath()
 	if err != nil {
 		return err
 	}
 
 	destModelPath := ParseModelPath(dest)
-	destPath, err := destModelPath.GetManifestPath(true)
+	destPath, err := destModelPath.GetManifestPath()
 	if err != nil {
 		return err
 	}
+	if err := os.MkdirAll(filepath.Dir(destPath), 0o755); err != nil {
+		return err
+	}
 
 	// copy the file
 	input, err := os.ReadFile(srcPath)
@@ -882,7 +884,7 @@ func DeleteModel(name string) error {
 		return err
 	}
 
-	fp, err := mp.GetManifestPath(false)
+	fp, err := mp.GetManifestPath()
 	if err != nil {
 		return err
 	}
@@ -975,46 +977,7 @@ func PushModel(ctx context.Context, name string, regOpts *RegistryOptions, fn fu
 	layers = append(layers, &manifest.Config)
 
 	for _, layer := range layers {
-		exists, err := checkBlobExistence(ctx, mp, layer.Digest, regOpts)
-		if err != nil {
-			return err
-		}
-
-		if exists {
-			fn(api.ProgressResponse{
-				Status:    "using existing layer",
-				Digest:    layer.Digest,
-				Total:     layer.Size,
-				Completed: layer.Size,
-			})
-			log.Printf("Layer %s already exists", layer.Digest)
-			continue
-		}
-
-		fn(api.ProgressResponse{
-			Status: "starting upload",
-			Digest: layer.Digest,
-			Total:  layer.Size,
-		})
-
-		location, chunkSize, err := startUpload(ctx, mp, layer, regOpts)
-		if err != nil {
-			log.Printf("couldn't start upload: %v", err)
-			return err
-		}
-
-		if strings.HasPrefix(filepath.Base(location.Path), "sha256:") {
-			layer.Digest = filepath.Base(location.Path)
-			fn(api.ProgressResponse{
-				Status:    "using existing layer",
-				Digest:    layer.Digest,
-				Total:     layer.Size,
-				Completed: layer.Size,
-			})
-			continue
-		}
-
-		if err := uploadBlob(ctx, location, layer, chunkSize, regOpts, fn); err != nil {
+		if err := uploadBlob(ctx, mp, layer, regOpts, fn); err != nil {
 			log.Printf("error uploading blob: %v", err)
 			return err
 		}
@@ -1031,7 +994,7 @@ func PushModel(ctx context.Context, name string, regOpts *RegistryOptions, fn fu
 
 	headers := make(http.Header)
 	headers.Set("Content-Type", "application/vnd.docker.distribution.manifest.v2+json")
-	resp, err := makeRequestWithRetry(ctx, "PUT", requestURL, headers, bytes.NewReader(manifestJSON), regOpts)
+	resp, err := makeRequestWithRetry(ctx, http.MethodPut, requestURL, headers, bytes.NewReader(manifestJSON), regOpts)
 	if err != nil {
 		return err
 	}
@@ -1121,10 +1084,13 @@ func PullModel(ctx context.Context, name string, regOpts *RegistryOptions, fn fu
 		return err
 	}
 
-	fp, err := mp.GetManifestPath(true)
+	fp, err := mp.GetManifestPath()
 	if err != nil {
 		return err
 	}
+	if err := os.MkdirAll(filepath.Dir(fp), 0o755); err != nil {
+		return err
+	}
 
 	err = os.WriteFile(fp, manifestJSON, 0o644)
 	if err != nil {
@@ -1150,22 +1116,12 @@ func pullModelManifest(ctx context.Context, mp ModelPath, regOpts *RegistryOptio
 
 	headers := make(http.Header)
 	headers.Set("Accept", "application/vnd.docker.distribution.manifest.v2+json")
-	resp, err := makeRequest(ctx, "GET", requestURL, headers, nil, regOpts)
+	resp, err := makeRequestWithRetry(ctx, http.MethodGet, requestURL, headers, nil, regOpts)
 	if err != nil {
-		log.Printf("couldn't get manifest: %v", err)
 		return nil, err
 	}
 	defer resp.Body.Close()
 
-	if resp.StatusCode >= http.StatusBadRequest {
-		if resp.StatusCode == http.StatusNotFound {
-			return nil, fmt.Errorf("model not found")
-		}
-
-		body, _ := io.ReadAll(resp.Body)
-		return nil, fmt.Errorf("on pull registry responded with code %d: %s", resp.StatusCode, body)
-	}
-
 	var m *ManifestV2
 	if err := json.NewDecoder(resp.Body).Decode(&m); err != nil {
 		return nil, err
@@ -1209,24 +1165,7 @@ func GetSHA256Digest(r io.Reader) (string, int64) {
 	return fmt.Sprintf("sha256:%x", h.Sum(nil)), n
 }
 
-// Function to check if a blob already exists in the Docker registry
-func checkBlobExistence(ctx context.Context, mp ModelPath, digest string, regOpts *RegistryOptions) (bool, error) {
-	requestURL := mp.BaseURL()
-	requestURL = requestURL.JoinPath("v2", mp.GetNamespaceRepository(), "blobs", digest)
-
-	resp, err := makeRequest(ctx, "HEAD", requestURL, nil, nil, regOpts)
-	if err != nil {
-		log.Printf("couldn't check for blob: %v", err)
-		return false, err
-	}
-	defer resp.Body.Close()
-
-	// Check for success: If the blob exists, the Docker registry will respond with a 200 OK
-	return resp.StatusCode < http.StatusBadRequest, nil
-}
-
 func makeRequestWithRetry(ctx context.Context, method string, requestURL *url.URL, headers http.Header, body io.ReadSeeker, regOpts *RegistryOptions) (*http.Response, error) {
-	var status string
 	for try := 0; try < maxRetries; try++ {
 		resp, err := makeRequest(ctx, method, requestURL, headers, body, regOpts)
 		if err != nil {
@@ -1234,8 +1173,6 @@ func makeRequestWithRetry(ctx context.Context, method string, requestURL *url.UR
 			return nil, err
 		}
 
-		status = resp.Status
-
 		switch {
 		case resp.StatusCode == http.StatusUnauthorized:
 			auth := resp.Header.Get("www-authenticate")
@@ -1247,21 +1184,25 @@ func makeRequestWithRetry(ctx context.Context, method string, requestURL *url.UR
 
 			regOpts.Token = token
 			if body != nil {
-				if _, err := body.Seek(0, io.SeekStart); err != nil {
-					return nil, err
-				}
+				body.Seek(0, io.SeekStart)
 			}
 
 			continue
+		case resp.StatusCode == http.StatusNotFound:
+			return nil, os.ErrNotExist
 		case resp.StatusCode >= http.StatusBadRequest:
-			body, _ := io.ReadAll(resp.Body)
-			return nil, fmt.Errorf("on upload registry responded with code %d: %s", resp.StatusCode, body)
+			body, err := io.ReadAll(resp.Body)
+			if err != nil {
+				return nil, fmt.Errorf("%d: %s", resp.StatusCode, err)
+			}
+
+			return nil, fmt.Errorf("%d: %s", resp.StatusCode, body)
 		default:
 			return resp, nil
 		}
 	}
 
-	return nil, fmt.Errorf("max retry exceeded: %v", status)
+	return nil, errMaxRetriesExceeded
 }
 
 func makeRequest(ctx context.Context, method string, requestURL *url.URL, headers http.Header, body io.Reader, regOpts *RegistryOptions) (*http.Response, error) {

server/modelpath.go

@@ -85,20 +85,27 @@ func (mp ModelPath) GetShortTagname() string {
 	return fmt.Sprintf("%s/%s/%s:%s", mp.Registry, mp.Namespace, mp.Repository, mp.Tag)
 }
 
-func (mp ModelPath) GetManifestPath(createDir bool) (string, error) {
+// modelsDir returns the value of the OLLAMA_MODELS environment variable or the user's home directory if OLLAMA_MODELS is not set.
+// The models directory is where Ollama stores its model files and manifests.
+func modelsDir() (string, error) {
+	if models, exists := os.LookupEnv("OLLAMA_MODELS"); exists {
+		return models, nil
+	}
 	home, err := os.UserHomeDir()
 	if err != nil {
 		return "", err
 	}
+	return filepath.Join(home, ".ollama", "models"), nil
+}
 
-	path := filepath.Join(home, ".ollama", "models", "manifests", mp.Registry, mp.Namespace, mp.Repository, mp.Tag)
-	if createDir {
-		if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
-			return "", err
-		}
+// GetManifestPath returns the path to the manifest file for the given model path, it is up to the caller to create the directory if it does not exist.
+func (mp ModelPath) GetManifestPath() (string, error) {
+	dir, err := modelsDir()
+	if err != nil {
+		return "", err
 	}
 
-	return path, nil
+	return filepath.Join(dir, "manifests", mp.Registry, mp.Namespace, mp.Repository, mp.Tag), nil
 }
 
 func (mp ModelPath) BaseURL() *url.URL {
@@ -109,12 +116,12 @@ func (mp ModelPath) BaseURL() *url.URL {
 }
 
 func GetManifestPath() (string, error) {
-	home, err := os.UserHomeDir()
+	dir, err := modelsDir()
 	if err != nil {
 		return "", err
 	}
 
-	path := filepath.Join(home, ".ollama", "models", "manifests")
+	path := filepath.Join(dir, "manifests")
 	if err := os.MkdirAll(path, 0o755); err != nil {
 		return "", err
 	}
@@ -123,7 +130,7 @@ func GetManifestPath() (string, error) {
 }
 
 func GetBlobsPath(digest string) (string, error) {
-	home, err := os.UserHomeDir()
+	dir, err := modelsDir()
 	if err != nil {
 		return "", err
 	}
@@ -132,7 +139,7 @@ func GetBlobsPath(digest string) (string, error) {
 		digest = strings.ReplaceAll(digest, ":", "-")
 	}
 
-	path := filepath.Join(home, ".ollama", "models", "blobs", digest)
+	path := filepath.Join(dir, "blobs", digest)
 	dirPath := filepath.Dir(path)
 	if digest == "" {
 		dirPath = path

server/routes.go

@@ -365,7 +365,9 @@ func PushModelHandler(c *gin.Context) {
 			Insecure: req.Insecure,
 		}
 
-		ctx := context.Background()
+		ctx, cancel := context.WithCancel(c.Request.Context())
+		defer cancel()
+
 		if err := PushModel(ctx, req.Name, regOpts, fn); err != nil {
 			ch <- gin.H{"error": err.Error()}
 		}
@@ -614,6 +616,22 @@ var defaultAllowOrigins = []string{
 }
 
 func Serve(ln net.Listener, allowOrigins []string) error {
+	if noprune := os.Getenv("OLLAMA_NOPRUNE"); noprune == "" {
+		// clean up unused layers and manifests
+		if err := PruneLayers(); err != nil {
+			return err
+		}
+
+		manifestsPath, err := GetManifestPath()
+		if err != nil {
+			return err
+		}
+
+		if err := PruneDirectory(manifestsPath); err != nil {
+			return err
+		}
+	}
+
 	config := cors.DefaultConfig()
 	config.AllowWildcard = true
 
@@ -679,7 +697,7 @@ func Serve(ln net.Listener, allowOrigins []string) error {
 	if runtime.GOOS == "linux" {
 		// check compatibility to log warnings
 		if _, err := llm.CheckVRAM(); err != nil {
-			log.Printf("Warning: GPU support may not enabled, check you have installed install GPU drivers: %v", err)
+			log.Printf("Warning: GPU support may not be enabled, check you have installed GPU drivers: %v", err)
 		}
 	}
 

server/upload.go

@@ -2,218 +2,367 @@ package server
 
 import (
 	"context"
+	"crypto/md5"
 	"errors"
 	"fmt"
+	"hash"
 	"io"
 	"log"
 	"net/http"
 	"net/url"
 	"os"
-	"strconv"
+	"strings"
 	"sync"
+	"sync/atomic"
+	"time"
 
 	"github.com/jmorganca/ollama/api"
+	"github.com/jmorganca/ollama/format"
+	"golang.org/x/sync/errgroup"
 )
 
+var blobUploadManager sync.Map
+
+type blobUpload struct {
+	*Layer
+
+	Total     int64
+	Completed atomic.Int64
+
+	Parts []blobUploadPart
+
+	nextURL chan *url.URL
+
+	context.CancelFunc
+
+	done       bool
+	err        error
+	references atomic.Int32
+}
+
+type blobUploadPart struct {
+	// N is the part number
+	N      int
+	Offset int64
+	Size   int64
+	hash.Hash
+}
+
 const (
-	redirectChunkSize int64 = 1024 * 1024 * 1024
-	regularChunkSize  int64 = 95 * 1024 * 1024
+	numUploadParts          = 64
+	minUploadPartSize int64 = 95 * 1000 * 1000
+	maxUploadPartSize int64 = 1000 * 1000 * 1000
 )
 
-func startUpload(ctx context.Context, mp ModelPath, layer *Layer, regOpts *RegistryOptions) (*url.URL, int64, error) {
-	requestURL := mp.BaseURL()
-	requestURL = requestURL.JoinPath("v2", mp.GetNamespaceRepository(), "blobs/uploads/")
-	if layer.From != "" {
+func (b *blobUpload) Prepare(ctx context.Context, requestURL *url.URL, opts *RegistryOptions) error {
+	p, err := GetBlobsPath(b.Digest)
+	if err != nil {
+		return err
+	}
+
+	if b.From != "" {
 		values := requestURL.Query()
-		values.Add("mount", layer.Digest)
-		values.Add("from", layer.From)
+		values.Add("mount", b.Digest)
+		values.Add("from", b.From)
 		requestURL.RawQuery = values.Encode()
 	}
 
-	resp, err := makeRequestWithRetry(ctx, "POST", requestURL, nil, nil, regOpts)
+	resp, err := makeRequestWithRetry(ctx, http.MethodPost, requestURL, nil, nil, opts)
 	if err != nil {
-		log.Printf("couldn't start upload: %v", err)
-		return nil, 0, err
+		return err
 	}
 	defer resp.Body.Close()
 
 	location := resp.Header.Get("Docker-Upload-Location")
-	chunkSize := redirectChunkSize
 	if location == "" {
 		location = resp.Header.Get("Location")
-		chunkSize = regularChunkSize
 	}
 
-	locationURL, err := url.Parse(location)
+	fi, err := os.Stat(p)
 	if err != nil {
-		return nil, 0, err
+		return err
 	}
 
-	return locationURL, chunkSize, nil
+	b.Total = fi.Size()
+
+	var size = b.Total / numUploadParts
+	switch {
+	case size < minUploadPartSize:
+		size = minUploadPartSize
+	case size > maxUploadPartSize:
+		size = maxUploadPartSize
+	}
+
+	var offset int64
+	for offset < fi.Size() {
+		if offset+size > fi.Size() {
+			size = fi.Size() - offset
+		}
+
+		// set part.N to the current number of parts
+		b.Parts = append(b.Parts, blobUploadPart{N: len(b.Parts), Offset: offset, Size: size, Hash: md5.New()})
+		offset += size
+	}
+
+	log.Printf("uploading %s in %d %s part(s)", b.Digest[7:19], len(b.Parts), format.HumanBytes(b.Parts[0].Size))
+
+	requestURL, err = url.Parse(location)
+	if err != nil {
+		return err
+	}
+
+	b.nextURL = make(chan *url.URL, 1)
+	b.nextURL <- requestURL
+	return nil
 }
 
-func uploadBlob(ctx context.Context, requestURL *url.URL, layer *Layer, chunkSize int64, regOpts *RegistryOptions, fn func(api.ProgressResponse)) error {
-	// TODO allow resumability
-	// TODO allow canceling uploads via DELETE
+// Run uploads blob parts to the upstream. If the upstream supports redirection, parts will be uploaded
+// in parallel as defined by Prepare. Otherwise, parts will be uploaded serially. Run sets b.err on error.
+func (b *blobUpload) Run(ctx context.Context, opts *RegistryOptions) {
+	defer blobUploadManager.Delete(b.Digest)
+	ctx, b.CancelFunc = context.WithCancel(ctx)
 
-	fp, err := GetBlobsPath(layer.Digest)
+	p, err := GetBlobsPath(b.Digest)
 	if err != nil {
-		return err
+		b.err = err
+		return
 	}
 
-	f, err := os.Open(fp)
+	f, err := os.Open(p)
 	if err != nil {
-		return err
+		b.err = err
+		return
 	}
 	defer f.Close()
 
-	pw := ProgressWriter{
-		status: fmt.Sprintf("uploading %s", layer.Digest),
-		digest: layer.Digest,
-		total:  layer.Size,
-		fn:     fn,
-	}
+	g, inner := errgroup.WithContext(ctx)
+	g.SetLimit(numUploadParts)
+	for i := range b.Parts {
+		part := &b.Parts[i]
+		select {
+		case <-inner.Done():
+		case requestURL := <-b.nextURL:
+			g.Go(func() error {
+				for try := 0; try < maxRetries; try++ {
+					r := io.NewSectionReader(f, part.Offset, part.Size)
+					err := b.uploadChunk(inner, http.MethodPatch, requestURL, r, part, opts)
+					switch {
+					case errors.Is(err, context.Canceled):
+						return err
+					case errors.Is(err, errMaxRetriesExceeded):
+						return err
+					case err != nil:
+						log.Printf("%s part %d attempt %d failed: %v, retrying", b.Digest[7:19], part.N, try, err)
+						continue
+					}
 
-	for offset := int64(0); offset < layer.Size; {
-		chunk := layer.Size - offset
-		if chunk > chunkSize {
-			chunk = chunkSize
-		}
+					return nil
+				}
 
-		resp, err := uploadBlobChunk(ctx, http.MethodPatch, requestURL, f, offset, chunk, regOpts, &pw)
-		if err != nil {
-			fn(api.ProgressResponse{
-				Status:    fmt.Sprintf("error uploading chunk: %v", err),
-				Digest:    layer.Digest,
-				Total:     layer.Size,
-				Completed: offset,
+				return errMaxRetriesExceeded
 			})
-
-			return err
-		}
-
-		offset += chunk
-		location := resp.Header.Get("Docker-Upload-Location")
-		if location == "" {
-			location = resp.Header.Get("Location")
-		}
-
-		requestURL, err = url.Parse(location)
-		if err != nil {
-			return err
 		}
 	}
 
+	if err := g.Wait(); err != nil {
+		b.err = err
+		return
+	}
+
+	requestURL := <-b.nextURL
+
+	var sb strings.Builder
+	for _, part := range b.Parts {
+		sb.Write(part.Sum(nil))
+	}
+
+	md5sum := md5.Sum([]byte(sb.String()))
+
 	values := requestURL.Query()
-	values.Add("digest", layer.Digest)
+	values.Add("digest", b.Digest)
+	values.Add("etag", fmt.Sprintf("%x-%d", md5sum, len(b.Parts)))
 	requestURL.RawQuery = values.Encode()
 
 	headers := make(http.Header)
 	headers.Set("Content-Type", "application/octet-stream")
 	headers.Set("Content-Length", "0")
 
-	// finish the upload
-	resp, err := makeRequest(ctx, "PUT", requestURL, headers, nil, regOpts)
+	resp, err := makeRequestWithRetry(ctx, http.MethodPut, requestURL, headers, nil, opts)
+	if err != nil {
+		b.err = err
+		return
+	}
+	defer resp.Body.Close()
+
+	b.done = true
+}
+
+func (b *blobUpload) uploadChunk(ctx context.Context, method string, requestURL *url.URL, rs io.ReadSeeker, part *blobUploadPart, opts *RegistryOptions) error {
+	headers := make(http.Header)
+	headers.Set("Content-Type", "application/octet-stream")
+	headers.Set("Content-Length", fmt.Sprintf("%d", part.Size))
+	headers.Set("X-Redirect-Uploads", "1")
+
+	if method == http.MethodPatch {
+		headers.Set("Content-Range", fmt.Sprintf("%d-%d", part.Offset, part.Offset+part.Size-1))
+	}
+
+	buw := blobUploadWriter{blobUpload: b}
+	resp, err := makeRequest(ctx, method, requestURL, headers, io.TeeReader(rs, io.MultiWriter(&buw, part.Hash)), opts)
 	if err != nil {
-		log.Printf("couldn't finish upload: %v", err)
 		return err
 	}
 	defer resp.Body.Close()
 
-	if resp.StatusCode >= http.StatusBadRequest {
-		body, _ := io.ReadAll(resp.Body)
-		return fmt.Errorf("on finish upload registry responded with code %d: %v", resp.StatusCode, string(body))
-	}
-	return nil
-}
-
-func uploadBlobChunk(ctx context.Context, method string, requestURL *url.URL, r io.ReaderAt, offset, limit int64, opts *RegistryOptions, pw *ProgressWriter) (*http.Response, error) {
-	sectionReader := io.NewSectionReader(r, offset, limit)
-
-	headers := make(http.Header)
-	headers.Set("Content-Type", "application/octet-stream")
-	headers.Set("Content-Length", strconv.Itoa(int(limit)))
-	headers.Set("X-Redirect-Uploads", "1")
-
-	if method == http.MethodPatch {
-		headers.Set("Content-Range", fmt.Sprintf("%d-%d", offset, offset+sectionReader.Size()-1))
+	location := resp.Header.Get("Docker-Upload-Location")
+	if location == "" {
+		location = resp.Header.Get("Location")
 	}
 
-	for try := 0; try < maxRetries; try++ {
-		resp, err := makeRequest(ctx, method, requestURL, headers, io.TeeReader(sectionReader, pw), opts)
-		if err != nil && !errors.Is(err, io.EOF) {
-			return nil, err
+	nextURL, err := url.Parse(location)
+	if err != nil {
+		return err
+	}
+
+	switch {
+	case resp.StatusCode == http.StatusTemporaryRedirect:
+		b.nextURL <- nextURL
+
+		redirectURL, err := resp.Location()
+		if err != nil {
+			return err
 		}
-		defer resp.Body.Close()
 
-		switch {
-		case resp.StatusCode == http.StatusTemporaryRedirect:
-			location, err := resp.Location()
-			if err != nil {
-				return nil, err
-			}
-
-			pw.completed = offset
-			if _, err := uploadBlobChunk(ctx, http.MethodPut, location, r, offset, limit, nil, pw); err != nil {
-				// retry
-				log.Printf("retrying redirected upload: %v", err)
+		for try := 0; try < maxRetries; try++ {
+			rs.Seek(0, io.SeekStart)
+			b.Completed.Add(-buw.written)
+			buw.written = 0
+			part.Hash = md5.New()
+			err := b.uploadChunk(ctx, http.MethodPut, redirectURL, rs, part, nil)
+			switch {
+			case errors.Is(err, context.Canceled):
+				return err
+			case errors.Is(err, errMaxRetriesExceeded):
+				return err
+			case err != nil:
+				log.Printf("%s part %d attempt %d failed: %v, retrying", b.Digest[7:19], part.N, try, err)
 				continue
 			}
 
-			return resp, nil
-		case resp.StatusCode == http.StatusUnauthorized:
-			auth := resp.Header.Get("www-authenticate")
-			authRedir := ParseAuthRedirectString(auth)
-			token, err := getAuthToken(ctx, authRedir)
-			if err != nil {
-				return nil, err
-			}
-
-			opts.Token = token
-
-			pw.completed = offset
-			sectionReader = io.NewSectionReader(r, offset, limit)
-			continue
-		case resp.StatusCode >= http.StatusBadRequest:
-			body, _ := io.ReadAll(resp.Body)
-			return nil, fmt.Errorf("on upload registry responded with code %d: %s", resp.StatusCode, body)
+			return nil
 		}
 
-		return resp, nil
+		return errMaxRetriesExceeded
+
+	case resp.StatusCode == http.StatusUnauthorized:
+		auth := resp.Header.Get("www-authenticate")
+		authRedir := ParseAuthRedirectString(auth)
+		token, err := getAuthToken(ctx, authRedir)
+		if err != nil {
+			return err
+		}
+
+		opts.Token = token
+		fallthrough
+	case resp.StatusCode >= http.StatusBadRequest:
+		body, err := io.ReadAll(resp.Body)
+		if err != nil {
+			return err
+		}
+
+		rs.Seek(0, io.SeekStart)
+		b.Completed.Add(-buw.written)
+		buw.written = 0
+		return fmt.Errorf("http status %d %s: %s", resp.StatusCode, resp.Status, body)
 	}
 
-	return nil, fmt.Errorf("max retries exceeded")
+	if method == http.MethodPatch {
+		b.nextURL <- nextURL
+	}
+
+	return nil
 }
 
-type ProgressWriter struct {
-	status    string
-	digest    string
-	bucket    int64
-	completed int64
-	total     int64
-	fn        func(api.ProgressResponse)
-	mu        sync.Mutex
+func (b *blobUpload) acquire() {
+	b.references.Add(1)
 }
 
-func (pw *ProgressWriter) Write(b []byte) (int, error) {
-	pw.mu.Lock()
-	defer pw.mu.Unlock()
+func (b *blobUpload) release() {
+	if b.references.Add(-1) == 0 {
+		b.CancelFunc()
+	}
+}
 
-	n := len(b)
-	pw.bucket += int64(n)
+func (b *blobUpload) Wait(ctx context.Context, fn func(api.ProgressResponse)) error {
+	b.acquire()
+	defer b.release()
 
-	// throttle status updates to not spam the client
-	if pw.bucket >= 1024*1024 || pw.completed+pw.bucket >= pw.total {
-		pw.completed += pw.bucket
-		pw.fn(api.ProgressResponse{
-			Status:    pw.status,
-			Digest:    pw.digest,
-			Total:     pw.total,
-			Completed: pw.completed,
+	ticker := time.NewTicker(60 * time.Millisecond)
+	for {
+		select {
+		case <-ticker.C:
+		case <-ctx.Done():
+			return ctx.Err()
+		}
+
+		fn(api.ProgressResponse{
+			Status:    fmt.Sprintf("uploading %s", b.Digest),
+			Digest:    b.Digest,
+			Total:     b.Total,
+			Completed: b.Completed.Load(),
 		})
 
-		pw.bucket = 0
+		if b.done || b.err != nil {
+			return b.err
+		}
 	}
+}
 
+type blobUploadWriter struct {
+	written int64
+	*blobUpload
+}
+
+func (b *blobUploadWriter) Write(p []byte) (n int, err error) {
+	n = len(p)
+	b.written += int64(n)
+	b.Completed.Add(int64(n))
 	return n, nil
 }
+
+func uploadBlob(ctx context.Context, mp ModelPath, layer *Layer, opts *RegistryOptions, fn func(api.ProgressResponse)) error {
+	requestURL := mp.BaseURL()
+	requestURL = requestURL.JoinPath("v2", mp.GetNamespaceRepository(), "blobs", layer.Digest)
+
+	resp, err := makeRequestWithRetry(ctx, http.MethodHead, requestURL, nil, nil, opts)
+	switch {
+	case errors.Is(err, os.ErrNotExist):
+	case err != nil:
+		return err
+	default:
+		defer resp.Body.Close()
+		fn(api.ProgressResponse{
+			Status:    fmt.Sprintf("uploading %s", layer.Digest),
+			Digest:    layer.Digest,
+			Total:     layer.Size,
+			Completed: layer.Size,
+		})
+
+		return nil
+	}
+
+	data, ok := blobUploadManager.LoadOrStore(layer.Digest, &blobUpload{Layer: layer})
+	upload := data.(*blobUpload)
+	if !ok {
+		requestURL := mp.BaseURL()
+		requestURL = requestURL.JoinPath("v2", mp.GetNamespaceRepository(), "blobs/uploads/")
+		if err := upload.Prepare(ctx, requestURL, opts); err != nil {
+			blobUploadManager.Delete(layer.Digest)
+			return err
+		}
+
+		go upload.Run(context.Background(), opts)
+	}
+
+	return upload.Wait(ctx, fn)
+}