Merge pull request #884 from jmorganca/mxyng/update-submodules

bump submodules

Dockerfile

@@ -5,8 +5,8 @@ ARG GOFLAGS="'-ldflags=-w -s'"
 
 WORKDIR /go/src/github.com/jmorganca/ollama
 RUN apt-get update && apt-get install -y git build-essential cmake
-ADD https://dl.google.com/go/go1.21.1.linux-$TARGETARCH.tar.gz /tmp/go1.21.1.tar.gz
-RUN mkdir -p /usr/local && tar xz -C /usr/local </tmp/go1.21.1.tar.gz
+ADD https://dl.google.com/go/go1.21.3.linux-$TARGETARCH.tar.gz /tmp/go1.21.3.tar.gz
+RUN mkdir -p /usr/local && tar xz -C /usr/local </tmp/go1.21.3.tar.gz
 
 COPY . .
 ENV GOARCH=$TARGETARCH

Dockerfile.build

@@ -1,6 +1,5 @@
-
 # centos7 amd64 dependencies
-FROM --platform=linux/amd64 nvidia/cuda:11.8.0-devel-centos7 AS base-amd64
+FROM --platform=linux/amd64 nvidia/cuda:11.3.1-devel-centos7 AS base-amd64
 RUN yum install -y https://repo.ius.io/ius-release-el7.rpm centos-release-scl && \
     yum update -y && \
     yum install -y devtoolset-10-gcc devtoolset-10-gcc-c++ git236 wget
@@ -8,7 +7,7 @@ RUN wget "https://github.com/Kitware/CMake/releases/download/v3.27.6/cmake-3.27.
 ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
 
 # centos8 arm64 dependencies
-FROM --platform=linux/arm64 nvidia/cuda:11.4.3-devel-centos8 AS base-arm64
+FROM --platform=linux/arm64 nvidia/cuda-arm64:11.3.1-devel-centos8 AS base-arm64
 RUN sed -i -e 's/mirrorlist/#mirrorlist/g' -e 's|#baseurl=http://mirror.centos.org|baseurl=http://vault.centos.org|g' /etc/yum.repos.d/CentOS-*
 RUN yum install -y git cmake
 
@@ -17,8 +16,8 @@ ARG TARGETARCH
 ARG GOFLAGS="'-ldflags -w -s'"
 
 # install go
-ADD https://dl.google.com/go/go1.21.1.linux-$TARGETARCH.tar.gz /tmp/go1.21.1.tar.gz
-RUN mkdir -p /usr/local && tar xz -C /usr/local </tmp/go1.21.1.tar.gz
+ADD https://dl.google.com/go/go1.21.3.linux-$TARGETARCH.tar.gz /tmp/go1.21.3.tar.gz
+RUN mkdir -p /usr/local && tar xz -C /usr/local </tmp/go1.21.3.tar.gz
 
 # build the final binary
 WORKDIR /go/src/github.com/jmorganca/ollama

README.md

@@ -15,6 +15,10 @@ Get up and running with large language models locally.
 
 [Download](https://ollama.ai/download/Ollama-darwin.zip)
 
+### Windows
+
+Coming soon!
+
 ### Linux & WSL2
 
 ```
@@ -23,9 +27,9 @@ curl https://ollama.ai/install.sh | sh
 
 [Manual install instructions](https://github.com/jmorganca/ollama/blob/main/docs/linux.md)
 
-### Windows
+### Docker
 
-coming soon
+See the official [Docker image](https://hub.docker.com/r/ollama/ollama).
 
 ## Quickstart
 
@@ -56,11 +60,11 @@ Here are some example open-source models that can be downloaded:
 
 ## Customize your own model
 
-### Import from GGUF or GGML
+### Import from GGUF
 
-Ollama supports importing GGUF and GGML file formats in the Modelfile. This means if you have a model that is not in the Ollama library, you can create it, iterate on it, and upload it to the Ollama library to share with others when you are ready.
+Ollama supports importing GGUF models in the Modelfile:
 
-1. Create a file named Modelfile, and add a `FROM` instruction with the local filepath to the model you want to import.
+1. Create a file named `Modelfile`, with a `FROM` instruction with the local filepath to the model you want to import.
 
    ```
    FROM ./vicuna-33b.Q4_0.gguf
@@ -69,15 +73,19 @@ Ollama supports importing GGUF and GGML file formats in the Modelfile. This mean
 2. Create the model in Ollama
 
    ```
-   ollama create name -f path_to_modelfile
+   ollama create example -f Modelfile
    ```
 
 3. Run the model
 
    ```
-   ollama run name
+   ollama run example
    ```
 
+### Import from PyTorch or Safetensors
+
+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
@@ -195,7 +203,7 @@ Finally, in a separate shell, run a model:
 
 ## REST API
 
-> See the [API documentation](docs/api.md) for all endpoints.
+See the [API documentation](docs/api.md) for all endpoints.
 
 Ollama has an API for running and managing models. For example to generate text from a model:
 
@@ -222,3 +230,7 @@ curl -X POST http://localhost:11434/api/generate -d '{
 - [Typescript UI](https://github.com/ollama-interface/Ollama-Gui?tab=readme-ov-file)
 - [Dumbar](https://github.com/JerrySievert/Dumbar)
 - [Emacs client](https://github.com/zweifisch/ollama)
+- [oterm](https://github.com/ggozad/oterm)
+- [Ellama Emacs client](https://github.com/s-kostyaev/ellama)
+- [OllamaSharp for .NET](https://github.com/awaescher/OllamaSharp)
+- [Minimalistic React UI for Ollama Models](https://github.com/richawo/minimal-llm-ui)

api/client.go

@@ -14,6 +14,7 @@ import (
 	"runtime"
 	"strings"
 
+	"github.com/jmorganca/ollama/format"
 	"github.com/jmorganca/ollama/version"
 )
 
@@ -51,8 +52,10 @@ func ClientFromEnvironment() (*Client, error) {
 	host, port, err := net.SplitHostPort(hostport)
 	if err != nil {
 		host, port = "127.0.0.1", "11434"
-		if ip := net.ParseIP(strings.Trim(os.Getenv("OLLAMA_HOST"), "[]")); ip != nil {
+		if ip := net.ParseIP(strings.Trim(hostport, "[]")); ip != nil {
 			host = ip.String()
+		} else if hostport != "" {
+			host = hostport
 		}
 	}
 
@@ -127,7 +130,7 @@ func (c *Client) do(ctx context.Context, method, path string, reqData, respData
 	return nil
 }
 
-const maxBufferSize = 512 * 1000 // 512KB
+const maxBufferSize = 512 * format.KiloByte
 
 func (c *Client) stream(ctx context.Context, method, path string, data any, fn func([]byte) error) error {
 	var buf *bytes.Buffer

api/types.go

@@ -3,7 +3,6 @@ package api
 import (
 	"encoding/json"
 	"fmt"
-	"log"
 	"math"
 	"os"
 	"reflect"
@@ -162,15 +161,10 @@ func (r *GenerateResponse) Summary() {
 	}
 }
 
-type Options struct {
-	Seed int `json:"seed,omitempty"`
-
-	// Backend options
-	UseNUMA bool `json:"numa,omitempty"`
-
-	// Model options
+// Runner options which must be set when the model is loaded into memory
+type Runner struct {
+	UseNUMA            bool    `json:"numa,omitempty"`
 	NumCtx             int     `json:"num_ctx,omitempty"`
-	NumKeep            int     `json:"num_keep,omitempty"`
 	NumBatch           int     `json:"num_batch,omitempty"`
 	NumGQA             int     `json:"num_gqa,omitempty"`
 	NumGPU             int     `json:"num_gpu,omitempty"`
@@ -184,8 +178,15 @@ type Options struct {
 	EmbeddingOnly      bool    `json:"embedding_only,omitempty"`
 	RopeFrequencyBase  float32 `json:"rope_frequency_base,omitempty"`
 	RopeFrequencyScale float32 `json:"rope_frequency_scale,omitempty"`
+	NumThread          int     `json:"num_thread,omitempty"`
+}
 
-	// Predict options
+type Options struct {
+	Runner
+
+	// Predict options used at runtime
+	NumKeep          int      `json:"num_keep,omitempty"`
+	Seed             int      `json:"seed,omitempty"`
 	NumPredict       int      `json:"num_predict,omitempty"`
 	TopK             int      `json:"top_k,omitempty"`
 	TopP             float32  `json:"top_p,omitempty"`
@@ -201,8 +202,6 @@ type Options struct {
 	MirostatEta      float32  `json:"mirostat_eta,omitempty"`
 	PenalizeNewline  bool     `json:"penalize_newline,omitempty"`
 	Stop             []string `json:"stop,omitempty"`
-
-	NumThread int `json:"num_thread,omitempty"`
 }
 
 var ErrInvalidOpts = fmt.Errorf("invalid options")
@@ -238,44 +237,39 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
 						// when JSON unmarshals numbers, it uses float64, not int
 						field.SetInt(int64(t))
 					default:
-						log.Printf("could not convert model parameter %v of type %T to int, skipped", key, val)
+						return fmt.Errorf("option %q must be of type integer", key)
 					}
 				case reflect.Bool:
 					val, ok := val.(bool)
 					if !ok {
-						log.Printf("could not convert model parameter %v of type %T to bool, skipped", key, val)
-						continue
+						return fmt.Errorf("option %q must be of type boolean", key)
 					}
 					field.SetBool(val)
 				case reflect.Float32:
 					// JSON unmarshals to float64
 					val, ok := val.(float64)
 					if !ok {
-						log.Printf("could not convert model parameter %v of type %T to float32, skipped", key, val)
-						continue
+						return fmt.Errorf("option %q must be of type float32", key)
 					}
 					field.SetFloat(val)
 				case reflect.String:
 					val, ok := val.(string)
 					if !ok {
-						log.Printf("could not convert model parameter %v of type %T to string, skipped", key, val)
-						continue
+						return fmt.Errorf("option %q must be of type string", key)
 					}
 					field.SetString(val)
 				case reflect.Slice:
 					// JSON unmarshals to []interface{}, not []string
 					val, ok := val.([]interface{})
 					if !ok {
-						log.Printf("could not convert model parameter %v of type %T to slice, skipped", key, val)
-						continue
+						return fmt.Errorf("option %q must be of type array", key)
 					}
 					// convert []interface{} to []string
 					slice := make([]string, len(val))
 					for i, item := range val {
 						str, ok := item.(string)
 						if !ok {
-							log.Printf("could not convert model parameter %v of type %T to slice of strings, skipped", key, item)
-							continue
+							return fmt.Errorf("option %q must be of an array of strings", key)
 						}
 						slice[i] = str
 					}
@@ -315,20 +309,22 @@ func DefaultOptions() Options {
 		PenalizeNewline:  true,
 		Seed:             -1,
 
-		// options set when the model is loaded
-		NumCtx:             2048,
-		RopeFrequencyBase:  10000.0,
-		RopeFrequencyScale: 1.0,
-		NumBatch:           512,
-		NumGPU:             -1, // -1 here indicates that NumGPU should be set dynamically
-		NumGQA:             1,
-		NumThread:          0, // let the runtime decide
-		LowVRAM:            false,
-		F16KV:              true,
-		UseMLock:           false,
-		UseMMap:            true,
-		UseNUMA:            false,
-		EmbeddingOnly:      true,
+		Runner: Runner{
+			// options set when the model is loaded
+			NumCtx:             2048,
+			RopeFrequencyBase:  10000.0,
+			RopeFrequencyScale: 1.0,
+			NumBatch:           512,
+			NumGPU:             -1, // -1 here indicates that NumGPU should be set dynamically
+			NumGQA:             1,
+			NumThread:          0, // let the runtime decide
+			LowVRAM:            false,
+			F16KV:              true,
+			UseMLock:           false,
+			UseMMap:            true,
+			UseNUMA:            false,
+			EmbeddingOnly:      true,
+		},
 	}
 }
 

app/forge.config.ts

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

app/package.json

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

app/src/index.ts

@@ -162,13 +162,56 @@ app.on('before-quit', () => {
   }
 })
 
+const updateURL = `https://ollama.ai/api/update?os=${process.platform}&arch=${
+  process.arch
+}&version=${app.getVersion()}&id=${id()}`
+
+let latest = ''
+async function isNewReleaseAvailable() {
+  try {
+    const response = await fetch(updateURL)
+
+    if (!response.ok) {
+      return false
+    }
+
+    if (response.status === 204) {
+      return false
+    }
+
+    const data = await response.json()
+
+    const url = data?.url
+    if (!url) {
+      return false
+    }
+
+    if (latest === url) {
+      return false
+    }
+
+    latest = url
+
+    return true
+  } catch (error) {
+    logger.error(`update check failed - ${error}`)
+    return false
+  }
+}
+
+async function checkUpdate() {
+  const available = await isNewReleaseAvailable()
+  if (available) {
+    logger.info('checking for update')
+    autoUpdater.checkForUpdates()
+  }
+}
+
 function init() {
   if (app.isPackaged) {
-    autoUpdater.checkForUpdates()
+    checkUpdate()
     setInterval(() => {
-      if (!updateAvailable) {
-        autoUpdater.checkForUpdates()
-      }
+      checkUpdate()
     }, 60 * 60 * 1000)
   }
 
@@ -246,11 +289,7 @@ function id(): string {
   return uuid
 }
 
-autoUpdater.setFeedURL({
-  url: `https://ollama.ai/api/update?os=${process.platform}&arch=${
-    process.arch
-  }&version=${app.getVersion()}&id=${id()}`,
-})
+autoUpdater.setFeedURL({ url: updateURL })
 
 autoUpdater.on('error', e => {
   logger.error(`update check failed - ${e.message}`)

cmd/cmd.go

@@ -78,18 +78,12 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 				spinner.Stop()
 			}
 			currentDigest = resp.Digest
-			switch {
-			case strings.Contains(resp.Status, "embeddings"):
-				bar = progressbar.Default(resp.Total, resp.Status)
-				bar.Set64(resp.Completed)
-			default:
-				// pulling
-				bar = progressbar.DefaultBytes(
-					resp.Total,
-					resp.Status,
-				)
-				bar.Set64(resp.Completed)
-			}
+			// pulling
+			bar = progressbar.DefaultBytes(
+				resp.Total,
+				resp.Status,
+			)
+			bar.Set64(resp.Completed)
 		} else if resp.Digest == currentDigest && resp.Digest != "" {
 			bar.Set64(resp.Completed)
 		} else {
@@ -694,7 +688,12 @@ func generateInteractive(cmd *cobra.Command, model string) error {
 		case strings.HasPrefix(line, "/show"):
 			args := strings.Fields(line)
 			if len(args) > 1 {
-				resp, err := server.GetModelInfo(model)
+				client, err := api.ClientFromEnvironment()
+				if err != nil {
+					fmt.Println("error: couldn't connect to ollama server")
+					return err
+				}
+				resp, err := client.Show(cmd.Context(), &api.ShowRequest{Name: model})
 				if err != nil {
 					fmt.Println("error: couldn't get model")
 					return err
@@ -933,7 +932,7 @@ func NewCLI() *cobra.Command {
 	createCmd := &cobra.Command{
 		Use:     "create MODEL",
 		Short:   "Create a model from a Modelfile",
-		Args:    cobra.MinimumNArgs(1),
+		Args:    cobra.ExactArgs(1),
 		PreRunE: checkServerHeartbeat,
 		RunE:    CreateHandler,
 	}
@@ -943,7 +942,7 @@ func NewCLI() *cobra.Command {
 	showCmd := &cobra.Command{
 		Use:     "show MODEL",
 		Short:   "Show information for a model",
-		Args:    cobra.MinimumNArgs(1),
+		Args:    cobra.ExactArgs(1),
 		PreRunE: checkServerHeartbeat,
 		RunE:    ShowHandler,
 	}
@@ -970,13 +969,14 @@ func NewCLI() *cobra.Command {
 		Use:     "serve",
 		Aliases: []string{"start"},
 		Short:   "Start ollama",
+		Args:    cobra.ExactArgs(0),
 		RunE:    RunServer,
 	}
 
 	pullCmd := &cobra.Command{
 		Use:     "pull MODEL",
 		Short:   "Pull a model from a registry",
-		Args:    cobra.MinimumNArgs(1),
+		Args:    cobra.ExactArgs(1),
 		PreRunE: checkServerHeartbeat,
 		RunE:    PullHandler,
 	}
@@ -986,7 +986,7 @@ func NewCLI() *cobra.Command {
 	pushCmd := &cobra.Command{
 		Use:     "push MODEL",
 		Short:   "Push a model to a registry",
-		Args:    cobra.MinimumNArgs(1),
+		Args:    cobra.ExactArgs(1),
 		PreRunE: checkServerHeartbeat,
 		RunE:    PushHandler,
 	}
@@ -1002,15 +1002,15 @@ func NewCLI() *cobra.Command {
 	}
 
 	copyCmd := &cobra.Command{
-		Use:     "cp",
+		Use:     "cp SOURCE TARGET",
 		Short:   "Copy a model",
-		Args:    cobra.MinimumNArgs(2),
+		Args:    cobra.ExactArgs(2),
 		PreRunE: checkServerHeartbeat,
 		RunE:    CopyHandler,
 	}
 
 	deleteCmd := &cobra.Command{
-		Use:     "rm",
+		Use:     "rm MODEL [MODEL...]",
 		Short:   "Remove a model",
 		Args:    cobra.MinimumNArgs(1),
 		PreRunE: checkServerHeartbeat,

docs/api.md

@@ -355,7 +355,7 @@ curl -X POST http://localhost:11434/api/embeddings -d '{
 
 ```json
 {
-  "embeddings": [
+  "embedding": [
     0.5670403838157654, 0.009260174818336964, 0.23178744316101074, -0.2916173040866852, -0.8924556970596313,
     0.8785552978515625, -0.34576427936553955, 0.5742510557174683, -0.04222835972905159, -0.137906014919281
   ]

docs/faq.md

@@ -1,5 +1,21 @@
 # FAQ
 
+## How can I view the logs?
+
+On macOS:
+
+```
+cat ~/.ollama/logs/server.log
+```
+
+On Linux:
+
+```
+journalctl -u ollama
+```
+
+If you're running `ollama serve` directly, the logs will be printed to the console.
+
 ## How can I expose the Ollama server?
 
 ```bash
@@ -14,5 +30,5 @@ OLLAMA_ORIGINS=http://192.168.1.1:*,https://example.com ollama serve
 
 ## Where are models stored?
 
-* macOS: Raw model data is stored under `~/.ollama/models`.
-* Linux: Raw model data is stored under `/usr/share/ollama/.ollama/models`
+- macOS: Raw model data is stored under `~/.ollama/models`.
+- Linux: Raw model data is stored under `/usr/share/ollama/.ollama/models`

docs/import.md

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

docs/linux.md

@@ -80,4 +80,3 @@ To view logs of Ollama running as a startup service, run:
 ```bash
 journalctl -u ollama
 ```
-

docs/modelfile.md

@@ -12,7 +12,6 @@ A model file is the blueprint to create and share models with Ollama.
   - [FROM (Required)](#from-required)
     - [Build from llama2](#build-from-llama2)
     - [Build from a bin file](#build-from-a-bin-file)
-  - [EMBED](#embed)
   - [PARAMETER](#parameter)
     - [Valid Parameters and Values](#valid-parameters-and-values)
   - [TEMPLATE](#template)
@@ -91,17 +90,6 @@ FROM ./ollama-model.bin
 
 This bin file location should be specified as an absolute path or relative to the `Modelfile` location.
 
-### EMBED
-
-The `EMBED` instruction is used to add embeddings of files to a model. This is useful for adding custom data that the model can reference when generating an answer. Note that currently only text files are supported, formatted with each line as one embedding.
-
-```modelfile
-FROM <model name>:<tag>
-EMBED <file path>.txt
-EMBED <different file path>.txt
-EMBED <path to directory>/*.txt
-```
-
 ### PARAMETER
 
 The `PARAMETER` instruction defines a parameter that can be set when the model is run.
@@ -124,7 +112,7 @@ PARAMETER <parameter> <parametervalue>
 | repeat_last_n  | Sets how far back for the model to look back to prevent repetition. (Default: 64, 0 = disabled, -1 = num_ctx)                                                                                                                                           | int        | repeat_last_n 64     |
 | repeat_penalty | Sets how strongly to penalize repetitions. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. (Default: 1.1)                                                                     | float      | repeat_penalty 1.1   |
 | temperature    | The temperature of the model. Increasing the temperature will make the model answer more creatively. (Default: 0.8)                                                                                                                                     | float      | temperature 0.7      |
-| seed | Sets the random number seed to use for generation. Setting this to a specific number will make the model generate the same text for the same prompt. | int | seed 42 |
+| seed | Sets the random number seed to use for generation. Setting this to a specific number will make the model generate the same text for the same prompt. (Default: 0) | int | seed 42 |
 | stop           | Sets the stop sequences to use.                                                                                                                                                                                                                         | string     | stop "AI assistant:" |
 | tfs_z          | Tail free sampling is used to reduce the impact of less probable tokens from the output. A higher value (e.g., 2.0) will reduce the impact more, while a value of 1.0 disables this setting. (default: 1)                                               | float      | tfs_z 1              |
 | num_predict    | Maximum number of tokens to predict when generating text. (Default: 128, -1 = infinite generation, -2 = fill context)                                                                                                                                   | int        | num_predict 42       |

docs/quantize.md

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

examples/golang-simplegenerate/main.go

@@ -3,10 +3,10 @@ package main
 import (
 	"bytes"
 	"fmt"
-	"net/http"
-	"os"
 	"io"
 	"log"
+	"net/http"
+	"os"
 )
 
 func main() {
@@ -16,7 +16,7 @@ func main() {
 	if err != nil {
 		fmt.Print(err.Error())
 		os.Exit(1)
-	} 
+	}
 
 	responseData, err := io.ReadAll(resp.Body)
 	if err != nil {

examples/python-rag-newssummary/README.md

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

examples/python-rag-newssummary/requirements.txt

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

examples/python-rag-newssummary/summ.py

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

examples/python-rag-newssummary/utils.py

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

format/bytes.go

@@ -2,14 +2,21 @@ package format
 
 import "fmt"
 
+const (
+	Byte     = 1
+	KiloByte = Byte * 1000
+	MegaByte = KiloByte * 1000
+	GigaByte = MegaByte * 1000
+)
+
 func HumanBytes(b int64) string {
 	switch {
-	case b > 1000*1000*1000:
-		return fmt.Sprintf("%d GB", b/1000/1000/1000)
-	case b > 1000*1000:
-		return fmt.Sprintf("%d MB", b/1000/1000)
-	case b > 1000:
-		return fmt.Sprintf("%d KB", b/1000)
+	case b > GigaByte:
+		return fmt.Sprintf("%d GB", b/GigaByte)
+	case b > MegaByte:
+		return fmt.Sprintf("%d MB", b/MegaByte)
+	case b > KiloByte:
+		return fmt.Sprintf("%d KB", b/KiloByte)
 	default:
 		return fmt.Sprintf("%d B", b)
 	}

format/time_test.go

@@ -29,7 +29,7 @@ func TestHumanTime(t *testing.T) {
 	})
 
 	t.Run("soon", func(t *testing.T) {
-		v := now.Add(800*time.Millisecond)
+		v := now.Add(800 * time.Millisecond)
 		assertEqual(t, HumanTime(v, ""), "Less than a second from now")
 	})
 }

go.mod

@@ -45,7 +45,6 @@ require (
 	golang.org/x/sys v0.11.0 // indirect
 	golang.org/x/term v0.10.0
 	golang.org/x/text v0.10.0 // indirect
-	gonum.org/v1/gonum v0.13.0
 	google.golang.org/protobuf v1.30.0 // indirect
 	gopkg.in/yaml.v3 v3.0.1 // indirect
 )

go.sum

@@ -145,8 +145,6 @@ golang.org/x/text v0.10.0 h1:UpjohKhiEgNc0CSauXmwYftY1+LlaC75SJwh0SgCX58=
 golang.org/x/text v0.10.0/go.mod h1:TvPlkZtksWOMsz7fbANvkp4WM8x/WCo/om8BMLbz+aE=
 golang.org/x/tools v0.0.0-20180917221912-90fa682c2a6e/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
 golang.org/x/xerrors v0.0.0-20191204190536-9bdfabe68543/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
-gonum.org/v1/gonum v0.13.0 h1:a0T3bh+7fhRyqeNbiC3qVHYmkiQgit3wnNan/2c0HMM=
-gonum.org/v1/gonum v0.13.0/go.mod h1:/WPYRckkfWrhWefxyYTfrTtQR0KH4iyHNuzxqXAKyAU=
 google.golang.org/protobuf v1.26.0-rc.1/go.mod h1:jlhhOSvTdKEhbULTjvd4ARK9grFBp09yW+WbY/TyQbw=
 google.golang.org/protobuf v1.28.0/go.mod h1:HV8QOd/L58Z+nl8r43ehVNZIU/HEI6OcFqwMG9pJV4I=
 google.golang.org/protobuf v1.30.0 h1:kPPoIgf3TsEvrm0PFe15JQ+570QVxYzEvvHqChK+cng=

llm/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 +12,7 @@ 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 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 +12,7 @@ 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 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

@@ -13,7 +13,7 @@ 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 git -C gguf apply ../patches/0001-update-default-log-target.patch
 //go:generate cmake -S gguf -B gguf/build/cpu -DLLAMA_K_QUANTS=on
 //go:generate cmake --build gguf/build/cpu --target server --config Release
 //go:generate mv gguf/build/cpu/bin/server gguf/build/cpu/bin/ollama-runner

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 git -C gguf apply ../patches/0001-update-default-log-target.patch
 //go:generate cmake -S gguf -B gguf/build/cpu -DLLAMA_K_QUANTS=on
 //go:generate cmake --build gguf/build/cpu --target server --config Release
 //go:generate cmd /c move gguf\build\cpu\bin\Release\server.exe gguf\build\cpu\bin\Release\ollama-runner.exe

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

@@ -1,25 +0,0 @@
-From 07993bdc35345b67b27aa649a7c099ad42d80c4c Mon Sep 17 00:00:00 2001
-From: Michael Yang <mxyng@pm.me>
-Date: Thu, 21 Sep 2023 14:43:21 -0700
-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 2597ba0..b56549b 100644
---- a/common/common.cpp
-+++ b/common/common.cpp
-@@ -780,8 +780,6 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
-     }
- 
-     {
--        LOG("warming up the model with an empty run\n");
--
-         const std::vector<llama_token> tmp = { llama_token_bos(lctx), llama_token_eos(lctx), };
-         llama_eval(lctx, tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, params.n_threads);
-         llama_reset_timings(lctx);
--- 
-2.42.0
-

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

@@ -24,51 +24,53 @@ import (
 	"time"
 
 	"github.com/jmorganca/ollama/api"
+	"github.com/jmorganca/ollama/format"
 )
 
 //go:embed llama.cpp/*/build/*/bin/*
 var llamaCppEmbed embed.FS
 
 type ModelRunner struct {
-	Path string // path to the model runner executable
+	Path        string // path to the model runner executable
+	Accelerated bool
 }
 
 func chooseRunners(workDir, runnerType string) []ModelRunner {
 	buildPath := path.Join("llama.cpp", runnerType, "build")
-	var runners []string
+	var runners []ModelRunner
 
 	// set the runners based on the OS
 	// IMPORTANT: the order of the runners in the array is the priority order
 	switch runtime.GOOS {
 	case "darwin":
-		runners = []string{
-			path.Join(buildPath, "metal", "bin", "ollama-runner"),
-			path.Join(buildPath, "cpu", "bin", "ollama-runner"),
+		runners = []ModelRunner{
+			{Path: path.Join(buildPath, "metal", "bin", "ollama-runner")},
+			{Path: path.Join(buildPath, "cpu", "bin", "ollama-runner")},
 		}
 	case "linux":
-		runners = []string{
-			path.Join(buildPath, "cuda", "bin", "ollama-runner"),
-			path.Join(buildPath, "cpu", "bin", "ollama-runner"),
+		runners = []ModelRunner{
+			{Path: path.Join(buildPath, "cuda", "bin", "ollama-runner"), Accelerated: true},
+			{Path: path.Join(buildPath, "cpu", "bin", "ollama-runner")},
 		}
 	case "windows":
 		// TODO: select windows GPU runner here when available
-		runners = []string{
-			path.Join(buildPath, "cpu", "bin", "Release", "ollama-runner.exe"),
+		runners = []ModelRunner{
+			{Path: path.Join(buildPath, "cpu", "bin", "Release", "ollama-runner.exe")},
 		}
 	default:
 		log.Printf("unknown OS, running on CPU: %s", runtime.GOOS)
-		runners = []string{
-			path.Join(buildPath, "cpu", "bin", "ollama-runner"),
+		runners = []ModelRunner{
+			{Path: path.Join(buildPath, "cpu", "bin", "ollama-runner")},
 		}
 	}
 
 	runnerAvailable := false // if no runner files are found in the embed, this flag will cause a fast fail
 	for _, r := range runners {
 		// find all the files in the runner's bin directory
-		files, err := fs.Glob(llamaCppEmbed, path.Join(path.Dir(r), "*"))
+		files, err := fs.Glob(llamaCppEmbed, path.Join(path.Dir(r.Path), "*"))
 		if err != nil {
 			// this is expected, ollama may be compiled without all runners packed in
-			log.Printf("%s runner not found: %v", r, err)
+			log.Printf("%s runner not found: %v", r.Path, err)
 			continue
 		}
 
@@ -115,7 +117,10 @@ func chooseRunners(workDir, runnerType string) []ModelRunner {
 	localRunnersByPriority := []ModelRunner{}
 	for _, r := range runners {
 		// clean the ModelRunner paths so that they match the OS we are running on
-		localRunnersByPriority = append(localRunnersByPriority, ModelRunner{Path: filepath.Clean(path.Join(workDir, r))})
+		localRunnersByPriority = append(localRunnersByPriority, ModelRunner{
+			Path:        filepath.Clean(path.Join(workDir, r.Path)),
+			Accelerated: r.Accelerated,
+		})
 	}
 
 	return localRunnersByPriority
@@ -178,12 +183,12 @@ type llamaHyperparameters struct {
 }
 
 type Running struct {
-	Port     int
-	Cmd      *exec.Cmd
-	Cancel   context.CancelFunc
-	exitOnce sync.Once
-	exitCh   chan error // channel to receive the exit status of the subprocess
-	exitErr  error      // error returned by the subprocess
+	Port          int
+	Cmd           *exec.Cmd
+	Cancel        context.CancelFunc
+	exitOnce      sync.Once
+	exitCh        chan error // channel to receive the exit status of the subprocess
+	*StatusWriter            // captures error messages from the llama runner process
 }
 
 type llama struct {
@@ -193,7 +198,7 @@ type llama struct {
 
 var errNoGPU = errors.New("nvidia-smi command failed")
 
-// CheckVRAM returns the available VRAM in MiB on Linux machines with NVIDIA GPUs
+// CheckVRAM returns the free VRAM in bytes on Linux machines with NVIDIA GPUs
 func CheckVRAM() (int64, error) {
 	cmd := exec.Command("nvidia-smi", "--query-gpu=memory.free", "--format=csv,noheader,nounits")
 	var stdout bytes.Buffer
@@ -203,7 +208,7 @@ func CheckVRAM() (int64, error) {
 		return 0, errNoGPU
 	}
 
-	var free int64
+	var freeMiB int64
 	scanner := bufio.NewScanner(&stdout)
 	for scanner.Scan() {
 		line := scanner.Text()
@@ -212,10 +217,16 @@ func CheckVRAM() (int64, error) {
 			return 0, fmt.Errorf("failed to parse available VRAM: %v", err)
 		}
 
-		free += vram
+		freeMiB += vram
 	}
 
-	return free, nil
+	freeBytes := freeMiB * 1024 * 1024
+	if freeBytes < 2*format.GigaByte {
+		log.Printf("less than 2 GB VRAM available, falling back to CPU only")
+		freeMiB = 0
+	}
+
+	return freeBytes, nil
 }
 
 func NumGPU(numLayer, fileSizeBytes int64, opts api.Options) int {
@@ -223,7 +234,7 @@ func NumGPU(numLayer, fileSizeBytes int64, opts api.Options) int {
 		return opts.NumGPU
 	}
 	if runtime.GOOS == "linux" {
-		vramMib, err := CheckVRAM()
+		freeBytes, err := CheckVRAM()
 		if err != nil {
 			if err.Error() != "nvidia-smi command failed" {
 				log.Print(err.Error())
@@ -232,15 +243,13 @@ func NumGPU(numLayer, fileSizeBytes int64, opts api.Options) int {
 			return 0
 		}
 
-		freeVramBytes := int64(vramMib) * 1024 * 1024 // 1 MiB = 1024^2 bytes
-
 		// Calculate bytes per layer
 		// TODO: this is a rough heuristic, better would be to calculate this based on number of layers and context size
 		bytesPerLayer := fileSizeBytes / numLayer
 
-		// max number of layers we can fit in VRAM, subtract 5% to prevent consuming all available VRAM and running out of memory
-		layers := int(freeVramBytes/bytesPerLayer) * 95 / 100
-		log.Printf("%d MiB VRAM available, loading up to %d GPU layers", vramMib, layers)
+		// max number of layers we can fit in VRAM, subtract 8% to prevent consuming all available VRAM and running out of memory
+		layers := int(freeBytes/bytesPerLayer) * 92 / 100
+		log.Printf("%d MB VRAM available, loading up to %d GPU layers", freeBytes/(1024*1024), layers)
 
 		return layers
 	}
@@ -250,7 +259,8 @@ func NumGPU(numLayer, fileSizeBytes int64, opts api.Options) int {
 
 // StatusWriter is a writer that captures error messages from the llama runner process
 type StatusWriter struct {
-	ErrCh chan error
+	ErrCh      chan error
+	LastErrMsg string
 }
 
 func NewStatusWriter() *StatusWriter {
@@ -260,10 +270,18 @@ func NewStatusWriter() *StatusWriter {
 }
 
 func (w *StatusWriter) Write(b []byte) (int, error) {
+	var errMsg string
 	if _, after, ok := bytes.Cut(b, []byte("error:")); ok {
-		err := fmt.Errorf("llama runner: %s", after)
-		w.ErrCh <- err
+		errMsg = string(bytes.TrimSpace(after))
+	} else if _, after, ok := bytes.Cut(b, []byte("CUDA error")); ok {
+		errMsg = string(bytes.TrimSpace(after))
 	}
+
+	if errMsg != "" {
+		w.LastErrMsg = errMsg
+		w.ErrCh <- fmt.Errorf("llama runner: %s", errMsg)
+	}
+
 	return os.Stderr.Write(b)
 }
 
@@ -277,13 +295,14 @@ func newLlama(model string, adapters []string, runners []ModelRunner, numLayers
 		return nil, errors.New("ollama supports only one lora adapter, but multiple were provided")
 	}
 
+	numGPU := NumGPU(numLayers, fileInfo.Size(), opts)
 	params := []string{
 		"--model", model,
 		"--ctx-size", fmt.Sprintf("%d", opts.NumCtx),
 		"--rope-freq-base", fmt.Sprintf("%f", opts.RopeFrequencyBase),
 		"--rope-freq-scale", fmt.Sprintf("%f", opts.RopeFrequencyScale),
 		"--batch-size", fmt.Sprintf("%d", opts.NumBatch),
-		"--n-gpu-layers", fmt.Sprintf("%d", NumGPU(numLayers, fileInfo.Size(), opts)),
+		"--n-gpu-layers", fmt.Sprintf("%d", numGPU),
 		"--embedding",
 	}
 
@@ -317,6 +336,11 @@ func newLlama(model string, adapters []string, runners []ModelRunner, numLayers
 
 	// start the llama.cpp server with a retry in case the port is already in use
 	for _, runner := range runners {
+		if runner.Accelerated && numGPU == 0 {
+			log.Printf("skipping accelerated runner because num_gpu=0")
+			continue
+		}
+
 		if _, err := os.Stat(runner.Path); err != nil {
 			log.Printf("llama runner not found: %v", err)
 			continue
@@ -345,7 +369,13 @@ func newLlama(model string, adapters []string, runners []ModelRunner, numLayers
 		// monitor the llama runner process and signal when it exits
 		go func() {
 			err := llm.Cmd.Wait()
-			llm.exitErr = err
+			// default to printing the exit message of the command process, it will probably just say 'exit staus 1'
+			errMsg := err.Error()
+			// try to set a better error message if llama runner logs captured an error
+			if statusWriter.LastErrMsg != "" {
+				errMsg = statusWriter.LastErrMsg
+			}
+			log.Println(errMsg)
 			// llm.Cmd.Wait() can only be called once, use this exit channel to signal that the process has exited
 			llm.exitOnce.Do(func() {
 				close(llm.exitCh)
@@ -415,10 +445,9 @@ func (llm *llama) Close() {
 
 	// wait for the command to exit to prevent race conditions with the next run
 	<-llm.exitCh
-	err := llm.exitErr
 
-	if err != nil {
-		log.Printf("llama runner stopped with error: %v", err)
+	if llm.StatusWriter != nil && llm.StatusWriter.LastErrMsg != "" {
+		log.Printf("llama runner stopped with error: %v", llm.StatusWriter.LastErrMsg)
 	} else {
 		log.Print("llama runner stopped successfully")
 	}
@@ -428,71 +457,21 @@ func (llm *llama) SetOptions(opts api.Options) {
 	llm.Options = opts
 }
 
-type GenerationSettings struct {
-	FrequencyPenalty float64       `json:"frequency_penalty"`
-	IgnoreEOS        bool          `json:"ignore_eos"`
-	LogitBias        []interface{} `json:"logit_bias"`
-	Mirostat         int           `json:"mirostat"`
-	MirostatEta      float64       `json:"mirostat_eta"`
-	MirostatTau      float64       `json:"mirostat_tau"`
-	Model            string        `json:"model"`
-	NCtx             int           `json:"n_ctx"`
-	NKeep            int           `json:"n_keep"`
-	NPredict         int           `json:"n_predict"`
-	NProbs           int           `json:"n_probs"`
-	PenalizeNl       bool          `json:"penalize_nl"`
-	PresencePenalty  float64       `json:"presence_penalty"`
-	RepeatLastN      int           `json:"repeat_last_n"`
-	RepeatPenalty    float64       `json:"repeat_penalty"`
-	Seed             uint32        `json:"seed"`
-	Stop             []string      `json:"stop"`
-	Stream           bool          `json:"stream"`
-	Temp             float64       `json:"temp"`
-	TfsZ             float64       `json:"tfs_z"`
-	TopK             int           `json:"top_k"`
-	TopP             float64       `json:"top_p"`
-	TypicalP         float64       `json:"typical_p"`
-}
-
-type Timings struct {
-	PredictedN  int     `json:"predicted_n"`
-	PredictedMS float64 `json:"predicted_ms"`
-	PromptN     int     `json:"prompt_n"`
-	PromptMS    float64 `json:"prompt_ms"`
-}
-
-type Prediction struct {
+type prediction struct {
 	Content string `json:"content"`
 	Model   string `json:"model"`
 	Prompt  string `json:"prompt"`
 	Stop    bool   `json:"stop"`
 
-	Timings `json:"timings"`
+	Timings struct {
+		PredictedN  int     `json:"predicted_n"`
+		PredictedMS float64 `json:"predicted_ms"`
+		PromptN     int     `json:"prompt_n"`
+		PromptMS    float64 `json:"prompt_ms"`
+	}
 }
 
-type PredictRequest struct {
-	Prompt           string   `json:"prompt"`
-	Stream           bool     `json:"stream"`
-	NPredict         int      `json:"n_predict"`
-	NKeep            int      `json:"n_keep"`
-	Temperature      float32  `json:"temperature"`
-	TopK             int      `json:"top_k"`
-	TopP             float32  `json:"top_p"`
-	TfsZ             float32  `json:"tfs_z"`
-	TypicalP         float32  `json:"typical_p"`
-	RepeatLastN      int      `json:"repeat_last_n"`
-	RepeatPenalty    float32  `json:"repeat_penalty"`
-	PresencePenalty  float32  `json:"presence_penalty"`
-	FrequencyPenalty float32  `json:"frequency_penalty"`
-	Mirostat         int      `json:"mirostat"`
-	MirostatTau      float32  `json:"mirostat_tau"`
-	MirostatEta      float32  `json:"mirostat_eta"`
-	PenalizeNl       bool     `json:"penalize_nl"`
-	Seed             int      `json:"seed"`
-	Stop             []string `json:"stop,omitempty"`
-}
-
-const maxBufferSize = 512 * 1000 // 512KB
+const maxBufferSize = 512 * format.KiloByte
 
 func (llm *llama) Predict(ctx context.Context, prevContext []int, prompt string, fn func(api.GenerateResponse)) error {
 	prevConvo, err := llm.Decode(ctx, prevContext)
@@ -500,39 +479,46 @@ func (llm *llama) Predict(ctx context.Context, prevContext []int, prompt string,
 		return err
 	}
 
+	// Remove leading spaces from prevConvo if present
+	prevConvo = strings.TrimPrefix(prevConvo, " ")
+
 	var nextContext strings.Builder
 	nextContext.WriteString(prevConvo)
 	nextContext.WriteString(prompt)
 
+	request := map[string]any{
+		"prompt":            nextContext.String(),
+		"stream":            true,
+		"n_predict":         llm.NumPredict,
+		"n_keep":            llm.NumKeep,
+		"temperature":       llm.Temperature,
+		"top_k":             llm.TopK,
+		"top_p":             llm.TopP,
+		"tfs_z":             llm.TFSZ,
+		"typical_p":         llm.TypicalP,
+		"repeat_last_n":     llm.RepeatLastN,
+		"repeat_penalty":    llm.RepeatPenalty,
+		"presence_penalty":  llm.PresencePenalty,
+		"frequency_penalty": llm.FrequencyPenalty,
+		"mirostat":          llm.Mirostat,
+		"mirostat_tau":      llm.MirostatTau,
+		"mirostat_eta":      llm.MirostatEta,
+		"penalize_nl":       llm.PenalizeNewline,
+		"seed":              llm.Seed,
+		"stop":              llm.Stop,
+	}
+
+	// Handling JSON marshaling with special characters unescaped.
+	buffer := &bytes.Buffer{}
+	enc := json.NewEncoder(buffer)
+	enc.SetEscapeHTML(false)
+
+	if err := enc.Encode(request); err != nil {
+		return fmt.Errorf("failed to marshal data: %v", err)
+	}
+
 	endpoint := fmt.Sprintf("http://127.0.0.1:%d/completion", llm.Port)
-	predReq := PredictRequest{
-		Prompt:           nextContext.String(),
-		Stream:           true,
-		NPredict:         llm.NumPredict,
-		NKeep:            llm.NumKeep,
-		Temperature:      llm.Temperature,
-		TopK:             llm.TopK,
-		TopP:             llm.TopP,
-		TfsZ:             llm.TFSZ,
-		TypicalP:         llm.TypicalP,
-		RepeatLastN:      llm.RepeatLastN,
-		RepeatPenalty:    llm.RepeatPenalty,
-		PresencePenalty:  llm.PresencePenalty,
-		FrequencyPenalty: llm.FrequencyPenalty,
-		Mirostat:         llm.Mirostat,
-		MirostatTau:      llm.MirostatTau,
-		MirostatEta:      llm.MirostatEta,
-		PenalizeNl:       llm.PenalizeNewline,
-		Seed:             llm.Seed,
-		Stop:             llm.Stop,
-	}
-
-	data, err := json.Marshal(predReq)
-	if err != nil {
-		return fmt.Errorf("error marshaling data: %v", err)
-	}
-
-	req, err := http.NewRequestWithContext(ctx, http.MethodPost, endpoint, bytes.NewBuffer(data))
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, endpoint, buffer)
 	if err != nil {
 		return fmt.Errorf("error creating POST request: %v", err)
 	}
@@ -563,16 +549,14 @@ func (llm *llama) Predict(ctx context.Context, prevContext []int, prompt string,
 			// This handles the request cancellation
 			return ctx.Err()
 		default:
-			line := scanner.Text()
-			if line == "" {
+			line := scanner.Bytes()
+			if len(line) == 0 {
 				continue
 			}
 
-			// Read data from the server-side event stream
-			if strings.HasPrefix(line, "data: ") {
-				evt := line[6:]
-				var p Prediction
-				if err := json.Unmarshal([]byte(evt), &p); err != nil {
+			if evt, ok := bytes.CutPrefix(line, []byte("data: ")); ok {
+				var p prediction
+				if err := json.Unmarshal(evt, &p); err != nil {
 					return fmt.Errorf("error unmarshaling llm prediction response: %v", err)
 				}
 
@@ -590,10 +574,10 @@ func (llm *llama) Predict(ctx context.Context, prevContext []int, prompt string,
 					fn(api.GenerateResponse{
 						Done:               true,
 						Context:            embd,
-						PromptEvalCount:    p.PromptN,
-						PromptEvalDuration: parseDurationMs(p.PromptMS),
-						EvalCount:          p.PredictedN,
-						EvalDuration:       parseDurationMs(p.PredictedMS),
+						PromptEvalCount:    p.Timings.PromptN,
+						PromptEvalDuration: parseDurationMs(p.Timings.PromptMS),
+						EvalCount:          p.Timings.PredictedN,
+						EvalDuration:       parseDurationMs(p.Timings.PredictedMS),
 					})
 
 					return nil
@@ -603,6 +587,14 @@ func (llm *llama) Predict(ctx context.Context, prevContext []int, prompt string,
 	}
 
 	if err := scanner.Err(); err != nil {
+		if strings.Contains(err.Error(), "unexpected EOF") {
+			// this means the llama runner subprocess crashed
+			llm.Close()
+			if llm.StatusWriter != nil && llm.StatusWriter.LastErrMsg != "" {
+				return fmt.Errorf("llama runner exited: %v", llm.StatusWriter.LastErrMsg)
+			}
+			return fmt.Errorf("llama runner exited, you may not have enough available memory to run this model")
+		}
 		return fmt.Errorf("error reading llm response: %v", err)
 	}
 
@@ -699,9 +691,6 @@ func (llm *llama) Decode(ctx context.Context, tokens []int) (string, error) {
 		return "", fmt.Errorf("unmarshal encode response: %w", err)
 	}
 
-	// decoded content contains a leading whitespace
-	decoded.Content, _ = strings.CutPrefix(decoded.Content, "")
-
 	return decoded.Content, nil
 }
 

llm/llm.go

@@ -10,6 +10,7 @@ import (
 	"github.com/pbnjay/memory"
 
 	"github.com/jmorganca/ollama/api"
+	"github.com/jmorganca/ollama/format"
 )
 
 type LLM interface {
@@ -55,39 +56,30 @@ func New(workDir, model string, adapters []string, opts api.Options) (LLM, error
 				opts.NumGPU = 0
 			}
 		}
-	}
 
-	totalResidentMemory := memory.TotalMemory()
-	switch ggml.ModelType() {
-	case "3B", "7B":
-		if ggml.FileType() == "F16" && totalResidentMemory < 16*1000*1000 {
-			return nil, fmt.Errorf("F16 model requires at least 16 GB of memory")
-		} else if totalResidentMemory < 8*1000*1000 {
-			return nil, fmt.Errorf("model requires at least 8 GB of memory")
+		var requiredMemory int64
+		var f16Multiplier int64 = 2
+
+		switch ggml.ModelType() {
+		case "3B", "7B":
+			requiredMemory = 8 * format.GigaByte
+		case "13B":
+			requiredMemory = 16 * format.GigaByte
+		case "30B", "34B", "40B":
+			requiredMemory = 32 * format.GigaByte
+		case "65B", "70B":
+			requiredMemory = 64 * format.GigaByte
+		case "180B":
+			requiredMemory = 128 * format.GigaByte
+			f16Multiplier = 4
 		}
-	case "13B":
-		if ggml.FileType() == "F16" && totalResidentMemory < 32*1000*1000 {
-			return nil, fmt.Errorf("F16 model requires at least 32 GB of memory")
-		} else if totalResidentMemory < 16*1000*1000 {
-			return nil, fmt.Errorf("model requires at least 16 GB of memory")
-		}
-	case "30B", "34B", "40B":
-		if ggml.FileType() == "F16" && totalResidentMemory < 64*1000*1000 {
-			return nil, fmt.Errorf("F16 model requires at least 64 GB of memory")
-		} else if totalResidentMemory < 32*1000*1000 {
-			return nil, fmt.Errorf("model requires at least 32 GB of memory")
-		}
-	case "65B", "70B":
-		if ggml.FileType() == "F16" && totalResidentMemory < 128*1000*1000 {
-			return nil, fmt.Errorf("F16 model requires at least 128 GB of memory")
-		} else if totalResidentMemory < 64*1000*1000 {
-			return nil, fmt.Errorf("model requires at least 64 GB of memory")
-		}
-	case "180B":
-		if ggml.FileType() == "F16" && totalResidentMemory < 512*1000*1000 {
-			return nil, fmt.Errorf("F16 model requires at least 512GB of memory")
-		} else if totalResidentMemory < 128*1000*1000 {
-			return nil, fmt.Errorf("model requires at least 128GB of memory")
+
+		systemMemory := int64(memory.TotalMemory())
+
+		if ggml.FileType() == "F16" && requiredMemory*f16Multiplier > systemMemory {
+			return nil, fmt.Errorf("F16 model requires at least %s of total memory", format.HumanBytes(requiredMemory))
+		} else if requiredMemory > systemMemory {
+			return nil, fmt.Errorf("model requires at least %s of total memory", format.HumanBytes(requiredMemory))
 		}
 	}
 

parser/parser.go

@@ -40,7 +40,7 @@ func Parse(reader io.Reader) ([]Command, error) {
 			command.Args = string(fields[1])
 			// copy command for validation
 			modelCommand = command
-		case "LICENSE", "TEMPLATE", "SYSTEM", "PROMPT", "EMBED", "ADAPTER":
+		case "LICENSE", "TEMPLATE", "SYSTEM", "PROMPT", "ADAPTER":
 			command.Name = string(bytes.ToLower(fields[0]))
 			command.Args = string(fields[1])
 		case "PARAMETER":
@@ -51,6 +51,8 @@ func Parse(reader io.Reader) ([]Command, error) {
 
 			command.Name = string(fields[0])
 			command.Args = string(fields[1])
+		case "EMBED":
+			return nil, fmt.Errorf("deprecated command: EMBED is no longer supported, use the /embed API endpoint instead")
 		default:
 			if !bytes.HasPrefix(fields[0], []byte("#")) {
 				// log a warning for unknown commands

scripts/install.sh

@@ -26,7 +26,8 @@ require() {
 
 [ "$(uname -s)" = "Linux" ] || error 'This script is intended to run on Linux only.'
 
-case "$(uname -m)" in
+ARCH=$(uname -m)
+case "$ARCH" in
     x86_64) ARCH="amd64" ;;
     aarch64|arm64) ARCH="arm64" ;;
     *) error "Unsupported architecture: $ARCH" ;;

scripts/push_docker.sh

@@ -0,0 +1,15 @@
+#!/bin/sh
+
+set -eu
+
+export VERSION=${VERSION:-0.0.0}
+export GOFLAGS="'-ldflags=-w -s \"-X=github.com/jmorganca/ollama/version.Version=$VERSION\" \"-X=github.com/jmorganca/ollama/server.mode=release\"'"
+
+docker buildx build \
+    --push \
+    --platform=linux/arm64,linux/amd64 \
+    --build-arg=VERSION \
+    --build-arg=GOFLAGS \
+    -f Dockerfile \
+    -t ollama/ollama -t ollama/ollama:$VERSION \
+    .

server/auth.go

@@ -106,6 +106,7 @@ func getAuthToken(ctx context.Context, redirData AuthRedirect) (string, error) {
 	resp, err := makeRequest(ctx, "GET", redirectURL, headers, nil, nil)
 	if err != nil {
 		log.Printf("couldn't get token: %q", err)
+		return "", err
 	}
 	defer resp.Body.Close()
 

server/images.go

@@ -1,7 +1,6 @@
 package server
 
 import (
-	"bufio"
 	"bytes"
 	"context"
 	"crypto/sha256"
@@ -26,7 +25,6 @@ import (
 	"github.com/jmorganca/ollama/api"
 	"github.com/jmorganca/ollama/llm"
 	"github.com/jmorganca/ollama/parser"
-	"github.com/jmorganca/ollama/vector"
 	"github.com/jmorganca/ollama/version"
 )
 
@@ -47,12 +45,10 @@ type Model struct {
 	System        string
 	License       []string
 	Digest        string
-	ConfigDigest  string
 	Options       map[string]interface{}
-	Embeddings    []vector.Embedding
 }
 
-func (m *Model) Prompt(request api.GenerateRequest, embedding string) (string, error) {
+func (m *Model) Prompt(request api.GenerateRequest) (string, error) {
 	t := m.Template
 	if request.Template != "" {
 		t = request.Template
@@ -67,7 +63,6 @@ func (m *Model) Prompt(request api.GenerateRequest, embedding string) (string, e
 		First  bool
 		System string
 		Prompt string
-		Embed  string
 
 		// deprecated: versions <= 0.0.7 used this to omit the system prompt
 		Context []int
@@ -77,7 +72,6 @@ func (m *Model) Prompt(request api.GenerateRequest, embedding string) (string, e
 	vars.System = m.System
 	vars.Prompt = request.Prompt
 	vars.Context = request.Context
-	vars.Embed = embedding
 
 	if request.System != "" {
 		vars.System = request.System
@@ -171,12 +165,11 @@ func GetModel(name string) (*Model, error) {
 	}
 
 	model := &Model{
-		Name:         mp.GetFullTagname(),
-		ShortName:    mp.GetShortTagname(),
-		Digest:       digest,
-		ConfigDigest: manifest.Config.Digest,
-		Template:     "{{ .Prompt }}",
-		License:      []string{},
+		Name:      mp.GetFullTagname(),
+		ShortName: mp.GetShortTagname(),
+		Digest:    digest,
+		Template:  "{{ .Prompt }}",
+		License:   []string{},
 	}
 
 	for _, layer := range manifest.Layers {
@@ -190,15 +183,9 @@ func GetModel(name string) (*Model, error) {
 			model.ModelPath = filename
 			model.OriginalModel = layer.From
 		case "application/vnd.ollama.image.embed":
-			file, err := os.Open(filename)
-			if err != nil {
-				return nil, fmt.Errorf("failed to open file: %s", filename)
-			}
-			defer file.Close()
-
-			if err = json.NewDecoder(file).Decode(&model.Embeddings); err != nil {
-				return nil, err
-			}
+			// Deprecated in versions  > 0.1.2
+			// TODO: remove this warning in a future version
+			log.Print("WARNING: model contains embeddings, but embeddings in modelfiles have been deprecated and will be ignored.")
 		case "application/vnd.ollama.image.adapter":
 			model.AdapterPaths = append(model.AdapterPaths, filename)
 		case "application/vnd.ollama.image.template":
@@ -265,7 +252,7 @@ func filenameWithPath(path, f string) (string, error) {
 	return f, nil
 }
 
-func CreateModel(ctx context.Context, workDir, name string, path string, fn func(resp api.ProgressResponse)) error {
+func CreateModel(ctx context.Context, name string, path string, fn func(resp api.ProgressResponse)) error {
 	mp := ParseModelPath(name)
 
 	var manifest *ManifestV2
@@ -310,13 +297,11 @@ func CreateModel(ctx context.Context, workDir, name string, path string, fn func
 	var layers []*LayerReader
 	params := make(map[string][]string)
 	var sourceParams map[string]any
-	embed := EmbeddingParams{fn: fn}
 	for _, c := range commands {
 		log.Printf("[%s] - %s\n", c.Name, c.Args)
 		switch c.Name {
 		case "model":
 			fn(api.ProgressResponse{Status: "looking for model"})
-			embed.model = c.Args
 
 			mp := ParseModelPath(c.Args)
 			mf, _, err := GetManifest(mp)
@@ -340,7 +325,6 @@ func CreateModel(ctx context.Context, workDir, name string, path string, fn func
 						return err
 					}
 				} else {
-					embed.model = modelFile
 					// create a model from this specified file
 					fn(api.ProgressResponse{Status: "creating model layer"})
 					file, err := os.Open(modelFile)
@@ -421,12 +405,6 @@ func CreateModel(ctx context.Context, workDir, name string, path string, fn func
 					layers = append(layers, newLayer)
 				}
 			}
-		case "embed":
-			embedFilePath, err := filenameWithPath(path, c.Args)
-			if err != nil {
-				return err
-			}
-			embed.files = append(embed.files, embedFilePath)
 		case "adapter":
 			fn(api.ProgressResponse{Status: fmt.Sprintf("creating model %s layer", c.Name)})
 
@@ -517,18 +495,8 @@ func CreateModel(ctx context.Context, workDir, name string, path string, fn func
 		}
 		l.MediaType = "application/vnd.ollama.image.params"
 		layers = append(layers, l)
-
-		// apply these parameters to the embedding options, in case embeddings need to be generated using this model
-		embed.opts = formattedParams
 	}
 
-	// generate the embedding layers
-	embeddingLayers, err := embeddingLayers(workDir, embed)
-	if err != nil {
-		return err
-	}
-	layers = append(layers, embeddingLayers...)
-
 	digests, err := getLayerDigests(layers)
 	if err != nil {
 		return err
@@ -572,146 +540,6 @@ func CreateModel(ctx context.Context, workDir, name string, path string, fn func
 	return nil
 }
 
-type EmbeddingParams struct {
-	model string
-	opts  map[string]interface{}
-	files []string // paths to files to embed
-	fn    func(resp api.ProgressResponse)
-}
-
-// embeddingLayers loads the associated LLM and generates the embeddings to be stored from an input file
-func embeddingLayers(workDir string, e EmbeddingParams) ([]*LayerReader, error) {
-	layers := []*LayerReader{}
-	if len(e.files) > 0 {
-		// check if the model is a file path or a model name
-		model, err := GetModel(e.model)
-		if err != nil {
-			if !strings.Contains(err.Error(), "couldn't open file") {
-				return nil, fmt.Errorf("unexpected error opening model to generate embeddings: %v", err)
-			}
-			// the model may be a file path, create a model from this file
-			model = &Model{ModelPath: e.model}
-		}
-
-		if err := load(context.Background(), workDir, model, e.opts, defaultSessionDuration); err != nil {
-			return nil, fmt.Errorf("load model to generate embeddings: %v", err)
-		}
-
-		// this will be used to check if we already have embeddings for a file
-		modelInfo, err := os.Stat(model.ModelPath)
-		if err != nil {
-			return nil, fmt.Errorf("failed to get model file info: %v", err)
-		}
-
-		addedFiles := make(map[string]bool) // keep track of files that have already been added
-		for _, filePattern := range e.files {
-			matchingFiles, err := filepath.Glob(filePattern)
-			if err != nil {
-				return nil, fmt.Errorf("could not find files with pattern %s: %w", filePattern, err)
-			}
-
-			for _, filePath := range matchingFiles {
-				if addedFiles[filePath] {
-					continue
-				}
-				addedFiles[filePath] = true
-				// check if we already have embeddings for this file path
-				layerIdentifier := fmt.Sprintf("%s:%s:%s:%d", filePath, e.model, modelInfo.ModTime().Format("2006-01-02 15:04:05"), modelInfo.Size())
-				digest, _ := GetSHA256Digest(strings.NewReader(layerIdentifier))
-				existing, err := existingFileEmbeddings(digest)
-				if err != nil {
-					return nil, fmt.Errorf("failed to check existing embeddings for file %s: %v", filePath, err)
-				}
-
-				// TODO: check file type
-				f, err := os.Open(filePath)
-				if err != nil {
-					return nil, fmt.Errorf("could not open embed file: %w", err)
-				}
-				scanner := bufio.NewScanner(f)
-				scanner.Split(bufio.ScanLines)
-
-				data := []string{}
-				for scanner.Scan() {
-					data = append(data, scanner.Text())
-				}
-				f.Close()
-
-				// the digest of the file is set here so that the client knows a new operation is in progress
-				fileDigest, _ := GetSHA256Digest(bytes.NewReader([]byte(filePath)))
-
-				embeddings := []vector.Embedding{}
-				for i, d := range data {
-					if strings.TrimSpace(d) == "" {
-						continue
-					}
-					e.fn(api.ProgressResponse{
-						Status:    fmt.Sprintf("creating embeddings for file %s", filePath),
-						Digest:    fileDigest,
-						Total:     int64(len(data) - 1),
-						Completed: int64(i),
-					})
-					if len(existing[d]) > 0 {
-						// already have an embedding for this line
-						embeddings = append(embeddings, vector.Embedding{Data: d, Vector: existing[d]})
-						continue
-					}
-					embed, err := loaded.llm.Embedding(context.Background(), d)
-					if err != nil {
-						log.Printf("failed to generate embedding for '%s' line %d: %v", filePath, i+1, err)
-						continue
-					}
-					embeddings = append(embeddings, vector.Embedding{Data: d, Vector: embed})
-				}
-
-				b, err := json.Marshal(embeddings)
-				if err != nil {
-					return nil, fmt.Errorf("failed to encode embeddings: %w", err)
-				}
-				r := bytes.NewReader(b)
-
-				layer := &LayerReader{
-					Layer: Layer{
-						MediaType: "application/vnd.ollama.image.embed",
-						Digest:    digest,
-						Size:      r.Size(),
-					},
-					Reader: r,
-				}
-
-				layers = append(layers, layer)
-			}
-		}
-	}
-	return layers, nil
-}
-
-// existingFileEmbeddings checks if we already have embeddings for a file and loads them into a look-up map
-func existingFileEmbeddings(digest string) (map[string][]float64, error) {
-	path, err := GetBlobsPath(digest)
-	if err != nil {
-		return nil, fmt.Errorf("embeddings blobs path: %w", err)
-	}
-	existingFileEmbeddings := make(map[string][]float64)
-	if _, err := os.Stat(path); err == nil {
-		// already have some embeddings for this file, load embeddings previously generated
-		file, err := os.Open(path)
-		if err != nil {
-			return nil, fmt.Errorf("failed to open existing embedding file: %s", err)
-		}
-		defer file.Close()
-
-		existing := []vector.Embedding{}
-		if err = json.NewDecoder(file).Decode(&existing); err != nil {
-			return nil, err
-		}
-		for _, e := range existing {
-			existingFileEmbeddings[e.Data] = e.Vector
-		}
-	}
-	return existingFileEmbeddings, nil
-}
-
 func removeLayerFromLayers(layers []*LayerReader, mediaType string) []*LayerReader {
 	return slices.DeleteFunc(layers, func(layer *LayerReader) bool {
 		return layer.MediaType == mediaType
@@ -727,8 +555,7 @@ func SaveLayers(layers []*LayerReader, fn func(resp api.ProgressResponse), force
 		}
 
 		_, err = os.Stat(fp)
-		// note: embed layers are always written since their digest doesnt indicate anything about the contents
-		if os.IsNotExist(err) || force || layer.MediaType == "application/vnd.ollama.image.embed" {
+		if os.IsNotExist(err) || force {
 			fn(api.ProgressResponse{Status: fmt.Sprintf("writing layer %s", layer.Digest)})
 
 			out, err := os.Create(fp)
@@ -1069,51 +896,27 @@ func DeleteModel(name string) error {
 }
 
 func ShowModelfile(model *Model) (string, error) {
-	type modelTemplate struct {
+	var mt struct {
 		*Model
-		From   string
-		Params string
+		From       string
+		Parameters map[string][]any
 	}
 
-	var params []string
+	mt.Parameters = make(map[string][]any)
 	for k, v := range model.Options {
-		switch val := v.(type) {
-		case string:
-			params = append(params, fmt.Sprintf("PARAMETER %s %s", k, val))
-		case int:
-			params = append(params, fmt.Sprintf("PARAMETER %s %s", k, strconv.Itoa(val)))
-		case float64:
-			params = append(params, fmt.Sprintf("PARAMETER %s %s", k, strconv.FormatFloat(val, 'f', 0, 64)))
-		case bool:
-			params = append(params, fmt.Sprintf("PARAMETER %s %s", k, strconv.FormatBool(val)))
-		case []interface{}:
-			for _, nv := range val {
-				switch nval := nv.(type) {
-				case string:
-					params = append(params, fmt.Sprintf("PARAMETER %s %s", k, nval))
-				case int:
-					params = append(params, fmt.Sprintf("PARAMETER %s %s", k, strconv.Itoa(nval)))
-				case float64:
-					params = append(params, fmt.Sprintf("PARAMETER %s %s", k, strconv.FormatFloat(nval, 'f', 0, 64)))
-				case bool:
-					params = append(params, fmt.Sprintf("PARAMETER %s %s", k, strconv.FormatBool(nval)))
-				default:
-					log.Printf("unknown type: %s", reflect.TypeOf(nv).String())
-				}
-			}
-		default:
-			log.Printf("unknown type: %s", reflect.TypeOf(v).String())
+		if s, ok := v.([]any); ok {
+			mt.Parameters[k] = s
+			continue
 		}
+
+		mt.Parameters[k] = []any{v}
 	}
 
-	mt := modelTemplate{
-		Model:  model,
-		From:   model.OriginalModel,
-		Params: strings.Join(params, "\n"),
-	}
+	mt.Model = model
+	mt.From = model.ModelPath
 
-	if mt.From == "" {
-		mt.From = model.ModelPath
+	if model.OriginalModel != "" {
+		mt.From = model.OriginalModel
 	}
 
 	modelFile := `# Modelfile generated by "ollama show"
@@ -1122,12 +925,20 @@ func ShowModelfile(model *Model) (string, error) {
 
 FROM {{ .From }}
 TEMPLATE """{{ .Template }}"""
+
+{{- if .System }}
 SYSTEM """{{ .System }}"""
-{{ .Params }}
-`
-	for _, l := range mt.Model.AdapterPaths {
-		modelFile += fmt.Sprintf("ADAPTER %s\n", l)
-	}
+{{- end }}
+
+{{- range $adapter := .AdapterPaths }}
+ADAPTER {{ $adapter }}
+{{- end }}
+
+{{- range $k, $v := .Parameters }}
+{{- range $parameter := $v }}
+PARAMETER {{ $k }} {{ printf "%#v" $parameter }}
+{{- end }}
+{{- end }}`
 
 	tmpl, err := template.New("").Parse(modelFile)
 	if err != nil {

server/images_test.go

@@ -12,7 +12,7 @@ func TestModelPrompt(t *testing.T) {
 		Template: "a{{ .Prompt }}b",
 		Prompt:   "<h1>",
 	}
-	s, err := m.Prompt(req, "")
+	s, err := m.Prompt(req)
 	if err != nil {
 		t.Fatal(err)
 	}

server/routes.go

@@ -23,11 +23,10 @@ import (
 
 	"github.com/gin-contrib/cors"
 	"github.com/gin-gonic/gin"
-	"gonum.org/v1/gonum/mat"
 
 	"github.com/jmorganca/ollama/api"
 	"github.com/jmorganca/ollama/llm"
-	"github.com/jmorganca/ollama/vector"
+	"github.com/jmorganca/ollama/version"
 )
 
 var mode string = gin.DebugMode
@@ -47,14 +46,13 @@ func init() {
 var loaded struct {
 	mu sync.Mutex
 
-	llm        llm.LLM
-	Embeddings []vector.Embedding
+	runner llm.LLM
 
 	expireAt    time.Time
 	expireTimer *time.Timer
 
-	digest  string
-	options api.Options
+	*Model
+	*api.Options
 }
 
 var defaultSessionDuration = 5 * time.Minute
@@ -72,66 +70,52 @@ func load(ctx context.Context, workDir string, model *Model, reqOpts map[string]
 	}
 
 	// check if the loaded model is still running in a subprocess, in case something unexpected happened
-	if loaded.llm != nil {
-		if err := loaded.llm.Ping(ctx); err != nil {
+	if loaded.runner != nil {
+		if err := loaded.runner.Ping(ctx); err != nil {
 			log.Print("loaded llm process not responding, closing now")
 			// the subprocess is no longer running, so close it
-			loaded.llm.Close()
-			loaded.llm = nil
-			loaded.digest = ""
+			loaded.runner.Close()
+			loaded.runner = nil
+			loaded.Model = nil
+			loaded.Options = nil
 		}
 	}
 
-	if model.Digest != loaded.digest || !reflect.DeepEqual(loaded.options, opts) {
-		if loaded.llm != nil {
+	needLoad := loaded.runner == nil || // is there a model loaded?
+		loaded.ModelPath != model.ModelPath || // has the base model changed?
+		!reflect.DeepEqual(loaded.AdapterPaths, model.AdapterPaths) || // have the adapters changed?
+		!reflect.DeepEqual(loaded.Options.Runner, opts.Runner) // have the runner options changed?
+
+	if needLoad {
+		if loaded.runner != nil {
 			log.Println("changing loaded model")
-			loaded.llm.Close()
-			loaded.llm = nil
-			loaded.digest = ""
+			loaded.runner.Close()
+			loaded.runner = nil
+			loaded.Model = nil
+			loaded.Options = nil
 		}
 
-		if model.Embeddings != nil && len(model.Embeddings) > 0 {
-			opts.EmbeddingOnly = true // this is requried to generate embeddings, completions will still work
-			loaded.Embeddings = model.Embeddings
-		}
-
-		llmModel, err := llm.New(workDir, model.ModelPath, model.AdapterPaths, opts)
+		llmRunner, err := llm.New(workDir, model.ModelPath, model.AdapterPaths, opts)
 		if err != nil {
+			// some older models are not compatible with newer versions of llama.cpp
+			// show a generalized compatibility error until there is a better way to
+			// check for model compatibility
+			if strings.Contains(err.Error(), "failed to load model") {
+				err = fmt.Errorf("%v: this model may be incompatible with your version of Ollama. If you previously pulled this model, try updating it by running `ollama pull %s`", err, model.ShortName)
+			}
+
 			return err
 		}
 
-		// set cache values before modifying opts
-		loaded.llm = llmModel
-		loaded.digest = model.Digest
-		loaded.options = opts
-
-		if opts.NumKeep < 0 {
-			promptWithSystem, err := model.Prompt(api.GenerateRequest{}, "")
-			if err != nil {
-				return err
-			}
-
-			promptNoSystem, err := model.Prompt(api.GenerateRequest{Context: []int{0}}, "")
-			if err != nil {
-				return err
-			}
-
-			tokensWithSystem, err := llmModel.Encode(ctx, promptWithSystem)
-			if err != nil {
-				return err
-			}
-
-			tokensNoSystem, err := llmModel.Encode(ctx, promptNoSystem)
-			if err != nil {
-				return err
-			}
-
-			opts.NumKeep = len(tokensWithSystem) - len(tokensNoSystem)
-
-			llmModel.SetOptions(opts)
-		}
+		loaded.Model = model
+		loaded.runner = llmRunner
+		loaded.Options = &opts
 	}
 
+	// update options for the loaded llm
+	// TODO(mxyng): this isn't thread safe, but it should be fine for now
+	loaded.runner.SetOptions(opts)
+
 	loaded.expireAt = time.Now().Add(sessionDuration)
 
 	if loaded.expireTimer == nil {
@@ -143,13 +127,13 @@ func load(ctx context.Context, workDir string, model *Model, reqOpts map[string]
 				return
 			}
 
-			if loaded.llm == nil {
-				return
+			if loaded.runner != nil {
+				loaded.runner.Close()
 			}
 
-			loaded.llm.Close()
-			loaded.llm = nil
-			loaded.digest = ""
+			loaded.runner = nil
+			loaded.Model = nil
+			loaded.Options = nil
 		})
 	}
 
@@ -164,8 +148,18 @@ func GenerateHandler(c *gin.Context) {
 	checkpointStart := time.Now()
 
 	var req api.GenerateRequest
-	if err := c.ShouldBindJSON(&req); err != nil {
-		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+	err := c.ShouldBindJSON(&req)
+	switch {
+	case errors.Is(err, io.EOF):
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
+		return
+	case err != nil:
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		return
+	}
+
+	if req.Model == "" {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "model is required"})
 		return
 	}
 
@@ -195,22 +189,7 @@ func GenerateHandler(c *gin.Context) {
 
 	checkpointLoaded := time.Now()
 
-	embedding := ""
-	if model.Embeddings != nil && len(model.Embeddings) > 0 {
-		promptEmbed, err := loaded.llm.Embedding(c.Request.Context(), req.Prompt)
-		if err != nil {
-			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-			return
-		}
-		// TODO: set embed_top from specified parameters in modelfile
-		embed_top := 3
-		topK := vector.TopK(embed_top, mat.NewVecDense(len(promptEmbed), promptEmbed), loaded.Embeddings)
-		for _, e := range topK {
-			embedding = fmt.Sprintf("%s %s", embedding, e.Embedding.Data)
-		}
-	}
-
-	prompt, err := model.Prompt(req, embedding)
+	prompt, err := model.Prompt(req)
 	if err != nil {
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 		return
@@ -219,6 +198,12 @@ func GenerateHandler(c *gin.Context) {
 	ch := make(chan any)
 	go func() {
 		defer close(ch)
+		// an empty request loads the model
+		if req.Prompt == "" && req.Template == "" && req.System == "" {
+			ch <- api.GenerateResponse{CreatedAt: time.Now().UTC(), Model: req.Model, Done: true}
+			return
+		}
+
 		fn := func(r api.GenerateResponse) {
 			loaded.expireAt = time.Now().Add(sessionDuration)
 			loaded.expireTimer.Reset(sessionDuration)
@@ -233,13 +218,8 @@ func GenerateHandler(c *gin.Context) {
 			ch <- r
 		}
 
-		// an empty request loads the model
-		if req.Prompt == "" && req.Template == "" && req.System == "" {
-			ch <- api.GenerateResponse{Model: req.Model, Done: true}
-		} else {
-			if err := loaded.llm.Predict(c.Request.Context(), req.Context, prompt, fn); err != nil {
-				ch <- gin.H{"error": err.Error()}
-			}
+		if err := loaded.runner.Predict(c.Request.Context(), req.Context, prompt, fn); err != nil {
+			ch <- gin.H{"error": err.Error()}
 		}
 	}()
 
@@ -268,8 +248,18 @@ func EmbeddingHandler(c *gin.Context) {
 	defer loaded.mu.Unlock()
 
 	var req api.EmbeddingRequest
-	if err := c.ShouldBindJSON(&req); err != nil {
-		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+	err := c.ShouldBindJSON(&req)
+	switch {
+	case errors.Is(err, io.EOF):
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
+		return
+	case err != nil:
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		return
+	}
+
+	if req.Model == "" {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "model is required"})
 		return
 	}
 
@@ -285,12 +275,12 @@ func EmbeddingHandler(c *gin.Context) {
 		return
 	}
 
-	if !loaded.options.EmbeddingOnly {
+	if !loaded.Options.EmbeddingOnly {
 		c.JSON(http.StatusBadRequest, gin.H{"error": "embedding option must be set to true"})
 		return
 	}
 
-	embedding, err := loaded.llm.Embedding(c.Request.Context(), req.Prompt)
+	embedding, err := loaded.runner.Embedding(c.Request.Context(), req.Prompt)
 	if err != nil {
 		log.Printf("embedding generation failed: %v", err)
 		c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to generate embedding"})
@@ -305,8 +295,18 @@ func EmbeddingHandler(c *gin.Context) {
 
 func PullModelHandler(c *gin.Context) {
 	var req api.PullRequest
-	if err := c.ShouldBindJSON(&req); err != nil {
-		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+	err := c.ShouldBindJSON(&req)
+	switch {
+	case errors.Is(err, io.EOF):
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
+		return
+	case err != nil:
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		return
+	}
+
+	if req.Name == "" {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "name is required"})
 		return
 	}
 
@@ -339,8 +339,18 @@ func PullModelHandler(c *gin.Context) {
 
 func PushModelHandler(c *gin.Context) {
 	var req api.PushRequest
-	if err := c.ShouldBindJSON(&req); err != nil {
-		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+	err := c.ShouldBindJSON(&req)
+	switch {
+	case errors.Is(err, io.EOF):
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
+		return
+	case err != nil:
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		return
+	}
+
+	if req.Name == "" {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "name is required"})
 		return
 	}
 
@@ -371,12 +381,20 @@ func PushModelHandler(c *gin.Context) {
 
 func CreateModelHandler(c *gin.Context) {
 	var req api.CreateRequest
-	if err := c.ShouldBindJSON(&req); err != nil {
-		c.JSON(http.StatusBadRequest, gin.H{"message": err.Error()})
+	err := c.ShouldBindJSON(&req)
+	switch {
+	case errors.Is(err, io.EOF):
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
+		return
+	case err != nil:
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
 		return
 	}
 
-	workDir := c.GetString("workDir")
+	if req.Name == "" || req.Path == "" {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "name and path are required"})
+		return
+	}
 
 	ch := make(chan any)
 	go func() {
@@ -388,7 +406,7 @@ func CreateModelHandler(c *gin.Context) {
 		ctx, cancel := context.WithCancel(c.Request.Context())
 		defer cancel()
 
-		if err := CreateModel(ctx, workDir, req.Name, req.Path, fn); err != nil {
+		if err := CreateModel(ctx, req.Name, req.Path, fn); err != nil {
 			ch <- gin.H{"error": err.Error()}
 		}
 	}()
@@ -403,8 +421,18 @@ func CreateModelHandler(c *gin.Context) {
 
 func DeleteModelHandler(c *gin.Context) {
 	var req api.DeleteRequest
-	if err := c.ShouldBindJSON(&req); err != nil {
-		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+	err := c.ShouldBindJSON(&req)
+	switch {
+	case errors.Is(err, io.EOF):
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
+		return
+	case err != nil:
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		return
+	}
+
+	if req.Name == "" {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "name is required"})
 		return
 	}
 
@@ -433,8 +461,18 @@ func DeleteModelHandler(c *gin.Context) {
 
 func ShowModelHandler(c *gin.Context) {
 	var req api.ShowRequest
-	if err := c.ShouldBindJSON(&req); err != nil {
-		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+	err := c.ShouldBindJSON(&req)
+	switch {
+	case errors.Is(err, io.EOF):
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
+		return
+	case err != nil:
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		return
+	}
+
+	if req.Name == "" {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "name is required"})
 		return
 	}
 
@@ -503,7 +541,7 @@ func GetModelInfo(name string) (*api.ShowResponse, error) {
 }
 
 func ListModelsHandler(c *gin.Context) {
-	var models []api.ModelResponse
+	models := make([]api.ModelResponse, 0)
 	fp, err := GetManifestPath()
 	if err != nil {
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
@@ -544,8 +582,18 @@ func ListModelsHandler(c *gin.Context) {
 
 func CopyModelHandler(c *gin.Context) {
 	var req api.CopyRequest
-	if err := c.ShouldBindJSON(&req); err != nil {
-		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+	err := c.ShouldBindJSON(&req)
+	switch {
+	case errors.Is(err, io.EOF):
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
+		return
+	case err != nil:
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		return
+	}
+
+	if req.Source == "" || req.Destination == "" {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "source add destination are required"})
 		return
 	}
 
@@ -611,7 +659,7 @@ func Serve(ln net.Listener, allowOrigins []string) error {
 		r.Handle(method, "/api/tags", ListModelsHandler)
 	}
 
-	log.Printf("Listening on %s", ln.Addr())
+	log.Printf("Listening on %s (version %s)", ln.Addr(), version.Version)
 	s := &http.Server{
 		Handler: r,
 	}
@@ -621,8 +669,8 @@ func Serve(ln net.Listener, allowOrigins []string) error {
 	signal.Notify(signals, syscall.SIGINT, syscall.SIGTERM)
 	go func() {
 		<-signals
-		if loaded.llm != nil {
-			loaded.llm.Close()
+		if loaded.runner != nil {
+			loaded.runner.Close()
 		}
 		os.RemoveAll(workDir)
 		os.Exit(0)

vector/store.go

@@ -1,69 +0,0 @@
-package vector
-
-import (
-	"container/heap"
-	"sort"
-
-	"gonum.org/v1/gonum/mat"
-)
-
-type Embedding struct {
-	Vector []float64 // the embedding vector
-	Data   string    // the data represted by the embedding
-}
-
-type EmbeddingSimilarity struct {
-	Embedding  Embedding // the embedding that was used to calculate the similarity
-	Similarity float64   // the similarity between the embedding and the query
-}
-
-type Heap []EmbeddingSimilarity
-
-func (h Heap) Len() int           { return len(h) }
-func (h Heap) Less(i, j int) bool { return h[i].Similarity < h[j].Similarity }
-func (h Heap) Swap(i, j int)      { h[i], h[j] = h[j], h[i] }
-func (h *Heap) Push(e any) {
-	*h = append(*h, e.(EmbeddingSimilarity))
-}
-
-func (h *Heap) Pop() interface{} {
-	old := *h
-	n := len(old)
-	x := old[n-1]
-	*h = old[0 : n-1]
-	return x
-}
-
-// cosineSimilarity is a measure that calculates the cosine of the angle between two vectors.
-// This value will range from -1 to 1, where 1 means the vectors are identical.
-func cosineSimilarity(vec1, vec2 *mat.VecDense) float64 {
-	dotProduct := mat.Dot(vec1, vec2)
-	norms := mat.Norm(vec1, 2) * mat.Norm(vec2, 2)
-
-	if norms == 0 {
-		return 0
-	}
-	return dotProduct / norms
-}
-
-func TopK(k int, query *mat.VecDense, embeddings []Embedding) []EmbeddingSimilarity {
-	h := &Heap{}
-	heap.Init(h)
-	for _, emb := range embeddings {
-		similarity := cosineSimilarity(query, mat.NewVecDense(len(emb.Vector), emb.Vector))
-		heap.Push(h, EmbeddingSimilarity{Embedding: emb, Similarity: similarity})
-		if h.Len() > k {
-			heap.Pop(h)
-		}
-	}
-
-	topK := make([]EmbeddingSimilarity, 0, h.Len())
-	for h.Len() > 0 {
-		topK = append(topK, heap.Pop(h).(EmbeddingSimilarity))
-	}
-	sort.Slice(topK, func(i, j int) bool {
-		return topK[i].Similarity > topK[j].Similarity
-	})
-
-	return topK
-}