add model_format config var

Go is required and not installed by default.

.dockerignore

@@ -1,7 +1,4 @@
-build
-llama/build
-.venv
 .vscode
 ollama
 app
-web
+llm/llama.cpp/ggml

.gitignore

@@ -5,4 +5,4 @@
 .swp
 dist
 ollama
-/ggml-metal.metal
+ggml-metal.metal

.gitmodules

@@ -0,0 +1,4 @@
+[submodule "llm/llama.cpp/ggml"]
+	path = llm/llama.cpp/ggml
+	url = https://github.com/ggerganov/llama.cpp.git
+	ignore = dirty

Dockerfile

@@ -1,15 +1,21 @@
-FROM golang:1.20
+FROM golang:alpine
+
 WORKDIR /go/src/github.com/jmorganca/ollama
+RUN apk add --no-cache git build-base cmake
+
 COPY . .
-RUN CGO_ENABLED=1 go build -ldflags '-linkmode external -extldflags "-static"' .
+RUN go generate ./... && go build -ldflags '-linkmode external -extldflags "-static"' .
 
 FROM alpine
-COPY --from=0 /go/src/github.com/jmorganca/ollama/ollama /bin/ollama
-EXPOSE 11434
+ENV OLLAMA_HOST 0.0.0.0
+RUN apk add --no-cache libstdc++
+
 ARG USER=ollama
 ARG GROUP=ollama
-RUN addgroup -g 1000 $GROUP && adduser -u 1000 -DG $GROUP $USER
+RUN addgroup $GROUP && adduser -D -G $GROUP $USER
+
+COPY --from=0 /go/src/github.com/jmorganca/ollama/ollama /bin/ollama
+
 USER $USER:$GROUP
 ENTRYPOINT ["/bin/ollama"]
-ENV OLLAMA_HOST 0.0.0.0
 CMD ["serve"]

README.md

@@ -29,7 +29,9 @@ ollama run llama2
 
 ## Model library
 
-`ollama` includes a library of open-source models:
+Ollama supports a list of open-source models available on [ollama.ai/library](https://ollama.ai/library 'ollama model library')
+
+Here are some example open-source models that can be downloaded:
 
 | Model                    | Parameters | Size  | Download                        |
 | ------------------------ | ---------- | ----- | ------------------------------- |
@@ -37,16 +39,26 @@ ollama run llama2
 | Llama2 13B               | 13B        | 7.3GB | `ollama pull llama2:13b`        |
 | Llama2 70B               | 70B        | 39GB  | `ollama pull llama2:70b`        |
 | Llama2 Uncensored        | 7B         | 3.8GB | `ollama pull llama2-uncensored` |
-| Orca Mini                | 3B         | 1.9GB | `ollama pull orca`              |
+| Code Llama               | 7B         | 3.8GB | `ollama pull codellama`         |
+| Orca Mini                | 3B         | 1.9GB | `ollama pull orca-mini`         |
 | Vicuna                   | 7B         | 3.8GB | `ollama pull vicuna`            |
-| Nous-Hermes              | 13B        | 7.3GB | `ollama pull nous-hermes`       |
+| Nous-Hermes              | 7B         | 3.8GB | `ollama pull nous-hermes`       |
+| Nous-Hermes 13B          | 13B        | 7.3GB | `ollama pull nous-hermes:13b`   |
 | Wizard Vicuna Uncensored | 13B        | 7.3GB | `ollama pull wizard-vicuna`     |
 
 > Note: You should have at least 8 GB of RAM to run the 3B models, 16 GB to run the 7B models, and 32 GB to run the 13B models.
 
 ## Examples
 
-### Run a model
+### Pull a public model
+
+```
+ollama pull llama2
+```
+
+> This command can also be used to update a local model. Only updated changes will be pulled.
+
+### Run a model interactively
 
 ```
 ollama run llama2
@@ -63,7 +75,39 @@ For multiline input, you can wrap text with `"""`:
 I'm a basic program that prints the famous "Hello, world!" message to the console.
 ```
 
-### Create a custom model
+### Run a model non-interactively
+
+```
+$ ollama run llama2 'tell me a joke'
+ Sure! Here's a quick one:
+ Why did the scarecrow win an award? Because he was outstanding in his field!
+```
+
+```
+$ cat <<EOF >prompts.txt
+tell me a joke about llamas
+tell me another one
+EOF
+$ ollama run llama2 <prompts.txt
+>>> tell me a joke about llamas
+ Why did the llama refuse to play hide-and-seek?
+ nobody likes to be hided!
+
+>>> tell me another one
+ Sure, here's another one:
+
+Why did the llama go to the bar?
+To have a hay-often good time!
+```
+
+### Run a model on contents of a text file
+
+```
+$ ollama run llama2 "summarize this file:" "$(cat README.md)"
+ Ollama is a lightweight, extensible framework for building and running language models on the local machine. It provides a simple API for creating, running, and managing models, as well as a library of pre-built models that can be easily used in a variety of applications.
+```
+
+### Customize a model
 
 Pull a base model:
 
@@ -71,8 +115,6 @@ Pull a base model:
 ollama pull llama2
 ```
 
-> To update a model to the latest version, run `ollama pull llama2` again. The model will be updated (if necessary).
-
 Create a `Modelfile`:
 
 ```
@@ -98,23 +140,23 @@ Hello! It's your friend Mario.
 
 For more examples, see the [examples](./examples) directory. For more information on creating a Modelfile, see the [Modelfile](./docs/modelfile.md) documentation.
 
-### Pull a model from the registry
-
-```
-ollama pull orca
-```
-
 ### Listing local models
 
 ```
 ollama list
 ```
 
+### Removing local models
+
+```
+ollama rm llama2
+```
+
 ## Model packages
 
 ### Overview
 
-Ollama bundles model weights, configuration, and data into a single package, defined by a [Modelfile](./docs/modelfile.md).
+Ollama bundles model weights, configurations, and data into a single package, defined by a [Modelfile](./docs/modelfile.md).
 
 <picture>
   <source media="(prefers-color-scheme: dark)" height="480" srcset="https://github.com/jmorganca/ollama/assets/251292/2fd96b5f-191b-45c1-9668-941cfad4eb70">
@@ -123,17 +165,27 @@ Ollama bundles model weights, configuration, and data into a single package, def
 
 ## Building
 
+Install `cmake` and `go`:
+
 ```
+brew install cmake
+brew install go
+```
+
+Then generate dependencies and build:
+
+```
+go generate ./...
 go build .
 ```
 
-To run it start the server:
+Next, start the server:
 
 ```
-./ollama serve &
+./ollama serve
 ```
 
-Finally, run a model!
+Finally, in a separate shell, run a model:
 
 ```
 ./ollama run llama2
@@ -152,10 +204,12 @@ curl -X POST http://localhost:11434/api/generate -d '{
 }'
 ```
 
-## Tools using Ollama
+## Community Projects using Ollama
 
 - [LangChain](https://python.langchain.com/docs/integrations/llms/ollama) and [LangChain.js](https://js.langchain.com/docs/modules/model_io/models/llms/integrations/ollama) with a question-answering [example](https://js.langchain.com/docs/use_cases/question_answering/local_retrieval_qa).
 - [Continue](https://github.com/continuedev/continue) - embeds Ollama inside Visual Studio Code. The extension lets you highlight code to add to the prompt, ask questions in the sidebar, and generate code inline.
+- [LiteLLM](https://github.com/BerriAI/litellm) a lightweight python package to simplify LLM API calls
 - [Discord AI Bot](https://github.com/mekb-turtle/discord-ai-bot) - interact with Ollama as a chatbot on Discord.
 - [Raycast Ollama](https://github.com/MassimilianoPasquini97/raycast_ollama) - Raycast extension to use Ollama for local llama inference on Raycast.
 - [Simple HTML UI for Ollama](https://github.com/rtcfirefly/ollama-ui)
+- [Emacs client](https://github.com/zweifisch/ollama) for Ollama

api/client.go

@@ -9,16 +9,27 @@ import (
 	"io"
 	"net/http"
 	"net/url"
+	"os"
+	"runtime"
+	"strings"
+
+	"github.com/jmorganca/ollama/version"
+)
+
+const DefaultHost = "127.0.0.1:11434"
+
+var (
+	envHost = os.Getenv("OLLAMA_HOST")
 )
 
 type Client struct {
-	base    url.URL
+	Base    url.URL
 	HTTP    http.Client
 	Headers http.Header
 }
 
 func checkError(resp *http.Response, body []byte) error {
-	if resp.StatusCode >= 200 && resp.StatusCode < 400 {
+	if resp.StatusCode < http.StatusBadRequest {
 		return nil
 	}
 
@@ -33,16 +44,34 @@ func checkError(resp *http.Response, body []byte) error {
 	return apiError
 }
 
-func NewClient(hosts ...string) *Client {
-	host := "127.0.0.1:11434"
-	if len(hosts) > 0 {
-		host = hosts[0]
+// Host returns the default host to use for the client. It is determined in the following order:
+// 1. The OLLAMA_HOST environment variable
+// 2. The default host (localhost:11434)
+func Host() string {
+	if envHost != "" {
+		return envHost
+	}
+	return DefaultHost
+}
+
+// FromEnv creates a new client using Host() as the host. An error is returns
+// if the host is invalid.
+func FromEnv() (*Client, error) {
+	h := Host()
+	if !strings.HasPrefix(h, "http://") && !strings.HasPrefix(h, "https://") {
+		h = "http://" + h
 	}
 
-	return &Client{
-		base: url.URL{Scheme: "http", Host: host},
-		HTTP: http.Client{},
+	u, err := url.Parse(h)
+	if err != nil {
+		return nil, fmt.Errorf("could not parse host: %w", err)
 	}
+
+	if u.Port() == "" {
+		u.Host += ":11434"
+	}
+
+	return &Client{Base: *u, HTTP: http.Client{}}, nil
 }
 
 func (c *Client) do(ctx context.Context, method, path string, reqData, respData any) error {
@@ -57,21 +86,21 @@ func (c *Client) do(ctx context.Context, method, path string, reqData, respData
 		reqBody = bytes.NewReader(data)
 	}
 
-	url := c.base.JoinPath(path).String()
-
-	req, err := http.NewRequestWithContext(ctx, method, url, reqBody)
+	requestURL := c.Base.JoinPath(path)
+	request, err := http.NewRequestWithContext(ctx, method, requestURL.String(), reqBody)
 	if err != nil {
 		return err
 	}
 
-	req.Header.Set("Content-Type", "application/json")
-	req.Header.Set("Accept", "application/json")
+	request.Header.Set("Content-Type", "application/json")
+	request.Header.Set("Accept", "application/json")
+	request.Header.Set("User-Agent", fmt.Sprintf("ollama/%s (%s %s) Go/%s", version.Version, runtime.GOARCH, runtime.GOOS, runtime.Version()))
 
 	for k, v := range c.Headers {
-		req.Header[k] = v
+		request.Header[k] = v
 	}
 
-	respObj, err := c.HTTP.Do(req)
+	respObj, err := c.HTTP.Do(request)
 	if err != nil {
 		return err
 	}
@@ -105,13 +134,15 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 		buf = bytes.NewBuffer(bts)
 	}
 
-	request, err := http.NewRequestWithContext(ctx, method, c.base.JoinPath(path).String(), buf)
+	requestURL := c.Base.JoinPath(path)
+	request, err := http.NewRequestWithContext(ctx, method, requestURL.String(), buf)
 	if err != nil {
 		return err
 	}
 
 	request.Header.Set("Content-Type", "application/json")
 	request.Header.Set("Accept", "application/json")
+	request.Header.Set("User-Agent", fmt.Sprintf("ollama/%s (%s %s) Go/%s", version.Version, runtime.GOARCH, runtime.GOOS, runtime.Version()))
 
 	response, err := http.DefaultClient.Do(request)
 	if err != nil {
@@ -134,7 +165,7 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 			return fmt.Errorf(errorResponse.Error)
 		}
 
-		if response.StatusCode >= 400 {
+		if response.StatusCode >= http.StatusBadRequest {
 			return StatusError{
 				StatusCode:   response.StatusCode,
 				Status:       response.Status,
@@ -224,6 +255,14 @@ func (c *Client) Delete(ctx context.Context, req *DeleteRequest) error {
 	return nil
 }
 
+func (c *Client) Show(ctx context.Context, req *ShowRequest) (*ShowResponse, error) {
+	var resp ShowResponse
+	if err := c.do(ctx, http.MethodPost, "/api/show", req, &resp); err != nil {
+		return nil, err
+	}
+	return &resp, nil
+}
+
 func (c *Client) Heartbeat(ctx context.Context) error {
 	if err := c.do(ctx, http.MethodHead, "/", nil, nil); err != nil {
 		return err

api/types.go

@@ -7,7 +7,6 @@ import (
 	"math"
 	"os"
 	"reflect"
-	"runtime"
 	"strings"
 	"time"
 )
@@ -62,6 +61,18 @@ type DeleteRequest struct {
 	Name string `json:"name"`
 }
 
+type ShowRequest struct {
+	Name string `json:"name"`
+}
+
+type ShowResponse struct {
+	License    string `json:"license,omitempty"`
+	Modelfile  string `json:"modelfile,omitempty"`
+	Parameters string `json:"parameters,omitempty"`
+	Template   string `json:"template,omitempty"`
+	System     string `json:"system,omitempty"`
+}
+
 type CopyRequest struct {
 	Source      string `json:"source"`
 	Destination string `json:"destination"`
@@ -89,13 +100,14 @@ type PushRequest struct {
 }
 
 type ListResponse struct {
-	Models []ListResponseModel `json:"models"`
+	Models []ModelResponse `json:"models"`
 }
 
-type ListResponseModel struct {
+type ModelResponse struct {
 	Name       string    `json:"name"`
 	ModifiedAt time.Time `json:"modified_at"`
 	Size       int       `json:"size"`
+	Digest     string    `json:"digest"`
 }
 
 type TokenResponse struct {
@@ -112,8 +124,6 @@ type GenerateResponse struct {
 
 	TotalDuration      time.Duration `json:"total_duration,omitempty"`
 	LoadDuration       time.Duration `json:"load_duration,omitempty"`
-	SampleCount        int           `json:"sample_count,omitempty"`
-	SampleDuration     time.Duration `json:"sample_duration,omitempty"`
 	PromptEvalCount    int           `json:"prompt_eval_count,omitempty"`
 	PromptEvalDuration time.Duration `json:"prompt_eval_duration,omitempty"`
 	EvalCount          int           `json:"eval_count,omitempty"`
@@ -129,15 +139,6 @@ func (r *GenerateResponse) Summary() {
 		fmt.Fprintf(os.Stderr, "load duration:        %v\n", r.LoadDuration)
 	}
 
-	if r.SampleCount > 0 {
-		fmt.Fprintf(os.Stderr, "sample count:         %d token(s)\n", r.SampleCount)
-	}
-
-	if r.SampleDuration > 0 {
-		fmt.Fprintf(os.Stderr, "sample duration:      %s\n", r.SampleDuration)
-		fmt.Fprintf(os.Stderr, "sample rate:          %.2f tokens/s\n", float64(r.SampleCount)/r.SampleDuration.Seconds())
-	}
-
 	if r.PromptEvalCount > 0 {
 		fmt.Fprintf(os.Stderr, "prompt eval count:    %d token(s)\n", r.PromptEvalCount)
 	}
@@ -181,15 +182,16 @@ type Options struct {
 	RopeFrequencyScale float32 `json:"rope_frequency_scale,omitempty"`
 
 	// Predict options
-	RepeatLastN      int      `json:"repeat_last_n,omitempty"`
-	RepeatPenalty    float32  `json:"repeat_penalty,omitempty"`
-	FrequencyPenalty float32  `json:"frequency_penalty,omitempty"`
-	PresencePenalty  float32  `json:"presence_penalty,omitempty"`
-	Temperature      float32  `json:"temperature,omitempty"`
+	NumPredict       int      `json:"num_predict,omitempty"`
 	TopK             int      `json:"top_k,omitempty"`
 	TopP             float32  `json:"top_p,omitempty"`
 	TFSZ             float32  `json:"tfs_z,omitempty"`
 	TypicalP         float32  `json:"typical_p,omitempty"`
+	RepeatLastN      int      `json:"repeat_last_n,omitempty"`
+	Temperature      float32  `json:"temperature,omitempty"`
+	RepeatPenalty    float32  `json:"repeat_penalty,omitempty"`
+	PresencePenalty  float32  `json:"presence_penalty,omitempty"`
+	FrequencyPenalty float32  `json:"frequency_penalty,omitempty"`
 	Mirostat         int      `json:"mirostat,omitempty"`
 	MirostatTau      float32  `json:"mirostat_tau,omitempty"`
 	MirostatEta      float32  `json:"mirostat_eta,omitempty"`
@@ -216,19 +218,25 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
 		if opt, ok := jsonOpts[key]; ok {
 			field := valueOpts.FieldByName(opt.Name)
 			if field.IsValid() && field.CanSet() {
+				if val == nil {
+					continue
+				}
+
 				switch field.Kind() {
 				case reflect.Int:
-					// when JSON unmarshals numbers, it uses float64 by default, not int
-					val, ok := val.(float64)
-					if !ok {
-						log.Printf("could not convert model parmeter %v to int, skipped", key)
-						continue
+					switch t := val.(type) {
+					case int64:
+						field.SetInt(t)
+					case float64:
+						// when JSON unmarshals numbers, it uses float64, not int
+						field.SetInt(int64(t))
+					default:
+						log.Printf("could not convert model parameter %v to int, skipped", key)
 					}
-					field.SetInt(int64(val))
 				case reflect.Bool:
 					val, ok := val.(bool)
 					if !ok {
-						log.Printf("could not convert model parmeter %v to bool, skipped", key)
+						log.Printf("could not convert model parameter %v to bool, skipped", key)
 						continue
 					}
 					field.SetBool(val)
@@ -236,14 +244,14 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
 					// JSON unmarshals to float64
 					val, ok := val.(float64)
 					if !ok {
-						log.Printf("could not convert model parmeter %v to float32, skipped", key)
+						log.Printf("could not convert model parameter %v to float32, skipped", key)
 						continue
 					}
 					field.SetFloat(val)
 				case reflect.String:
 					val, ok := val.(string)
 					if !ok {
-						log.Printf("could not convert model parmeter %v to string, skipped", key)
+						log.Printf("could not convert model parameter %v to string, skipped", key)
 						continue
 					}
 					field.SetString(val)
@@ -251,7 +259,7 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
 					// JSON unmarshals to []interface{}, not []string
 					val, ok := val.([]interface{})
 					if !ok {
-						log.Printf("could not convert model parmeter %v to slice, skipped", key)
+						log.Printf("could not convert model parameter %v to slice, skipped", key)
 						continue
 					}
 					// convert []interface{} to []string
@@ -259,7 +267,7 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
 					for i, item := range val {
 						str, ok := item.(string)
 						if !ok {
-							log.Printf("could not convert model parmeter %v to slice of strings, skipped", key)
+							log.Printf("could not convert model parameter %v to slice of strings, skipped", key)
 							continue
 						}
 						slice[i] = str
@@ -307,7 +315,7 @@ func DefaultOptions() Options {
 		MirostatEta:      0.1,
 		PenalizeNewline:  true,
 
-		NumThread: runtime.NumCPU(),
+		NumThread: 0, // let the runtime decide
 	}
 }
 

app/forge.config.ts

@@ -27,7 +27,6 @@ const config: ForgeConfig = {
       path.join(__dirname, './assets/iconDarkTemplate@2x.png'),
       path.join(__dirname, './assets/iconDarkUpdateTemplate.png'),
       path.join(__dirname, './assets/iconDarkUpdateTemplate@2x.png'),
-      ...(process.platform === 'darwin' ? ['../llama/ggml-metal.metal'] : []),
     ],
     ...(process.env.SIGN
       ? {

app/src/index.ts

@@ -158,7 +158,7 @@ function restart() {
 app.on('before-quit', () => {
   if (proc) {
     proc.off('exit', restart)
-    proc.kill()
+    proc.kill('SIGINT') // send SIGINT signal to the server, which also stops any loaded llms
   }
 })
 

cmd/cmd.go

@@ -3,6 +3,9 @@ package cmd
 import (
 	"bufio"
 	"context"
+	"crypto/ed25519"
+	"crypto/rand"
+	"encoding/pem"
 	"errors"
 	"fmt"
 	"io"
@@ -11,6 +14,7 @@ import (
 	"net/http"
 	"os"
 	"os/exec"
+	"path"
 	"path/filepath"
 	"runtime"
 	"strings"
@@ -20,11 +24,13 @@ import (
 	"github.com/dustin/go-humanize"
 	"github.com/olekukonko/tablewriter"
 	"github.com/spf13/cobra"
+	"golang.org/x/crypto/ssh"
 
 	"github.com/jmorganca/ollama/api"
 	"github.com/jmorganca/ollama/format"
 	"github.com/jmorganca/ollama/progressbar"
 	"github.com/jmorganca/ollama/server"
+	"github.com/jmorganca/ollama/version"
 )
 
 func CreateHandler(cmd *cobra.Command, args []string) error {
@@ -34,7 +40,10 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}
 
-	client := api.NewClient()
+	client, err := api.FromEnv()
+	if err != nil {
+		return err
+	}
 
 	var spinner *Spinner
 
@@ -70,6 +79,7 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 			spinner = NewSpinner(resp.Status)
 			go spinner.Spin(100 * time.Millisecond)
 		}
+
 		return nil
 	}
 
@@ -79,13 +89,29 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 
 	if spinner != nil {
 		spinner.Stop()
+		if spinner.description != "success" {
+			return errors.New("unexpected end to create model")
+		}
 	}
 
 	return nil
 }
 
 func RunHandler(cmd *cobra.Command, args []string) error {
+	insecure, err := cmd.Flags().GetBool("insecure")
+	if err != nil {
+		return err
+	}
+
 	mp := server.ParseModelPath(args[0])
+	if err != nil {
+		return err
+	}
+
+	if mp.ProtocolScheme == "http" && !insecure {
+		return fmt.Errorf("insecure protocol http")
+	}
+
 	fp, err := mp.GetManifestPath(false)
 	if err != nil {
 		return err
@@ -94,7 +120,7 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	_, err = os.Stat(fp)
 	switch {
 	case errors.Is(err, os.ErrNotExist):
-		if err := pull(args[0], false); err != nil {
+		if err := pull(args[0], insecure); err != nil {
 			var apiStatusError api.StatusError
 			if !errors.As(err, &apiStatusError) {
 				return err
@@ -112,7 +138,10 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 }
 
 func PushHandler(cmd *cobra.Command, args []string) error {
-	client := api.NewClient()
+	client, err := api.FromEnv()
+	if err != nil {
+		return err
+	}
 
 	insecure, err := cmd.Flags().GetBool("insecure")
 	if err != nil {
@@ -144,11 +173,19 @@ func PushHandler(cmd *cobra.Command, args []string) error {
 	if err := client.Push(context.Background(), &request, fn); err != nil {
 		return err
 	}
+
+	if bar != nil && !bar.IsFinished() {
+		return errors.New("unexpected end to push model")
+	}
+
 	return nil
 }
 
 func ListHandler(cmd *cobra.Command, args []string) error {
-	client := api.NewClient()
+	client, err := api.FromEnv()
+	if err != nil {
+		return err
+	}
 
 	models, err := client.List(context.Background())
 	if err != nil {
@@ -159,12 +196,12 @@ func ListHandler(cmd *cobra.Command, args []string) error {
 
 	for _, m := range models.Models {
 		if len(args) == 0 || strings.HasPrefix(m.Name, args[0]) {
-			data = append(data, []string{m.Name, humanize.Bytes(uint64(m.Size)), format.HumanTime(m.ModifiedAt, "Never")})
+			data = append(data, []string{m.Name, m.Digest[:12], humanize.Bytes(uint64(m.Size)), format.HumanTime(m.ModifiedAt, "Never")})
 		}
 	}
 
 	table := tablewriter.NewWriter(os.Stdout)
-	table.SetHeader([]string{"NAME", "SIZE", "MODIFIED"})
+	table.SetHeader([]string{"NAME", "ID", "SIZE", "MODIFIED"})
 	table.SetHeaderAlignment(tablewriter.ALIGN_LEFT)
 	table.SetAlignment(tablewriter.ALIGN_LEFT)
 	table.SetHeaderLine(false)
@@ -178,18 +215,104 @@ func ListHandler(cmd *cobra.Command, args []string) error {
 }
 
 func DeleteHandler(cmd *cobra.Command, args []string) error {
-	client := api.NewClient()
-
-	req := api.DeleteRequest{Name: args[0]}
-	if err := client.Delete(context.Background(), &req); err != nil {
+	client, err := api.FromEnv()
+	if err != nil {
 		return err
 	}
-	fmt.Printf("deleted '%s'\n", args[0])
+
+	for _, name := range args {
+		req := api.DeleteRequest{Name: name}
+		if err := client.Delete(context.Background(), &req); err != nil {
+			return err
+		}
+		fmt.Printf("deleted '%s'\n", name)
+	}
+	return nil
+}
+
+func ShowHandler(cmd *cobra.Command, args []string) error {
+	client, err := api.FromEnv()
+	if err != nil {
+		return err
+	}
+
+	if len(args) != 1 {
+		return errors.New("missing model name")
+	}
+
+	license, errLicense := cmd.Flags().GetBool("license")
+	modelfile, errModelfile := cmd.Flags().GetBool("modelfile")
+	parameters, errParams := cmd.Flags().GetBool("parameters")
+	system, errSystem := cmd.Flags().GetBool("system")
+	template, errTemplate := cmd.Flags().GetBool("template")
+
+	for _, boolErr := range []error{errLicense, errModelfile, errParams, errSystem, errTemplate} {
+		if boolErr != nil {
+			return errors.New("error retrieving flags")
+		}
+	}
+
+	flagsSet := 0
+	showType := ""
+
+	if license {
+		flagsSet++
+		showType = "license"
+	}
+
+	if modelfile {
+		flagsSet++
+		showType = "modelfile"
+	}
+
+	if parameters {
+		flagsSet++
+		showType = "parameters"
+	}
+
+	if system {
+		flagsSet++
+		showType = "system"
+	}
+
+	if template {
+		flagsSet++
+		showType = "template"
+	}
+
+	if flagsSet > 1 {
+		return errors.New("only one of '--license', '--modelfile', '--parameters', '--system', or '--template' can be specified")
+	} else if flagsSet == 0 {
+		return errors.New("one of '--license', '--modelfile', '--parameters', '--system', or '--template' must be specified")
+	}
+
+	req := api.ShowRequest{Name: args[0]}
+	resp, err := client.Show(context.Background(), &req)
+	if err != nil {
+		return err
+	}
+
+	switch showType {
+	case "license":
+		fmt.Println(resp.License)
+	case "modelfile":
+		fmt.Println(resp.Modelfile)
+	case "parameters":
+		fmt.Println(resp.Parameters)
+	case "system":
+		fmt.Println(resp.System)
+	case "template":
+		fmt.Println(resp.Template)
+	}
+
 	return nil
 }
 
 func CopyHandler(cmd *cobra.Command, args []string) error {
-	client := api.NewClient()
+	client, err := api.FromEnv()
+	if err != nil {
+		return err
+	}
 
 	req := api.CopyRequest{Source: args[0], Destination: args[1]}
 	if err := client.Copy(context.Background(), &req); err != nil {
@@ -209,7 +332,10 @@ func PullHandler(cmd *cobra.Command, args []string) error {
 }
 
 func pull(model string, insecure bool) error {
-	client := api.NewClient()
+	client, err := api.FromEnv()
+	if err != nil {
+		return err
+	}
 
 	var currentDigest string
 	var bar *progressbar.ProgressBar
@@ -230,12 +356,18 @@ func pull(model string, insecure bool) error {
 			currentDigest = ""
 			fmt.Println(resp.Status)
 		}
+
 		return nil
 	}
 
 	if err := client.Pull(context.Background(), &request, fn); err != nil {
 		return err
 	}
+
+	if bar != nil && !bar.IsFinished() {
+		return errors.New("unexpected end to pull model")
+	}
+
 	return nil
 }
 
@@ -256,7 +388,10 @@ type generateContextKey string
 
 func generate(cmd *cobra.Command, model, prompt string) error {
 	if len(strings.TrimSpace(prompt)) > 0 {
-		client := api.NewClient()
+		client, err := api.FromEnv()
+		if err != nil {
+			return err
+		}
 
 		spinner := NewSpinner("")
 		go spinner.Spin(60 * time.Millisecond)
@@ -299,6 +434,10 @@ func generate(cmd *cobra.Command, model, prompt string) error {
 		fmt.Println()
 		fmt.Println()
 
+		if !latest.Done {
+			return errors.New("unexpected end of response")
+		}
+
 		verbose, err := cmd.Flags().GetBool("verbose")
 		if err != nil {
 			return err
@@ -316,20 +455,6 @@ func generate(cmd *cobra.Command, model, prompt string) error {
 	return nil
 }
 
-func showLayer(l *server.Layer) {
-	filename, err := server.GetBlobsPath(l.Digest)
-	if err != nil {
-		fmt.Println("Couldn't get layer's path")
-		return
-	}
-	bts, err := os.ReadFile(filename)
-	if err != nil {
-		fmt.Println("Couldn't read layer")
-		return
-	}
-	fmt.Println(string(bts))
-}
-
 func generateInteractive(cmd *cobra.Command, model string) error {
 	home, err := os.UserHomeDir()
 	if err != nil {
@@ -352,6 +477,8 @@ func generateInteractive(cmd *cobra.Command, model string) error {
 		),
 		readline.PcItem("/show",
 			readline.PcItem("license"),
+			readline.PcItem("modelfile"),
+			readline.PcItem("parameters"),
 			readline.PcItem("system"),
 			readline.PcItem("template"),
 		),
@@ -461,38 +588,28 @@ func generateInteractive(cmd *cobra.Command, model string) error {
 		case strings.HasPrefix(line, "/show"):
 			args := strings.Fields(line)
 			if len(args) > 1 {
-				mp := server.ParseModelPath(model)
-				manifest, err := server.GetManifest(mp)
+				resp, err := server.GetModelInfo(model)
 				if err != nil {
-					fmt.Println("error: couldn't get a manifest for this model")
+					fmt.Println("error: couldn't get model")
 					continue
 				}
+
 				switch args[1] {
 				case "license":
-					for _, l := range manifest.Layers {
-						if l.MediaType == "application/vnd.ollama.image.license" {
-							showLayer(l)
-						}
-					}
-					continue
+					fmt.Println(resp.License)
+				case "modelfile":
+					fmt.Println(resp.Modelfile)
+				case "parameters":
+					fmt.Println(resp.Parameters)
 				case "system":
-					for _, l := range manifest.Layers {
-						if l.MediaType == "application/vnd.ollama.image.system" {
-							showLayer(l)
-						}
-					}
-					continue
+					fmt.Println(resp.System)
 				case "template":
-					for _, l := range manifest.Layers {
-						if l.MediaType == "application/vnd.ollama.image.template" {
-							showLayer(l)
-						}
-					}
-					continue
+					fmt.Println(resp.Template)
 				default:
-					usage()
-					continue
+					fmt.Println("error: unknown command")
 				}
+
+				continue
 			} else {
 				usage()
 				continue
@@ -524,7 +641,7 @@ func generateBatch(cmd *cobra.Command, model string) error {
 }
 
 func RunServer(cmd *cobra.Command, _ []string) error {
-	var host, port = "127.0.0.1", "11434"
+	host, port := "127.0.0.1", "11434"
 
 	parts := strings.Split(os.Getenv("OLLAMA_HOST"), ":")
 	if ip := net.ParseIP(parts[0]); ip != nil {
@@ -540,6 +657,11 @@ func RunServer(cmd *cobra.Command, _ []string) error {
 		port = p
 	}
 
+	err := initializeKeypair()
+	if err != nil {
+		return err
+	}
+
 	ln, err := net.Listen("tcp", fmt.Sprintf("%s:%s", host, port))
 	if err != nil {
 		return err
@@ -553,6 +675,55 @@ func RunServer(cmd *cobra.Command, _ []string) error {
 	return server.Serve(ln, origins)
 }
 
+func initializeKeypair() error {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return err
+	}
+
+	privKeyPath := filepath.Join(home, ".ollama", "id_ed25519")
+	pubKeyPath := filepath.Join(home, ".ollama", "id_ed25519.pub")
+
+	_, err = os.Stat(privKeyPath)
+	if os.IsNotExist(err) {
+		fmt.Printf("Couldn't find '%s'. Generating new private key.\n", privKeyPath)
+		_, privKey, err := ed25519.GenerateKey(rand.Reader)
+		if err != nil {
+			return err
+		}
+
+		privKeyBytes, err := format.OpenSSHPrivateKey(privKey, "")
+		if err != nil {
+			return err
+		}
+
+		err = os.MkdirAll(path.Dir(privKeyPath), 0o700)
+		if err != nil {
+			return fmt.Errorf("could not create directory %w", err)
+		}
+
+		err = os.WriteFile(privKeyPath, pem.EncodeToMemory(privKeyBytes), 0o600)
+		if err != nil {
+			return err
+		}
+
+		sshPrivateKey, err := ssh.NewSignerFromKey(privKey)
+		if err != nil {
+			return err
+		}
+
+		pubKeyData := ssh.MarshalAuthorizedKey(sshPrivateKey.PublicKey())
+
+		err = os.WriteFile(pubKeyPath, pubKeyData, 0o644)
+		if err != nil {
+			return err
+		}
+
+		fmt.Printf("Your new public key is: \n\n%s\n", string(pubKeyData))
+	}
+	return nil
+}
+
 func startMacApp(client *api.Client) error {
 	exe, err := os.Executable()
 	if err != nil {
@@ -585,7 +756,10 @@ func startMacApp(client *api.Client) error {
 }
 
 func checkServerHeartbeat(_ *cobra.Command, _ []string) error {
-	client := api.NewClient()
+	client, err := api.FromEnv()
+	if err != nil {
+		return err
+	}
 	if err := client.Heartbeat(context.Background()); err != nil {
 		if !strings.Contains(err.Error(), "connection refused") {
 			return err
@@ -605,12 +779,14 @@ func NewCLI() *cobra.Command {
 	log.SetFlags(log.LstdFlags | log.Lshortfile)
 
 	rootCmd := &cobra.Command{
-		Use:          "ollama",
-		Short:        "Large language model runner",
-		SilenceUsage: true,
+		Use:           "ollama",
+		Short:         "Large language model runner",
+		SilenceUsage:  true,
+		SilenceErrors: true,
 		CompletionOptions: cobra.CompletionOptions{
 			DisableDefaultCmd: true,
 		},
+		Version: version.Version,
 	}
 
 	cobra.EnableCommandSorting = false
@@ -625,6 +801,20 @@ func NewCLI() *cobra.Command {
 
 	createCmd.Flags().StringP("file", "f", "Modelfile", "Name of the Modelfile (default \"Modelfile\")")
 
+	showCmd := &cobra.Command{
+		Use:     "show MODEL",
+		Short:   "Show information for a model",
+		Args:    cobra.MinimumNArgs(1),
+		PreRunE: checkServerHeartbeat,
+		RunE:    ShowHandler,
+	}
+
+	showCmd.Flags().Bool("license", false, "Show license of a model")
+	showCmd.Flags().Bool("modelfile", false, "Show Modelfile of a model")
+	showCmd.Flags().Bool("parameters", false, "Show parameters of a model")
+	showCmd.Flags().Bool("template", false, "Show template of a model")
+	showCmd.Flags().Bool("system", false, "Show system prompt of a model")
+
 	runCmd := &cobra.Command{
 		Use:     "run MODEL [PROMPT]",
 		Short:   "Run a model",
@@ -634,6 +824,7 @@ func NewCLI() *cobra.Command {
 	}
 
 	runCmd.Flags().Bool("verbose", false, "Show timings for response")
+	runCmd.Flags().Bool("insecure", false, "Use an insecure registry")
 
 	serveCmd := &cobra.Command{
 		Use:     "serve",
@@ -689,6 +880,7 @@ func NewCLI() *cobra.Command {
 	rootCmd.AddCommand(
 		serveCmd,
 		createCmd,
+		showCmd,
 		runCmd,
 		pullCmd,
 		pushCmd,

docs/README.md

@@ -3,3 +3,4 @@
 - [Modelfile](./modelfile.md)
 - [How to develop Ollama](./development.md)
 - [API](./api.md)
+- [Tutorials](./tutorials.md)

docs/api.md

@@ -8,12 +8,13 @@
 - [Copy a model](#copy-a-model)
 - [Delete a model](#delete-a-model)
 - [Pull a model](#pull-a-model)
+- [Generate embeddings](#generate-embeddings)
 
 ## Conventions
 
 ### Model names
 
-Model names follow a `model:tag` format. Some examples are `orca:3b-q4_1` and `llama2:70b`. The tag is optional and if not provided will default to `latest`. The tag is used to identify a specific version.
+Model names follow a `model:tag` format. Some examples are `orca-mini:3b-q4_1` and `llama2:70b`. The tag is optional and if not provided will default to `latest`. The tag is used to identify a specific version.
 
 ### Durations
 
@@ -37,6 +38,7 @@ Advanced parameters:
 - `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
 - `system`: system prompt to (overrides what is defined in the `Modelfile`)
 - `template`: the full prompt or prompt template (overrides what is defined in the `Modelfile`)
+- `context`: the context parameter returned from a previous request to `/generate`, this can be used to keep a short conversational memory
 
 ### Request
 
@@ -70,6 +72,7 @@ The final response in the stream also includes additional data about the generat
 - `prompt_eval_duration`: time spent in nanoseconds evaluating the prompt
 - `eval_count`: number of tokens the response
 - `eval_duration`: time in nanoseconds spent generating the response
+- `context`: an encoding of the conversation used in this response, this can be sent in the next request to keep a conversational memory
 
 To calculate how fast the response is generated in tokens per second (token/s), divide `eval_count` / `eval_duration`.
 
@@ -77,6 +80,7 @@ To calculate how fast the response is generated in tokens per second (token/s),
 {
   "model": "llama2:7b",
   "created_at": "2023-08-04T19:22:45.499127Z",
+  "context": [1, 2, 3],
   "done": true,
   "total_duration": 5589157167,
   "load_duration": 3013701500,
@@ -220,3 +224,40 @@ curl -X POST http://localhost:11434/api/pull -d '{
   "total": 2142590208
 }
 ```
+
+## Generate Embeddings
+
+```
+POST /api/embeddings
+```
+
+Generate embeddings from a model
+
+### Parameters
+
+- `model`: name of model to generate embeddings from
+- `prompt`: text to generate embeddings for
+
+Advanced parameters:
+
+- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
+
+### Request
+
+```
+curl -X POST http://localhost:11434/api/embeddings -d '{
+  "model": "llama2:7b",
+  "prompt": "Here is an article about llamas..."
+}'
+```
+
+### Response
+
+```json
+{
+  "embeddings": [
+    0.5670403838157654, 0.009260174818336964, 0.23178744316101074, -0.2916173040866852, -0.8924556970596313,
+    0.8785552978515625, -0.34576427936553955, 0.5742510557174683, -0.04222835972905159, -0.137906014919281
+  ]
+}
+```

docs/development.md

@@ -1,19 +1,21 @@
 # Development
 
+- Install cmake or (optionally, required tools for GPUs)
+- run `go generate ./...`
+- run `go build .`
+
 Install required tools:
 
 ```
-brew install go
+brew install go cmake gcc
 ```
 
-Enable CGO:
+Get the required libraries:
 
 ```
-export CGO_ENABLED=1
+go generate ./...
 ```
 
-You will also need a C/C++ compiler such as GCC for MacOS and Linux or Mingw-w64 GCC for Windows.
-
 Then build ollama:
 
 ```
@@ -25,20 +27,3 @@ Now you can run `ollama`:
 ```
 ./ollama
 ```
-
-## Releasing
-
-To release a new version of Ollama you'll need to set some environment variables:
-
-- `GITHUB_TOKEN`: your GitHub token
-- `APPLE_IDENTITY`: the Apple signing identity (macOS only)
-- `APPLE_ID`: your Apple ID
-- `APPLE_PASSWORD`: your Apple ID app-specific password
-- `APPLE_TEAM_ID`: the Apple team ID for the signing identity
-- `TELEMETRY_WRITE_KEY`: segment write key for telemetry
-
-Then run the publish script with the target version:
-
-```
-VERSION=0.0.2 ./scripts/publish.sh
-```

docs/modelfile.md

@@ -12,11 +12,13 @@ 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)
     - [Template Variables](#template-variables)
   - [SYSTEM](#system)
+  - [ADAPTER](#adapter)
   - [LICENSE](#license)
 - [Notes](#notes)
 
@@ -35,6 +37,7 @@ INSTRUCTION arguments
 | [`PARAMETER`](#parameter)           | Sets the parameters for how Ollama will run the model.        |
 | [`TEMPLATE`](#template)             | The full prompt template to be sent to the model.             |
 | [`SYSTEM`](#system)                 | Specifies the system prompt that will be set in the template. |
+| [`ADAPTER`](#adapter)               | Defines the (Q)LoRA adapters to apply to the model.           |
 | [`LICENSE`](#license)               | Specifies the legal license.                                  |
 
 ## Examples
@@ -88,6 +91,16 @@ FROM ./ollama-model.bin
 
 This bin file location should be specified as an absolute path or relative to the Modelfile location.
 
+### EMBED
+
+The EMBED instruction is used to add embeddings of files to a model. This is useful for adding custom data that the model can reference when generating an answer. Note that currently only text files are supported, formatted with each line as one embedding.
+```
+FROM <model name>:<tag>
+EMBED <file path>.txt
+EMBED <different file path>.txt
+EMBED <path to directory>/*.txt
+```
+
 ### PARAMETER
 
 The `PARAMETER` instruction defines a parameter that can be set when the model is run.
@@ -104,12 +117,13 @@ PARAMETER <parameter> <parametervalue>
 | mirostat_eta   | Influences how quickly the algorithm responds to feedback from the generated text. A lower learning rate will result in slower adjustments, while a higher learning rate will make the algorithm more responsive. (Default: 0.1)                        | float      | mirostat_eta 0.1     |
 | mirostat_tau   | Controls the balance between coherence and diversity of the output. A lower value will result in more focused and coherent text. (Default: 5.0)                                                                                                         | float      | mirostat_tau 5.0     |
 | num_ctx        | Sets the size of the context window used to generate the next token. (Default: 2048)                                                                                                                                                                    | int        | num_ctx 4096         |
+| num_gqa        | The number of GQA groups in the transformer layer. Required for some models, for example it is 8 for llama2:70b                                                                                                                                         | int        | num_gqa 1            |
 | num_gpu        | The number of GPUs to use. On macOS it defaults to 1 to enable metal support, 0 to disable.                                                                                                                                                             | int        | num_gpu 1            |
 | num_thread     | Sets the number of threads to use during computation. By default, Ollama will detect this for optimal performance. It is recommended to set this value to the number of physical CPU cores your system has (as opposed to the logical number of cores). | int        | num_thread 8         |
 | repeat_last_n  | Sets how far back for the model to look back to prevent repetition. (Default: 64, 0 = disabled, -1 = num_ctx)                                                                                                                                           | int        | repeat_last_n 64     |
 | repeat_penalty | Sets how strongly to penalize repetitions. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. (Default: 1.1)                                                                     | float      | repeat_penalty 1.1   |
 | temperature    | The temperature of the model. Increasing the temperature will make the model answer more creatively. (Default: 0.8)                                                                                                                                     | float      | temperature 0.7      |
-| stop           | Sets the stop tokens to use.                                                                                                                                                                                                                            | string     | stop "AI assistant:" |
+| 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              |
 | top_k          | Reduces the probability of generating nonsense. A higher value (e.g. 100) will give more diverse answers, while a lower value (e.g. 10) will be more conservative. (Default: 40)                                                                        | int        | top_k 40             |
 | top_p          | Works together with top-k. A higher value (e.g., 0.95) will lead to more diverse text, while a lower value (e.g., 0.5) will generate more focused and conservative text. (Default: 0.9)                                                                 | float      | top_p 0.9            |
@@ -150,6 +164,14 @@ The `SYSTEM` instruction specifies the system prompt to be used in the template,
 SYSTEM """<system message>"""
 ```
 
+### ADAPTER
+
+The `ADAPTER` instruction specifies the LoRA adapter to apply to the base model. The value of this instruction should be an absolute path or a path relative to the Modelfile and the file must be in a GGML file format. The adapter should be tuned from the base model otherwise the behaviour is undefined.
+
+```
+ADAPTER ./ollama-lora.bin
+```
+
 ### LICENSE
 
 The `LICENSE` instruction allows you to specify the legal license under which the model used with this Modelfile is shared or distributed.
@@ -163,4 +185,4 @@ LICENSE """
 ## Notes
 
 - the **modelfile is not case sensitive**. In the examples, we use uppercase for instructions to make it easier to distinguish it from arguments.
-- Instructions can be in any order. In the examples, we start with FROM instruction to keep it easily readable.
+- Instructions can be in any order. In the examples, we start with FROM instruction to keep it easily readable.

docs/tutorials.md

@@ -0,0 +1,8 @@
+# Tutorials
+
+Here is a list of ways you can use Ollama with other tools to build interesting applications.
+
+- [Using LangChain with Ollama in JavaScript](./tutorials/langchainjs.md)
+- [Using LangChain with Ollama in Python](./tutorials/langchainpy.md)
+
+Also be sure to check out the [examples](../examples) directory for more ways to use Ollama.

docs/tutorials/langchainjs.md

@@ -0,0 +1,73 @@
+# Using LangChain with Ollama using JavaScript
+
+In this tutorial, we are going to use JavaScript with LangChain and Ollama to learn about something just a touch more recent. In August 2023, there was a series of wildfires on Maui. There is no way an LLM trained before that time can know about this, since their training data would not include anything as recent as that. So we can find the [Wikipedia article about the fires](https://en.wikipedia.org/wiki/2023_Hawaii_wildfires) and ask questions about the contents.
+
+To get started, let's just use **LangChain** to ask a simple question to a model. To do this with JavaScript, we need to install **LangChain**:
+
+```bash
+npm install langchain
+```
+
+Now we can start building out our JavaScript:
+
+```javascript
+import { Ollama } from "langchain/llms/ollama";
+
+const ollama = new Ollama({
+  baseUrl: "http://localhost:11434",
+  model: "llama2",
+});
+
+const answer = await ollama.call(`why is the sky blue?`);
+
+console.log(answer);
+```
+
+That will get us the same thing as if we ran `ollama run llama2 "why is the sky blue"` in the terminal. But we want to load a document from the web to ask a question against. **Cheerio** is a great library for ingesting a webpage, and **LangChain** uses it in their **CheerioWebBaseLoader**. So let's build that part of the app.
+
+```javascript
+import { CheerioWebBaseLoader } from "langchain/document_loaders/web/cheerio";
+
+const loader = new CheerioWebBaseLoader("https://en.wikipedia.org/wiki/2023_Hawaii_wildfires");
+const data = loader.load();
+```
+
+That will load the document. Although this page is smaller than the Odyssey, it is certainly bigger than the context size for most LLMs. So we are going to need to split into smaller pieces, and then select just the pieces relevant to our question. This is a great use for a vector datastore. In this example, we will use the **MemoryVectorStore** that is part of **LangChain**. But there is one more thing we need to get the content into the datastore. We have to run an embeddings process that converts the tokens in the text into a series of vectors. And for that, we are going to use **Tensorflow**. There is a lot of stuff going on in this one. First, install the **Tensorflow** components that we need.
+
+```javascript
+npm install @tensorflow/tfjs-core@3.6.0 @tensorflow/tfjs-converter@3.6.0 @tensorflow-models/universal-sentence-encoder@1.3.3 @tensorflow/tfjs-node@4.10.0
+```
+
+If you just install those components without the version numbers, it will install the latest versions, but there are conflicts within **Tensorflow**, so you need to install the compatible versions.
+
+```javascript
+import { RecursiveCharacterTextSplitter } from "langchain/text_splitter"
+import { MemoryVectorStore } from "langchain/vectorstores/memory";
+import "@tensorflow/tfjs-node";
+import { TensorFlowEmbeddings } from "langchain/embeddings/tensorflow";
+
+// Split the text into 500 character chunks. And overlap each chunk by 20 characters
+const textSplitter = new RecursiveCharacterTextSplitter({
+ chunkSize: 500,
+ chunkOverlap: 20
+});
+const splitDocs = await textSplitter.splitDocuments(data);
+
+// Then use the TensorFlow Embedding to store these chunks in the datastore
+const vectorStore = await MemoryVectorStore.fromDocuments(splitDocs, new TensorFlowEmbeddings());
+```
+
+To connect the datastore to a question asked to a LLM, we need to use the concept at the heart of **LangChain**: the chain. Chains are a way to connect a number of activities together to accomplish a particular tasks. There are a number of chain types available, but for this tutorial we are using the **RetrievalQAChain**.
+
+```javascript
+import { RetrievalQAChain } from "langchain/chains";
+
+const retriever = vectorStore.asRetriever();
+const chain = RetrievalQAChain.fromLLM(ollama, retriever);
+const result = await chain.call({query: "When was Hawaii's request for a major disaster declaration approved?"});
+console.log(result.text)
+```
+
+So we created a retriever, which is a way to return the chunks that match a query from a datastore. And then connect the retriever and the model via a chain. Finally, we send a query to the chain, which results in an answer using our document as a source. The answer it returned was correct, August 10, 2023.
+
+And that is a simple introduction to what you can do with **LangChain** and **Ollama.**

docs/tutorials/langchainpy.md

@@ -0,0 +1,81 @@
+# Using LangChain with Ollama in Python
+
+Let's imagine we are studying the classics, such as **the Odyssey** by **Homer**. We might have a question about Neleus and his family. If you ask llama2 for that info, you may get something like:
+
+> I apologize, but I'm a large language model, I cannot provide information on individuals or families that do not exist in reality. Neleus is not a real person or character, and therefore does not have a family or any other personal details. My apologies for any confusion. Is there anything else I can help you with?
+
+This sounds like a typical censored response, but even llama2-uncensored gives a mediocre answer:
+
+> Neleus was a legendary king of Pylos and the father of Nestor, one of the Argonauts. His mother was Clymene, a sea nymph, while his father was Neptune, the god of the sea.
+
+So let's figure out how we can use **LangChain** with Ollama to ask our question to the actual document, the Odyssey by Homer, using Python.
+
+Let's start by asking a simple question that we can get an answer to from the **Llama2** model using **Ollama**. First, we need to install the **LangChain** package:
+
+`pip install langchain`
+
+Then we can create a model and ask the question:
+
+```python
+from langchain.llms import Ollama
+ollama = Ollama(base_url='http://localhost:11434',
+model="llama2")
+print(ollama("why is the sky blue"))
+```
+
+Notice that we are defining the model and the base URL for Ollama.
+
+Now let's load a document to ask questions against. I'll load up the Odyssey by Homer, which you can find at Project Gutenberg. We will need **WebBaseLoader** which is part of **LangChain** and loads text from any webpage. On my machine, I also needed to install **bs4** to get that to work, so run `pip install bs4`.
+
+```python
+from langchain.document_loaders import WebBaseLoader
+loader = WebBaseLoader("https://www.gutenberg.org/files/1727/1727-h/1727-h.htm")
+data = loader.load()
+```
+
+This file is pretty big. Just the preface is 3000 tokens. Which means the full document won't fit into the context for the model. So we need to split it up into smaller pieces.
+
+```python
+from langchain.text_splitter import RecursiveCharacterTextSplitter
+
+text_splitter=RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=0)
+all_splits = text_splitter.split_documents(data)
+```
+
+It's split up, but we have to find the relevant splits and then submit those to the model. We can do this by creating embeddings and storing them in a vector database. For now, we don't have embeddings built in to Ollama, though we will be adding that soon, so for now, we can use the GPT4All library for that. We will use ChromaDB in this example for a vector database. `pip install GPT4All chromadb`
+
+```python
+from langchain.embeddings import GPT4AllEmbeddings
+from langchain.vectorstores import Chroma
+vectorstore = Chroma.from_documents(documents=all_splits, embedding=GPT4AllEmbeddings())
+```
+
+Now let's ask a question from the document. **Who was Neleus, and who is in his family?** Neleus is a character in the Odyssey, and the answer can be found in our text.
+
+```python
+question="Who is Neleus and who is in Neleus' family?"
+docs = vectorstore.similarity_search(question)
+len(docs)
+```
+
+This will output the number of matches for chunks of data similar to the search.
+
+The next thing is to send the question and the relevant parts of the docs to the model to see if we can get a good answer. But we are stitching two parts of the process together, and that is called a chain. This means we need to define a chain:
+
+```python
+from langchain.chains import RetrievalQA
+qachain=RetrievalQA.from_chain_type(ollama, retriever=vectorstore.as_retriever())
+qachain({"query": question})
+```
+
+The answer received from this chain was:
+
+> Neleus is a character in Homer's "Odyssey" and is mentioned in the context of Penelope's suitors. Neleus is the father of Chloris, who is married to Neleus and bears him several children, including Nestor, Chromius, Periclymenus, and Pero. Amphinomus, the son of Nisus, is also mentioned as a suitor of Penelope and is known for his good natural disposition and agreeable conversation.
+
+It's not a perfect answer, as it implies Neleus married his daughter when actually Chloris "was the youngest daughter to Amphion son of Iasus and king of Minyan Orchomenus, and was Queen in Pylos".
+
+I updated the chunk_overlap for the text splitter to 20 and tried again and got a much better answer:
+
+> Neleus is a character in Homer's epic poem "The Odyssey." He is the husband of Chloris, who is the youngest daughter of Amphion son of Iasus and king of Minyan Orchomenus. Neleus has several children with Chloris, including Nestor, Chromius, Periclymenus, and Pero.
+
+And that is a much better answer.

examples/10tweets/Modelfile

@@ -0,0 +1,7 @@
+# Modelfile for creating a list of ten tweets from a topic
+# Run `ollama create 10tweets -f ./Modelfile` and then `ollama run 10tweets` and enter a topic
+
+FROM llama2
+SYSTEM """
+You are a content marketer who needs to come up with 10 short but succinct tweets. The answer should be a list of ten tweets. Each tweet can have a maximum of 280 characters and should include hashtags. Each user input will be a subject and you should expand it in ten creative ways. Never stop after just one tweet. Always include ten. 
+"""

examples/README.md

@@ -1,6 +1,6 @@
 # Examples
 
-This directory contains examples that can be created and run with `ollama`.
+This directory contains different examples of using Ollama
 
 To create a model:
 

examples/dockerit/Modelfile

@@ -0,0 +1,20 @@
+FROM llama2
+SYSTEM """
+You are an experienced Devops engineer focused on docker. When given specifications for a particular need or application you know the best way to host that within a docker container. For instance if someone tells you they want an nginx server to host files located at /web you will answer as follows
+
+---start
+FROM nginx:alpine
+COPY /myweb /usr/share/nginx/html
+EXPOSE 80
+---end
+
+Notice that the answer you should give is just the contents of the dockerfile with no explanation and there are three dashes and the word start at the beginning and 3 dashes and the word end. The full output can be piped into a file and run as is. Here is another example. The user will ask to launch a Postgres server with a password of abc123. And the response should be
+
+---start
+FROM postgres:latest
+ENV POSTGRES_PASSWORD=abc123
+EXPOSE 5432
+---end
+
+Again it's just the contents of the dockerfile and nothing else.
+"""

examples/dockerit/README.md

@@ -0,0 +1,15 @@
+# DockerIt
+
+DockerIt is a tool to help you build and run your application in a Docker container. It consists of a model that defines the system prompt and model weights to use, along with a python script to then build the container and run the image automatically. 
+
+## Caveats
+
+This is an simple example. It's assuming the Dockerfile content generated is going to work. In many cases, even with simple web servers, it fails when trying to copy files that don't exist. It's simply an example of what you could possibly do.
+
+## Example Usage
+
+```bash
+> python3 ./dockerit.py "simple postgres server with admin password set to 123"
+Enter the name of the image: matttest
+Container named happy_keller  started with id:  7c201bb6c30f02b356ddbc8e2a5af9d7d7d7b8c228519c9a501d15c0bd9d6b3e
+```

examples/dockerit/dockerit.py

@@ -0,0 +1,17 @@
+import requests, json, docker, io, sys
+inputDescription = " ".join(sys.argv[1:])
+imageName = input("Enter the name of the image: ")
+client = docker.from_env()
+s = requests.Session()
+output=""
+with s.post('http://localhost:11434/api/generate', json={'model': 'dockerit', 'prompt': inputDescription}, stream=True) as r:
+  for line in r.iter_lines():
+    if line:
+      j = json.loads(line)
+      if "response" in j:
+        output = output +j["response"]
+output = output[output.find("---start")+9:output.find("---end")-1]
+f = io.BytesIO(bytes(output, 'utf-8'))
+client.images.build(fileobj=f, tag=imageName)
+container = client.containers.run(imageName, detach=True)
+print("Container named", container.name, " started with id: ",container.id)

examples/dockerit/requirements.txt

@@ -0,0 +1 @@
+docker

examples/langchain-document/README.md

@@ -0,0 +1,21 @@
+# LangChain Document QA
+
+This example provides an interface for asking questions to a PDF document.
+
+## Setup
+
+```
+pip install -r requirements.txt
+```
+
+## Run
+
+```
+python main.py
+```
+
+A prompt will appear, where questions may be asked:
+
+```
+Query: How many locations does WeWork have?
+```

examples/langchain-document/main.py

@@ -0,0 +1,61 @@
+from langchain.document_loaders import OnlinePDFLoader
+from langchain.vectorstores import Chroma
+from langchain.embeddings import GPT4AllEmbeddings
+from langchain import PromptTemplate
+from langchain.llms import Ollama
+from langchain.callbacks.manager import CallbackManager
+from langchain.callbacks.streaming_stdout import StreamingStdOutCallbackHandler
+from langchain.chains import RetrievalQA
+import sys
+import os
+
+class SuppressStdout:
+    def __enter__(self):
+        self._original_stdout = sys.stdout
+        self._original_stderr = sys.stderr
+        sys.stdout = open(os.devnull, 'w')
+        sys.stderr = open(os.devnull, 'w')
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        sys.stdout.close()
+        sys.stdout = self._original_stdout
+        sys.stderr = self._original_stderr
+
+# load the pdf and split it into chunks
+loader = OnlinePDFLoader("https://d18rn0p25nwr6d.cloudfront.net/CIK-0001813756/975b3e9b-268e-4798-a9e4-2a9a7c92dc10.pdf")
+data = loader.load()
+
+from langchain.text_splitter import RecursiveCharacterTextSplitter
+text_splitter = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=0)
+all_splits = text_splitter.split_documents(data)
+
+with SuppressStdout():
+    vectorstore = Chroma.from_documents(documents=all_splits, embedding=GPT4AllEmbeddings())
+
+while True:
+    query = input("\nQuery: ")
+    if query == "exit":
+        break
+    if query.strip() == "":
+        continue
+
+    # Prompt
+    template = """Use the following pieces of context to answer the question at the end. 
+    If you don't know the answer, just say that you don't know, don't try to make up an answer. 
+    Use three sentences maximum and keep the answer as concise as possible. 
+    {context}
+    Question: {question}
+    Helpful Answer:"""
+    QA_CHAIN_PROMPT = PromptTemplate(
+        input_variables=["context", "question"],
+        template=template,
+    )
+
+    llm = Ollama(model="llama2:13b", callback_manager=CallbackManager([StreamingStdOutCallbackHandler()]))
+    qa_chain = RetrievalQA.from_chain_type(
+        llm,
+        retriever=vectorstore.as_retriever(),
+        chain_type_kwargs={"prompt": QA_CHAIN_PROMPT},
+    )
+
+    result = qa_chain({"query": query})

examples/langchain-document/requirements.txt

@@ -0,0 +1,109 @@
+absl-py==1.4.0
+aiohttp==3.8.5
+aiosignal==1.3.1
+anyio==3.7.1
+astunparse==1.6.3
+async-timeout==4.0.3
+attrs==23.1.0
+backoff==2.2.1
+beautifulsoup4==4.12.2
+bs4==0.0.1
+cachetools==5.3.1
+certifi==2023.7.22
+cffi==1.15.1
+chardet==5.2.0
+charset-normalizer==3.2.0
+Chroma==0.2.0
+chroma-hnswlib==0.7.2
+chromadb==0.4.5
+click==8.1.6
+coloredlogs==15.0.1
+cryptography==41.0.3
+dataclasses-json==0.5.14
+fastapi==0.99.1
+filetype==1.2.0
+flatbuffers==23.5.26
+frozenlist==1.4.0
+gast==0.4.0
+google-auth==2.22.0
+google-auth-oauthlib==1.0.0
+google-pasta==0.2.0
+gpt4all==1.0.8
+grpcio==1.57.0
+h11==0.14.0
+h5py==3.9.0
+httptools==0.6.0
+humanfriendly==10.0
+idna==3.4
+importlib-resources==6.0.1
+joblib==1.3.2
+keras==2.13.1
+langchain==0.0.261
+langsmith==0.0.21
+libclang==16.0.6
+lxml==4.9.3
+Markdown==3.4.4
+MarkupSafe==2.1.3
+marshmallow==3.20.1
+monotonic==1.6
+mpmath==1.3.0
+multidict==6.0.4
+mypy-extensions==1.0.0
+nltk==3.8.1
+numexpr==2.8.5
+numpy==1.24.3
+oauthlib==3.2.2
+onnxruntime==1.15.1
+openapi-schema-pydantic==1.2.4
+opt-einsum==3.3.0
+overrides==7.4.0
+packaging==23.1
+pdf2image==1.16.3
+pdfminer==20191125
+pdfminer.six==20221105
+Pillow==10.0.0
+posthog==3.0.1
+protobuf==4.24.0
+pulsar-client==3.2.0
+pyasn1==0.5.0
+pyasn1-modules==0.3.0
+pycparser==2.21
+pycryptodome==3.18.0
+pydantic==1.10.12
+PyPika==0.48.9
+python-dateutil==2.8.2
+python-dotenv==1.0.0
+python-magic==0.4.27
+PyYAML==6.0.1
+regex==2023.8.8
+requests==2.31.0
+requests-oauthlib==1.3.1
+rsa==4.9
+six==1.16.0
+sniffio==1.3.0
+soupsieve==2.4.1
+SQLAlchemy==2.0.19
+starlette==0.27.0
+sympy==1.12
+tabulate==0.9.0
+tenacity==8.2.2
+tensorboard==2.13.0
+tensorboard-data-server==0.7.1
+tensorflow==2.13.0
+tensorflow-estimator==2.13.0
+tensorflow-hub==0.14.0
+tensorflow-macos==2.13.0
+termcolor==2.3.0
+tokenizers==0.13.3
+tqdm==4.66.1
+typing-inspect==0.9.0
+typing_extensions==4.5.0
+unstructured==0.9.2
+urllib3==1.26.16
+uvicorn==0.23.2
+uvloop==0.17.0
+watchfiles==0.19.0
+websockets==11.0.3
+Werkzeug==2.3.6
+wrapt==1.15.0
+yarl==1.9.2

examples/langchain-web-summary/README.md

@@ -0,0 +1,15 @@
+# LangChain Web Summarization
+
+This example summarizes a website
+
+## Setup
+
+```
+pip install -r requirements.txt
+```
+
+## Run
+
+```
+python main.py
+```

examples/langchain-web-summary/main.py

@@ -0,0 +1,12 @@
+from langchain.llms import Ollama
+from langchain.document_loaders import WebBaseLoader
+from langchain.chains.summarize import load_summarize_chain
+
+loader = WebBaseLoader("https://ollama.ai/blog/run-llama2-uncensored-locally")
+docs = loader.load()
+
+llm = Ollama(model="llama2")
+chain = load_summarize_chain(llm, chain_type="stuff")
+
+result = chain.run(docs)
+print(result)

examples/langchain-web-summary/requirements.txt

@@ -0,0 +1,2 @@
+langchain==0.0.259
+bs4==0.0.1

examples/langchain/README.md

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

examples/langchain/main.py

@@ -0,0 +1,4 @@
+from langchain.llms import Ollama
+llm = Ollama(model="llama2")
+res = llm.predict("hello")
+print (res)

examples/langchain/requirements.txt

@@ -0,0 +1 @@
+langchain==0.0.259

examples/privategpt/.gitignore

@@ -0,0 +1,170 @@
+# OSX
+.DS_STORE
+
+# Models
+models/
+
+# Local Chroma db
+.chroma/
+db/
+
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/#use-with-ide
+.pdm.toml
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+#.idea/

examples/privategpt/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

examples/privategpt/README.md

@@ -0,0 +1,91 @@
+# PrivateGPT with Llama 2 uncensored
+
+https://github.com/jmorganca/ollama/assets/3325447/20cf8ec6-ff25-42c6-bdd8-9be594e3ce1b
+
+> Note: this example is a slightly modified version of PrivateGPT using models such as Llama 2 Uncensored. All credit for PrivateGPT goes to Iván Martínez who is the creator of it, and you can find his GitHub repo [here](https://github.com/imartinez/privateGPT).
+
+### Setup
+
+Set up a virtual environment (optional):
+
+```
+python3 -m venv .venv
+source .venv/bin/activate
+```
+
+Install the Python dependencies:
+
+```shell
+pip install -r requirements.txt
+```
+
+Pull the model you'd like to use:
+
+```
+ollama pull llama2-uncensored
+```
+
+### Getting WeWork's latest quarterly earnings report (10-Q)
+
+```
+mkdir source_documents
+curl https://d18rn0p25nwr6d.cloudfront.net/CIK-0001813756/975b3e9b-268e-4798-a9e4-2a9a7c92dc10.pdf -o source_documents/wework.pdf
+```
+
+### Ingesting files
+
+```shell
+python ingest.py
+```
+
+Output should look like this:
+
+```shell
+Creating new vectorstore
+Loading documents from source_documents
+Loading new documents: 100%|██████████████████████| 1/1 [00:01<00:00,  1.73s/it]
+Loaded 1 new documents from source_documents
+Split into 90 chunks of text (max. 500 tokens each)
+Creating embeddings. May take some minutes...
+Using embedded DuckDB with persistence: data will be stored in: db
+Ingestion complete! You can now run privateGPT.py to query your documents
+```
+
+### Ask questions
+
+```shell
+python privateGPT.py
+
+Enter a query: How many locations does WeWork have?
+
+> Answer (took 17.7 s.):
+As of June 2023, WeWork has 777 locations worldwide, including 610 Consolidated Locations (as defined in the section entitled Key Performance Indicators).
+```
+
+### Try a different model:
+
+```
+ollama pull llama2:13b
+MODEL=llama2:13b python privateGPT.py
+```
+
+## Adding more files
+
+Put any and all your files into the `source_documents` directory
+
+The supported extensions are:
+
+- `.csv`: CSV,
+- `.docx`: Word Document,
+- `.doc`: Word Document,
+- `.enex`: EverNote,
+- `.eml`: Email,
+- `.epub`: EPub,
+- `.html`: HTML File,
+- `.md`: Markdown,
+- `.msg`: Outlook Message,
+- `.odt`: Open Document Text,
+- `.pdf`: Portable Document Format (PDF),
+- `.pptx` : PowerPoint Document,
+- `.ppt` : PowerPoint Document,
+- `.txt`: Text file (UTF-8),

examples/privategpt/constants.py

@@ -0,0 +1,12 @@
+import os
+from chromadb.config import Settings
+
+# Define the folder for storing database
+PERSIST_DIRECTORY = os.environ.get('PERSIST_DIRECTORY', 'db')
+
+# Define the Chroma settings
+CHROMA_SETTINGS = Settings(
+        chroma_db_impl='duckdb+parquet',
+        persist_directory=PERSIST_DIRECTORY,
+        anonymized_telemetry=False
+)

examples/privategpt/ingest.py

@@ -0,0 +1,161 @@
+#!/usr/bin/env python3
+import os
+import glob
+from typing import List
+from multiprocessing import Pool
+from tqdm import tqdm
+
+from langchain.document_loaders import (
+    CSVLoader,
+    EverNoteLoader,
+    PyMuPDFLoader,
+    TextLoader,
+    UnstructuredEmailLoader,
+    UnstructuredEPubLoader,
+    UnstructuredHTMLLoader,
+    UnstructuredMarkdownLoader,
+    UnstructuredODTLoader,
+    UnstructuredPowerPointLoader,
+    UnstructuredWordDocumentLoader,
+)
+
+from langchain.text_splitter import RecursiveCharacterTextSplitter
+from langchain.vectorstores import Chroma
+from langchain.embeddings import HuggingFaceEmbeddings
+from langchain.docstore.document import Document
+from constants import CHROMA_SETTINGS
+
+
+# Load environment variables
+persist_directory = os.environ.get('PERSIST_DIRECTORY', 'db')
+source_directory = os.environ.get('SOURCE_DIRECTORY', 'source_documents')
+embeddings_model_name = os.environ.get('EMBEDDINGS_MODEL_NAME', 'all-MiniLM-L6-v2')
+chunk_size = 500
+chunk_overlap = 50
+
+# Custom document loaders
+class MyElmLoader(UnstructuredEmailLoader):
+    """Wrapper to fallback to text/plain when default does not work"""
+
+    def load(self) -> List[Document]:
+        """Wrapper adding fallback for elm without html"""
+        try:
+            try:
+                doc = UnstructuredEmailLoader.load(self)
+            except ValueError as e:
+                if 'text/html content not found in email' in str(e):
+                    # Try plain text
+                    self.unstructured_kwargs["content_source"]="text/plain"
+                    doc = UnstructuredEmailLoader.load(self)
+                else:
+                    raise
+        except Exception as e:
+            # Add file_path to exception message
+            raise type(e)(f"{self.file_path}: {e}") from e
+
+        return doc
+
+
+# Map file extensions to document loaders and their arguments
+LOADER_MAPPING = {
+    ".csv": (CSVLoader, {}),
+    # ".docx": (Docx2txtLoader, {}),
+    ".doc": (UnstructuredWordDocumentLoader, {}),
+    ".docx": (UnstructuredWordDocumentLoader, {}),
+    ".enex": (EverNoteLoader, {}),
+    ".eml": (MyElmLoader, {}),
+    ".epub": (UnstructuredEPubLoader, {}),
+    ".html": (UnstructuredHTMLLoader, {}),
+    ".md": (UnstructuredMarkdownLoader, {}),
+    ".odt": (UnstructuredODTLoader, {}),
+    ".pdf": (PyMuPDFLoader, {}),
+    ".ppt": (UnstructuredPowerPointLoader, {}),
+    ".pptx": (UnstructuredPowerPointLoader, {}),
+    ".txt": (TextLoader, {"encoding": "utf8"}),
+    # Add more mappings for other file extensions and loaders as needed
+}
+
+
+def load_single_document(file_path: str) -> List[Document]:
+    ext = "." + file_path.rsplit(".", 1)[-1]
+    if ext in LOADER_MAPPING:
+        loader_class, loader_args = LOADER_MAPPING[ext]
+        loader = loader_class(file_path, **loader_args)
+        return loader.load()
+
+    raise ValueError(f"Unsupported file extension '{ext}'")
+
+def load_documents(source_dir: str, ignored_files: List[str] = []) -> List[Document]:
+    """
+    Loads all documents from the source documents directory, ignoring specified files
+    """
+    all_files = []
+    for ext in LOADER_MAPPING:
+        all_files.extend(
+            glob.glob(os.path.join(source_dir, f"**/*{ext}"), recursive=True)
+        )
+    filtered_files = [file_path for file_path in all_files if file_path not in ignored_files]
+
+    with Pool(processes=os.cpu_count()) as pool:
+        results = []
+        with tqdm(total=len(filtered_files), desc='Loading new documents', ncols=80) as pbar:
+            for i, docs in enumerate(pool.imap_unordered(load_single_document, filtered_files)):
+                results.extend(docs)
+                pbar.update()
+
+    return results
+
+def process_documents(ignored_files: List[str] = []) -> List[Document]:
+    """
+    Load documents and split in chunks
+    """
+    print(f"Loading documents from {source_directory}")
+    documents = load_documents(source_directory, ignored_files)
+    if not documents:
+        print("No new documents to load")
+        exit(0)
+    print(f"Loaded {len(documents)} new documents from {source_directory}")
+    text_splitter = RecursiveCharacterTextSplitter(chunk_size=chunk_size, chunk_overlap=chunk_overlap)
+    texts = text_splitter.split_documents(documents)
+    print(f"Split into {len(texts)} chunks of text (max. {chunk_size} tokens each)")
+    return texts
+
+def does_vectorstore_exist(persist_directory: str) -> bool:
+    """
+    Checks if vectorstore exists
+    """
+    if os.path.exists(os.path.join(persist_directory, 'index')):
+        if os.path.exists(os.path.join(persist_directory, 'chroma-collections.parquet')) and os.path.exists(os.path.join(persist_directory, 'chroma-embeddings.parquet')):
+            list_index_files = glob.glob(os.path.join(persist_directory, 'index/*.bin'))
+            list_index_files += glob.glob(os.path.join(persist_directory, 'index/*.pkl'))
+            # At least 3 documents are needed in a working vectorstore
+            if len(list_index_files) > 3:
+                return True
+    return False
+
+def main():
+    # Create embeddings
+    embeddings = HuggingFaceEmbeddings(model_name=embeddings_model_name)
+
+    if does_vectorstore_exist(persist_directory):
+        # Update and store locally vectorstore
+        print(f"Appending to existing vectorstore at {persist_directory}")
+        db = Chroma(persist_directory=persist_directory, embedding_function=embeddings, client_settings=CHROMA_SETTINGS)
+        collection = db.get()
+        texts = process_documents([metadata['source'] for metadata in collection['metadatas']])
+        print(f"Creating embeddings. May take some minutes...")
+        db.add_documents(texts)
+    else:
+        # Create and store locally vectorstore
+        print("Creating new vectorstore")
+        texts = process_documents()
+        print(f"Creating embeddings. May take some minutes...")
+        db = Chroma.from_documents(texts, embeddings, persist_directory=persist_directory, client_settings=CHROMA_SETTINGS)
+    db.persist()
+    db = None
+
+    print(f"Ingestion complete! You can now run privateGPT.py to query your documents")
+
+
+if __name__ == "__main__":
+    main()

examples/privategpt/privateGPT.py

@@ -0,0 +1,71 @@
+#!/usr/bin/env python3
+from langchain.chains import RetrievalQA
+from langchain.embeddings import HuggingFaceEmbeddings
+from langchain.callbacks.streaming_stdout import StreamingStdOutCallbackHandler
+from langchain.vectorstores import Chroma
+from langchain.llms import Ollama
+import os
+import argparse
+import time
+
+model = os.environ.get("MODEL", "llama2-uncensored")
+# For embeddings model, the example uses a sentence-transformers model
+# https://www.sbert.net/docs/pretrained_models.html 
+# "The all-mpnet-base-v2 model provides the best quality, while all-MiniLM-L6-v2 is 5 times faster and still offers good quality."
+embeddings_model_name = os.environ.get("EMBEDDINGS_MODEL_NAME", "all-MiniLM-L6-v2")
+persist_directory = os.environ.get("PERSIST_DIRECTORY", "db")
+target_source_chunks = int(os.environ.get('TARGET_SOURCE_CHUNKS',4))
+
+from constants import CHROMA_SETTINGS
+
+def main():
+    # Parse the command line arguments
+    args = parse_arguments()
+    embeddings = HuggingFaceEmbeddings(model_name=embeddings_model_name)
+    db = Chroma(persist_directory=persist_directory, embedding_function=embeddings, client_settings=CHROMA_SETTINGS)
+    retriever = db.as_retriever(search_kwargs={"k": target_source_chunks})
+    # activate/deactivate the streaming StdOut callback for LLMs
+    callbacks = [] if args.mute_stream else [StreamingStdOutCallbackHandler()]
+
+    llm = Ollama(model=model, callbacks=callbacks)
+
+    qa = RetrievalQA.from_chain_type(llm=llm, chain_type="stuff", retriever=retriever, return_source_documents= not args.hide_source)
+    # Interactive questions and answers
+    while True:
+        query = input("\nEnter a query: ")
+        if query == "exit":
+            break
+        if query.strip() == "":
+            continue
+
+        # Get the answer from the chain
+        start = time.time()
+        res = qa(query)
+        answer, docs = res['result'], [] if args.hide_source else res['source_documents']
+        end = time.time()
+
+        # Print the result
+        print("\n\n> Question:")
+        print(query)
+        print(answer)
+
+        # Print the relevant sources used for the answer
+        for document in docs:
+            print("\n> " + document.metadata["source"] + ":")
+            print(document.page_content)
+
+def parse_arguments():
+    parser = argparse.ArgumentParser(description='privateGPT: Ask questions to your documents without an internet connection, '
+                                                 'using the power of LLMs.')
+    parser.add_argument("--hide-source", "-S", action='store_true',
+                        help='Use this flag to disable printing of source documents used for answers.')
+
+    parser.add_argument("--mute-stream", "-M",
+                        action='store_true',
+                        help='Use this flag to disable the streaming StdOut callback for LLMs.')
+
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    main()

examples/privategpt/pyproject.toml

@@ -0,0 +1,26 @@
+[tool.poetry]
+name = "privategpt"
+version = "0.1.0"
+description = ""
+authors = ["Ivan Martinez <ivanmartit@gmail.com>"]
+license = "Apache Version 2.0"
+readme = "README.md"
+
+[tool.poetry.dependencies]
+python = "^3.10"
+langchain = "0.0.261"
+gpt4all = "^1.0.3"
+chromadb = "^0.3.26"
+PyMuPDF = "^1.22.5"
+python-dotenv = "^1.0.0"
+unstructured = "^0.8.0"
+extract-msg = "^0.41.5"
+tabulate = "^0.9.0"
+pandoc = "^2.3"
+pypandoc = "^1.11"
+tqdm = "^4.65.0"
+sentence-transformers = "^2.2.2"
+
+[build-system]
+requires = ["poetry-core"]
+build-backend = "poetry.core.masonry.api"

examples/python/client.py

@@ -0,0 +1,38 @@
+import json
+import requests
+
+# NOTE: ollama must be running for this to work, start the ollama app or run `ollama serve`
+model = 'llama2' # TODO: update this for whatever model you wish to use
+
+def generate(prompt, context):
+    r = requests.post('http://localhost:11434/api/generate',
+                      json={
+                          'model': model,
+                          'prompt': prompt,
+                          'context': context,
+                      },
+                      stream=True)
+    r.raise_for_status()
+
+    for line in r.iter_lines():
+        body = json.loads(line)
+        response_part = body.get('response', '')
+        # the response streams one token at a time, print that as we recieve it
+        print(response_part, end='', flush=True)
+
+        if 'error' in body:
+            raise Exception(body['error'])
+
+        if body.get('done', False):
+            return body['context']
+
+def main():
+    context = [] # the context stores a conversation history, you can use this to make the model more context aware
+    while True:
+        user_input = input("Enter a prompt: ")
+        print()
+        context = generate(user_input, context)
+        print()
+
+if __name__ == "__main__":
+    main()

examples/sentiments/Modelfile

@@ -0,0 +1,28 @@
+# Modelfile for creating a sentiment analyzer. 
+# Run `ollama create sentiments -f pathtofile` and then `ollama run sentiments` and enter a topic
+
+FROM orca
+TEMPLATE """
+{{- if .First }}
+### System:
+{{ .System }}
+{{- end }}
+### User: 
+I hate it when my phone dies
+### Response: 
+NEGATIVE
+### User: 
+He is awesome
+### Response: 
+POSITIVE
+### User: 
+This is the link to the article
+### Response: 
+NEUTRAL
+### User:
+{{ .Prompt }}
+
+### Response:
+"""
+
+SYSTEM """You are a sentiment analyzer. You will receive text and output only one word, either POSITIVE or NEGATIVE or NEUTRAL, depending on the sentiment of the text."""

examples/sentiments/Readme.md

@@ -0,0 +1,25 @@
+# Sentiments Modelfile
+
+This is a simple sentiments analyzer using the Orca model. When you pull Orca from the registry, it has a Template already defined that looks like this:
+
+```Modelfile
+{{- if .First }}
+### System:
+{{ .System }}
+{{- end }}
+
+### User:
+{{ .Prompt }}
+
+### Response:
+```
+
+If we just wanted to have the text:
+
+```Plaintext
+You are a sentiment analyzer. You will receive text and output only one word, either POSITIVE or NEGATIVE or NEUTRAL, depending on the sentiment of the text.
+```
+
+then we could have put this in a SYSTEM block. But we want to provide examples which require updating the full Template. Any Modelfile you create will inherit all the settings from the source model. But in this example, we are overriding the Template.
+
+When providing examples for the input and output, you should include the way the model usually provides information. Since the Orca model expects a user prompt to appear after ### User: and the response is after ### Response, we should format our examples like that as well. If we were using the Llama 2 model, the format would be a bit different.

format/openssh.go

@@ -0,0 +1,102 @@
+// Copyright 2012 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Code originally from https://go-review.googlesource.com/c/crypto/+/218620
+
+// TODO: replace with upstream once the above change is merged and released.
+
+package format
+
+import (
+	"crypto"
+	"crypto/ed25519"
+	"crypto/rand"
+	"encoding/binary"
+	"encoding/pem"
+	"fmt"
+
+	"golang.org/x/crypto/ssh"
+)
+
+const privateKeyAuthMagic = "openssh-key-v1\x00"
+
+type openSSHEncryptedPrivateKey struct {
+	CipherName string
+	KDFName    string
+	KDFOptions string
+	KeysCount  uint32
+	PubKey     []byte
+	KeyBlocks  []byte
+}
+
+type openSSHPrivateKey struct {
+	Check1  uint32
+	Check2  uint32
+	Keytype string
+	Rest    []byte `ssh:"rest"`
+}
+
+type openSSHEd25519PrivateKey struct {
+	Pub     []byte
+	Priv    []byte
+	Comment string
+	Pad     []byte `ssh:"rest"`
+}
+
+func OpenSSHPrivateKey(key crypto.PrivateKey, comment string) (*pem.Block, error) {
+	var check uint32
+	if err := binary.Read(rand.Reader, binary.BigEndian, &check); err != nil {
+		return nil, err
+	}
+
+	var pk1 openSSHPrivateKey
+	pk1.Check1 = check
+	pk1.Check2 = check
+
+	var w openSSHEncryptedPrivateKey
+	w.KeysCount = 1
+
+	if k, ok := key.(*ed25519.PrivateKey); ok {
+		key = *k
+	}
+
+	switch k := key.(type) {
+	case ed25519.PrivateKey:
+		pub, priv := k[32:], k
+		key := openSSHEd25519PrivateKey{
+			Pub:     pub,
+			Priv:    priv,
+			Comment: comment,
+		}
+
+		pk1.Keytype = ssh.KeyAlgoED25519
+		pk1.Rest = ssh.Marshal(key)
+
+		w.PubKey = ssh.Marshal(struct {
+			KeyType string
+			Pub     []byte
+		}{
+			ssh.KeyAlgoED25519, pub,
+		})
+	default:
+		return nil, fmt.Errorf("ssh: unknown key type %T", k)
+	}
+
+	w.KeyBlocks = openSSHPadding(ssh.Marshal(pk1), 8)
+
+	w.CipherName, w.KDFName, w.KDFOptions = "none", "none", ""
+
+	return &pem.Block{
+		Type:  "OPENSSH PRIVATE KEY",
+		Bytes: append([]byte(privateKeyAuthMagic), ssh.Marshal(w)...),
+	}, nil
+}
+
+func openSSHPadding(block []byte, blocksize int) []byte {
+	for i, j := 0, len(block); (j+i)%blocksize != 0; i++ {
+		block = append(block, byte(i+1))
+	}
+
+	return block
+}

go.mod

@@ -32,14 +32,16 @@ require (
 	github.com/mattn/go-isatty v0.0.19 // indirect
 	github.com/modern-go/concurrent v0.0.0-20180306012644-bacd9c7ef1dd // indirect
 	github.com/modern-go/reflect2 v1.0.2 // indirect
+	github.com/pbnjay/memory v0.0.0-20210728143218-7b4eea64cf58
 	github.com/pelletier/go-toml/v2 v2.0.8 // indirect
 	github.com/spf13/pflag v1.0.5 // indirect
 	github.com/twitchyliquid64/golang-asm v0.15.1 // indirect
 	github.com/ugorji/go/codec v1.2.11 // indirect
 	golang.org/x/arch v0.3.0 // indirect
-	golang.org/x/crypto v0.10.0 // indirect
+	golang.org/x/crypto v0.10.0
+	golang.org/x/exp v0.0.0-20230817173708-d852ddb80c63
 	golang.org/x/net v0.10.0 // indirect
-	golang.org/x/sys v0.10.0 // indirect
+	golang.org/x/sys v0.11.0 // indirect
 	golang.org/x/term v0.10.0
 	golang.org/x/text v0.10.0 // indirect
 	gonum.org/v1/gonum v0.13.0

go.sum

@@ -78,6 +78,8 @@ github.com/modern-go/reflect2 v1.0.2 h1:xBagoLtFs94CBntxluKeaWgTMpvLxC4ur3nMaC9G
 github.com/modern-go/reflect2 v1.0.2/go.mod h1:yWuevngMOJpCy52FWWMvUC8ws7m/LJsjYzDa0/r8luk=
 github.com/olekukonko/tablewriter v0.0.5 h1:P2Ga83D34wi1o9J6Wh1mRuqd4mF/x/lgBS7N7AbDhec=
 github.com/olekukonko/tablewriter v0.0.5/go.mod h1:hPp6KlRPjbx+hW8ykQs1w3UBbZlj6HuIJcUGPhkA7kY=
+github.com/pbnjay/memory v0.0.0-20210728143218-7b4eea64cf58 h1:onHthvaw9LFnH4t2DcNVpwGmV9E1BkGknEliJkfwQj0=
+github.com/pbnjay/memory v0.0.0-20210728143218-7b4eea64cf58/go.mod h1:DXv8WO4yhMYhSNPKjeNKa5WY9YCIEBRbNzFFPJbWO6Y=
 github.com/pelletier/go-toml/v2 v2.0.1/go.mod h1:r9LEWfGN8R5k0VXJ+0BkIe7MYkRdwZOjgMj2KwnJFUo=
 github.com/pelletier/go-toml/v2 v2.0.8 h1:0ctb6s9mE31h0/lhu+J6OPmVeDxJn+kYnJc2jZR9tGQ=
 github.com/pelletier/go-toml/v2 v2.0.8/go.mod h1:vuYfssBdrU2XDZ9bYydBu6t+6a6PYNcZljzZR9VXg+4=
@@ -118,6 +120,9 @@ golang.org/x/arch v0.3.0/go.mod h1:5om86z9Hs0C8fWVUuoMHwpExlXzs5Tkyp9hOrfG7pp8=
 golang.org/x/crypto v0.0.0-20210711020723-a769d52b0f97/go.mod h1:GvvjBRRGRdwPK5ydBHafDWAxML/pGHZbMvKqRZ5+Abc=
 golang.org/x/crypto v0.10.0 h1:LKqV2xt9+kDzSTfOhx4FrkEBcMrAgHSYgzywV9zcGmM=
 golang.org/x/crypto v0.10.0/go.mod h1:o4eNf7Ede1fv+hwOwZsTHl9EsPFO6q6ZvYR8vYfY45I=
+golang.org/x/exp v0.0.0-20230321023759-10a507213a29 h1:ooxPy7fPvB4kwsA2h+iBNHkAbp/4JxTSwCmvdjEYmug=
+golang.org/x/exp v0.0.0-20230817173708-d852ddb80c63 h1:m64FZMko/V45gv0bNmrNYoDEq8U5YUhetc9cBWKS1TQ=
+golang.org/x/exp v0.0.0-20230817173708-d852ddb80c63/go.mod h1:0v4NqG35kSWCMzLaMeX+IQrlSnVE/bqGSyC2cz/9Le8=
 golang.org/x/net v0.0.0-20210226172049-e18ecbb05110/go.mod h1:m0MpNAwzfU5UDzcl9v0D8zg8gWTRqZa9RBIspLL5mdg=
 golang.org/x/net v0.10.0 h1:X2//UzNDwYmtCLn7To6G58Wr6f5ahEAQgKNzv9Y951M=
 golang.org/x/net v0.10.0/go.mod h1:0qNGK6F8kojg2nk9dLZ2mShWaEBan6FAoqfSigmmuDg=
@@ -128,8 +133,8 @@ golang.org/x/sys v0.0.0-20210806184541-e5e7981a1069/go.mod h1:oPkhp1MJrh7nUepCBc
 golang.org/x/sys v0.0.0-20220310020820-b874c991c1a5/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.0.0-20220704084225-05e143d24a9e/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.6.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
-golang.org/x/sys v0.10.0 h1:SqMFp9UcQJZa+pmYuAKjd9xq1f0j5rLcDIk0mj4qAsA=
-golang.org/x/sys v0.10.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
+golang.org/x/sys v0.11.0 h1:eG7RXZHdqOJ1i+0lgLgCpSXAp6M3LYlAo6osgSi0xOM=
+golang.org/x/sys v0.11.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/term v0.0.0-20201126162022-7de9c90e9dd1/go.mod h1:bj7SfCRtBDWHUb9snDiAeCFNEtKQo2Wmx5Cou7ajbmo=
 golang.org/x/term v0.10.0 h1:3R7pNqamzBraeqj/Tj8qt1aQ2HpmlC+Cx/qL/7hn4/c=
 golang.org/x/term v0.10.0/go.mod h1:lpqdcUyK/oCiQxvxVrppt5ggO2KCZ5QblwqPnfZ6d5o=

llama/ggml-alloc.c

@@ -1,567 +0,0 @@
-/**
- * llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12
- *
- * MIT License
- *
- * Copyright (c) 2023 Georgi Gerganov
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#include "ggml-alloc.h"
-#include "ggml.h"
-#include <assert.h>
-#include <stdarg.h>
-#include <stdio.h>
-#include <stdlib.h>
-#include <string.h>
-
-#define UNUSED(x) (void)(x)
-#define MAX(a, b) ((a) > (b) ? (a) : (b))
-
-//#define GGML_ALLOCATOR_DEBUG
-
-//#define AT_PRINTF printf
-#define AT_PRINTF(...) ((void)0)
-
-struct hash_node {
-    struct ggml_tensor * t;
-    int n_children;
-    int n_views;
-};
-
-static size_t hash(void * p) {
-    return (size_t)p % GGML_GRAPH_HASHTABLE_SIZE;
-}
-
-static struct hash_node * hash_get(struct hash_node hash_table[], struct ggml_tensor * t) {
-    size_t h = hash(t);
-
-    // linear probing
-    size_t i = h;
-    while (hash_table[i].t != NULL) {
-        if (hash_table[i].t == t) {
-            return &hash_table[i];
-        }
-        i = (i + 1) % GGML_GRAPH_HASHTABLE_SIZE;
-        if (i == h) {
-            // hash table is full
-            GGML_ASSERT(false);
-        }
-    }
-
-    hash_table[i].t = t;
-    return &hash_table[i];
-}
-
-// TODO: GGML_PAD ?
-static size_t aligned_offset(const void * buffer, size_t offset, size_t alignment) {
-    assert(alignment && !(alignment & (alignment - 1))); // power of 2
-    size_t align = (alignment - (((uintptr_t)buffer + offset) % alignment)) % alignment;
-    return offset + align;
-}
-
-struct free_block {
-    void * addr;
-    size_t size;
-};
-
-#define MAX_FREE_BLOCKS 128
-
-struct ggml_allocr {
-    void * data;
-    size_t size;
-    size_t alignment;
-    int n_free_blocks;
-    struct free_block free_blocks[MAX_FREE_BLOCKS];
-    struct hash_node hash_table[GGML_GRAPH_HASHTABLE_SIZE];
-    size_t max_size;
-    bool measure;
-
-#ifdef GGML_ALLOCATOR_DEBUG
-    struct ggml_tensor * allocated_tensors[1024];
-#endif
-};
-
-#ifdef GGML_ALLOCATOR_DEBUG
-static void add_allocated_tensor(struct ggml_allocator * alloc, struct ggml_tensor * tensor) {
-    for (int i = 0; i < 1024; i++) {
-        if (alloc->allocated_tensors[i] == NULL) {
-            alloc->allocated_tensors[i] = tensor;
-            return;
-        }
-    }
-    GGML_ASSERT(!"out of allocated_tensors");
-}
-static void remove_allocated_tensor(struct ggml_allocator * alloc, struct ggml_tensor * tensor) {
-    for (int i = 0; i < 1024; i++) {
-        if (alloc->allocated_tensors[i] == tensor ||
-            (alloc->allocated_tensors[i] != NULL && alloc->allocated_tensors[i]->data == tensor->data)) {
-            alloc->allocated_tensors[i] = NULL;
-            return;
-        }
-    }
-    printf("tried to free tensor %s not found\n", tensor->name);
-    GGML_ASSERT(!"tensor not found");
-}
-#endif
-
-
-static size_t ggml_allocator_get_alloc_size(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
-    return ggml_nbytes(tensor);
-
-    UNUSED(alloc);
-}
-
-void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
-    size_t size = ggml_allocator_get_alloc_size(alloc, tensor);
-    size = aligned_offset(NULL, size, alloc->alignment);
-
-    AT_PRINTF("%s: allocating %s (%zu bytes) - ", __func__, tensor->name, size);
-
-    size_t max_avail = 0;
-
-    // find the best fitting free block
-    int best_fit_block = -1;
-    size_t best_fit_size = SIZE_MAX;
-    for (int i = 0; i < alloc->n_free_blocks; i++) {
-        struct free_block * block = &alloc->free_blocks[i];
-        max_avail = MAX(max_avail, block->size);
-        if (block->size >= size && block->size <= best_fit_size) {
-            best_fit_block = i;
-            best_fit_size = block->size;
-        }
-    }
-
-    AT_PRINTF("block %d\n", best_fit_block);
-
-    if (best_fit_block == -1) {
-        fprintf(stderr, "%s: not enough space in the buffer (needed %zu, largest block available %zu)\n",
-                __func__, size, max_avail);
-        GGML_ASSERT(!"not enough space in the buffer");
-        return;
-    }
-    struct free_block * block = &alloc->free_blocks[best_fit_block];
-    void * addr = block->addr;
-    block->addr = (char*)block->addr + size;
-    block->size -= size;
-    if (block->size == 0) {
-        // remove block if empty
-        alloc->n_free_blocks--;
-        for (int j = best_fit_block; j < alloc->n_free_blocks; j++) {
-            alloc->free_blocks[j] = alloc->free_blocks[j+1];
-        }
-    }
-
-    tensor->data = addr;
-
-#ifdef GGML_ALLOCATOR_DEBUG
-    add_allocated_tensor(alloc, tensor);
-    size_t cur_max = (char*)addr - (char*)alloc->data + size;
-    if (cur_max > alloc->max_size) {
-        printf("max_size = %.2f MB: tensors: ", cur_max / 1024.0 / 1024.0);
-        for (int i = 0; i < 1024; i++) {
-            if (alloc->allocated_tensors[i]) {
-                printf("%s (%.2f MB) ", alloc->allocated_tensors[i]->name, ggml_nbytes(alloc->allocated_tensors[i]) / 1024.0 / 1024.0);
-            }
-        }
-        printf("\n");
-    }
-#endif
-
-    alloc->max_size = MAX(alloc->max_size, (char*)addr - (char*)alloc->data + size);
-}
-
-// this is a very naive implementation, but for our case the number of free blocks should be very small
-static void ggml_allocator_free_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
-    void * ptr = tensor->data;
-
-    if (ptr < alloc->data || (char*)ptr >= (char*)alloc->data + alloc->max_size) {
-        // the tensor was not allocated in this buffer
-        // this can happen because the graph allocator will try to free weights and other tensors from different buffers
-        // the easiest way to deal with this is just to ignore it
-        return;
-    }
-
-    size_t size = ggml_allocator_get_alloc_size(alloc, tensor);
-    size = aligned_offset(NULL, size, alloc->alignment);
-    AT_PRINTF("%s: freeing %s (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, size, alloc->n_free_blocks);
-
-#ifdef GGML_ALLOCATOR_DEBUG
-    remove_allocated_tensor(alloc, tensor);
-#endif
-
-    // see if we can merge with an existing block
-    for (int i = 0; i < alloc->n_free_blocks; i++) {
-        struct free_block * block = &alloc->free_blocks[i];
-        // check if ptr is at the end of the block
-        if ((char*)block->addr + block->size == ptr) {
-            block->size += size;
-            // check if we can merge with the next block
-            if (i < alloc->n_free_blocks - 1 && (char*)block->addr + block->size == alloc->free_blocks[i+1].addr) {
-                block->size += alloc->free_blocks[i+1].size;
-                alloc->n_free_blocks--;
-                for (int j = i+1; j < alloc->n_free_blocks; j++) {
-                    alloc->free_blocks[j] = alloc->free_blocks[j+1];
-                }
-            }
-            return;
-        }
-        // check if ptr is at the beginning of the block
-        if ((char*)ptr + size == block->addr) {
-            block->addr = ptr;
-            block->size += size;
-            // check if we can merge with the previous block
-            if (i > 0 && (char*)alloc->free_blocks[i-1].addr + alloc->free_blocks[i-1].size == block->addr) {
-                alloc->free_blocks[i-1].size += block->size;
-                alloc->n_free_blocks--;
-                for (int j = i; j < alloc->n_free_blocks; j++) {
-                    alloc->free_blocks[j] = alloc->free_blocks[j+1];
-                }
-            }
-            return;
-        }
-    }
-    // otherwise, add a new block
-    GGML_ASSERT(alloc->n_free_blocks < MAX_FREE_BLOCKS && "out of free blocks");
-    // insert the new block in the correct position to keep the array sorted by address (to make merging blocks faster)
-    int insert_pos = 0;
-    while (insert_pos < alloc->n_free_blocks && alloc->free_blocks[insert_pos].addr < ptr) {
-        insert_pos++;
-    }
-    // shift all blocks from insert_pos onward to make room for the new block
-    for (int i = alloc->n_free_blocks; i > insert_pos; i--) {
-        alloc->free_blocks[i] = alloc->free_blocks[i-1];
-    }
-    // insert the new block
-    alloc->free_blocks[insert_pos].addr = ptr;
-    alloc->free_blocks[insert_pos].size = size;
-    alloc->n_free_blocks++;
-}
-
-void ggml_allocr_reset(struct ggml_allocr * alloc) {
-    alloc->n_free_blocks = 1;
-    size_t align_offset = aligned_offset(alloc->data, 0, alloc->alignment);
-    alloc->free_blocks[0].addr = (char *)alloc->data + align_offset;
-    alloc->free_blocks[0].size = alloc->size - align_offset;
-}
-
-struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment) {
-    struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
-
-    *alloc = (struct ggml_allocr){
-        /*.data          = */ data,
-        /*.size          = */ size,
-        /*.alignment     = */ alignment,
-        /*.n_free_blocks = */ 0,
-        /*.free_blocks   = */ {{0}},
-        /*.hash_table    = */ {{0}},
-        /*.max_size      = */ 0,
-        /*.measure       = */ false,
-#ifdef GGML_ALLOCATOR_DEBUG
-        /*.allocated_tensors = */ = {0},
-#endif
-    };
-
-    ggml_allocr_reset(alloc);
-
-    return alloc;
-}
-
-// address and size of the buffer when measuring
-// it needs to be large enough to fit all the tensors, but it cannot overlap with other existing buffers
-static void * const MEASURE_BASE_ADDR = (void *) 0x1000;
-static const size_t MEASURE_MAX_SIZE  = 1ULL<<40; // 1 TB
-
-struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) {
-    struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
-
-    *alloc = (struct ggml_allocr){
-        /*.data          = */ MEASURE_BASE_ADDR,
-        /*.size          = */ MEASURE_MAX_SIZE,
-        /*.alignment     = */ alignment,
-        /*.n_free_blocks = */ 0,
-        /*.free_blocks   = */ {{0}},
-        /*.hash_table    = */ {{0}},
-        /*.max_size      = */ 0,
-        /*.measure       = */ true,
-#ifdef GGML_ALLOCATOR_DEBUG
-        /*.allocated_tensors = */ = {0},
-#endif
-    };
-
-    ggml_allocr_reset(alloc);
-
-    return alloc;
-}
-
-void ggml_allocr_free(struct ggml_allocr * alloc) {
-    free(alloc);
-}
-
-bool ggml_allocr_is_measure(struct ggml_allocr * alloc) {
-    return alloc->measure;
-}
-
-//////////// compute graph allocator
-
-static bool ggml_is_view(struct ggml_tensor * t) {
-    return t->op == GGML_OP_RESHAPE || t->op == GGML_OP_VIEW || t->op == GGML_OP_TRANSPOSE ||
-           t->op == GGML_OP_PERMUTE || t->op == GGML_OP_CPY;
-}
-
-static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
-    if (a->type != b->type) {
-        return false;
-    }
-    for (int i = 0; i < GGML_MAX_DIMS; i++) {
-        if (a->ne[i] != b->ne[i]) {
-            return false;
-        }
-        if (a->nb[i] != b->nb[i]) {
-            return false;
-        }
-    }
-    return true;
-}
-
-static struct ggml_tensor * get_view_parent(struct ggml_tensor * t) {
-    switch (t->op) {
-        case GGML_OP_PERMUTE:
-        case GGML_OP_RESHAPE:
-        case GGML_OP_TRANSPOSE:
-        case GGML_OP_VIEW:
-            return t->src[0];
-        case GGML_OP_CPY:
-            return t->src[1];
-        default:
-            return NULL;
-    }
-}
-
-static struct ggml_tensor * get_view_source(struct ggml_tensor * t) {
-    struct ggml_tensor * parent = t;
-    do {
-        parent = get_view_parent(parent);
-    } while (ggml_is_view(parent));
-    return parent;
-}
-
-static bool ggml_op_can_inplace(enum ggml_op op) {
-    switch (op) {
-        case GGML_OP_SCALE:
-        case GGML_OP_DIAG_MASK_ZERO:
-        case GGML_OP_DIAG_MASK_INF:
-        case GGML_OP_ADD:
-        case GGML_OP_ADD1:
-        case GGML_OP_ACC:
-        case GGML_OP_SUB:
-        case GGML_OP_MUL:
-        case GGML_OP_DIV:
-        case GGML_OP_SQR:
-        case GGML_OP_SQRT:
-        case GGML_OP_LOG:
-        case GGML_OP_UNARY:
-        case GGML_OP_ROPE:
-        case GGML_OP_RMS_NORM:
-        case GGML_OP_SET:
-        case GGML_OP_SOFT_MAX:
-        case GGML_OP_CONT:
-            return true;
-
-        default:
-            return false;
-    }
-}
-
-static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node) {
-    struct hash_node * ht = alloc->hash_table;
-    if (node->data == NULL) {
-        if (ggml_is_view(node)) {
-            size_t offset;
-            switch(node->op) {
-                case GGML_OP_VIEW:
-                    memcpy(&offset, node->op_params, sizeof(size_t));
-                    node->data = (char *) node->src[0]->data + offset;
-                    break;
-                case GGML_OP_PERMUTE:
-                case GGML_OP_RESHAPE:
-                case GGML_OP_TRANSPOSE:
-                    node->data = node->src[0]->data;
-                    break;
-                case GGML_OP_CPY:
-                    node->data = node->src[1]->data;
-                    break;
-                default:
-                    GGML_ASSERT(!"unknown view op");
-                    break;
-            }
-        } else {
-            // see if we can reuse a parent's buffer (inplace)
-            if (ggml_op_can_inplace(node->op)) {
-                for (int i = 0; i < GGML_MAX_SRC; i++) {
-                    struct ggml_tensor * parent = node->src[i];
-                    if (parent == NULL) {
-                        break;
-                    }
-                    struct hash_node * p_hn = hash_get(ht, parent);
-                    if (parent->data != NULL && p_hn->n_children == 1 && p_hn->n_views == 0 && ggml_are_same_layout(node, parent)) {
-                        if (ggml_is_view(parent)) {
-                            struct ggml_tensor * view_src = get_view_source(parent);
-                            struct hash_node * view_src_hn = hash_get(ht, view_src);
-                            if (view_src_hn->n_views == 1 && view_src_hn->n_children == 0 && view_src->data == parent->data) {
-                                // TODO: the offset of the view parent must be kept to ensure that the op doesn't overwrite
-                                // the parent's data that it will need later (same layout requirement). the problem is that then
-                                // we cannot free the tensor because the original address of the allocation is lost.
-                                // adding a view_src pointer to the tensor would solve this and simplify the code dealing with views
-                                // for now, we only reuse the parent's data if the offset is zero (view_src->data == parent->data)
-                                AT_PRINTF("reusing view parent %s (%s) for %s\n", parent->name, view_src->name, node->name);
-                                node->data = parent->data;
-                                return;
-                            }
-                        }
-                        else {
-                            AT_PRINTF("reusing parent %s for %s\n", parent->name, node->name);
-                            node->data = parent->data;
-                        }
-                        return;
-                    }
-                }
-            }
-            ggml_allocr_alloc(alloc, node);
-        }
-    }
-}
-
-static size_t ggml_allocator_alloc_graph_tensors_n(
-    struct ggml_allocr * alloc,
-    struct ggml_cgraph ** graphs, int n_graphs,
-    struct ggml_tensor *** inputs, struct ggml_tensor *** outputs) {
-
-    // reset hash table
-    struct hash_node * ht = alloc->hash_table;
-    memset(ht, 0, sizeof(struct hash_node) * GGML_GRAPH_HASHTABLE_SIZE);
-
-    // count number of children and views
-    for (int g = 0; g < n_graphs; g++) {
-        struct ggml_cgraph * gf = graphs[g];
-        for (int i = 0; i < gf->n_nodes; i++) {
-            struct ggml_tensor * node = gf->nodes[i];
-
-            if (ggml_is_view(node)) {
-                struct ggml_tensor * view_src = get_view_source(node);
-                hash_get(ht, view_src)->n_views += 1;
-            }
-
-            for (int j = 0; j < GGML_MAX_SRC; j++) {
-                struct ggml_tensor * parent = node->src[j];
-                if (parent == NULL) {
-                    break;
-                }
-                hash_get(ht, parent)->n_children += 1;
-            }
-        }
-    }
-
-    // allocate tensors
-    for (int g = 0; g < n_graphs; g++) {
-        struct ggml_cgraph * gf = graphs[g];
-        AT_PRINTF("####### graph %d/%d\n", g, n_graphs);
-        // graph inputs are allocated first to ensure that they are not overwritten by each other
-        if (inputs != NULL && inputs[g] != NULL) {
-            for (int i = 0; inputs[g][i] != NULL; i++) {
-                struct ggml_tensor * input = inputs[g][i];
-                AT_PRINTF("input: %s\n", input->name);
-                allocate_node(alloc, input);
-            }
-        }
-        for (int i = 0; i < gf->n_nodes; i++) {
-            struct ggml_tensor * node = gf->nodes[i];
-
-            // allocate parents (leafs)
-            for (int j = 0; j < GGML_MAX_SRC; j++) {
-                struct ggml_tensor * parent = node->src[j];
-                if (parent == NULL) {
-                    break;
-                }
-                allocate_node(alloc, parent);
-            }
-
-            // allocate node
-            allocate_node(alloc, node);
-
-            AT_PRINTF("exec: %s (%s) <= ", ggml_op_name(node->op), node->name);
-            for (int j = 0; j < GGML_MAX_SRC; j++) {
-                struct ggml_tensor * parent = node->src[j];
-                if (parent == NULL) {
-                    break;
-                }
-                AT_PRINTF("%s", parent->name);
-                if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) {
-                    AT_PRINTF(", ");
-                }
-            }
-            AT_PRINTF("\n");
-
-            // update parents
-            for (int j = 0; j < GGML_MAX_SRC; j++) {
-                struct ggml_tensor * parent = node->src[j];
-                if (parent == NULL) {
-                    break;
-                }
-                struct hash_node * p_hn = hash_get(ht, parent);
-                p_hn->n_children -= 1;
-
-                //AT_PRINTF("parent %s: %d children, %d views\n", parent->name, parent->n_children, parent->n_views);
-
-                if (p_hn->n_children == 0 && p_hn->n_views == 0) {
-                    if (ggml_is_view(parent)) {
-                        struct ggml_tensor * view_src = get_view_source(parent);
-                        struct hash_node * view_src_hn = hash_get(ht, view_src);
-                        view_src_hn->n_views -= 1;
-                        AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src->n_children, view_src->n_views);
-                        if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src->data != node->data) {
-                            ggml_allocator_free_tensor(alloc, view_src);
-                        }
-                    }
-                    else {
-                        if (parent->data != node->data) {
-                            ggml_allocator_free_tensor(alloc, parent);
-                        }
-                    }
-                }
-            }
-            AT_PRINTF("\n");
-        }
-        // free graph outputs here that wouldn't be freed otherwise because they have no children
-        if (outputs != NULL && outputs[g] != NULL) {
-            for (int i = 0; outputs[g][i] != NULL; i++) {
-                struct ggml_tensor * output = outputs[g][i];
-                AT_PRINTF("output: %s\n", output->name);
-                ggml_allocator_free_tensor(alloc, output);
-            }
-        }
-    }
-
-    return alloc->max_size;
-}
-
-size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph) {
-    return ggml_allocator_alloc_graph_tensors_n(alloc, &graph, 1, NULL, NULL);
-}

llama/ggml-alloc.h

@@ -1,48 +0,0 @@
-/**
- * llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12
- *
- * MIT License
- *
- * Copyright (c) 2023 Georgi Gerganov
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#pragma once
-
-#include "ggml.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-
-GGML_API struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment);
-GGML_API struct ggml_allocr * ggml_allocr_new_measure(size_t alignment);
-
-GGML_API void   ggml_allocr_free(struct ggml_allocr * alloc);
-GGML_API bool   ggml_allocr_is_measure(struct ggml_allocr * alloc);
-GGML_API void   ggml_allocr_reset(struct ggml_allocr * alloc);
-GGML_API void   ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor);
-GGML_API size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph);
-
-
-#ifdef  __cplusplus
-}
-#endif

llama/ggml-cuda.h

@@ -1,63 +0,0 @@
-/**
- * llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12
- *
- * MIT License
- *
- * Copyright (c) 2023 Georgi Gerganov
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#pragma once
-
-#include "ggml.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-#define GGML_CUDA_MAX_DEVICES       16
-
-void   ggml_init_cublas(void);
-void   ggml_cuda_set_tensor_split(const float * tensor_split);
-
-void   ggml_cuda_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-bool   ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-size_t ggml_cuda_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-void   ggml_cuda_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
-
-// TODO: export these with GGML_API
-void * ggml_cuda_host_malloc(size_t size);
-void   ggml_cuda_host_free(void * ptr);
-
-void   ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor);
-
-void   ggml_cuda_free_data(struct ggml_tensor * tensor);
-void   ggml_cuda_assign_buffers(struct ggml_tensor * tensor);
-void   ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);
-void   ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor);
-void   ggml_cuda_set_main_device(int main_device);
-void   ggml_cuda_set_mul_mat_q(bool mul_mat_q);
-void   ggml_cuda_set_scratch_size(size_t scratch_size);
-void   ggml_cuda_free_scratch(void);
-bool   ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
-
-#ifdef  __cplusplus
-}
-#endif

llama/ggml-metal.h

@@ -1,106 +0,0 @@
-//go:build darwin
-
-/**
- * llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12
- *
- * MIT License
- *
- * Copyright (c) 2023 Georgi Gerganov
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-// An interface allowing to compute ggml_cgraph with Metal
-//
-// This is a fully functional interface that extends ggml with GPU support for Apple devices.
-// A similar interface can be created for other GPU backends (e.g. Vulkan, CUDA, OpenCL, etc.)
-//
-// How it works?
-//
-// As long as your program can create and evaluate a ggml_cgraph on the CPU, you can use this
-// interface to evaluate the same graph on the GPU. Instead of using ggml_graph_compute(), you
-// use ggml_metal_graph_compute() (or ggml_vulkan_graph_compute(), etc.)
-//
-// You only need to make sure that all memory buffers that you used during the graph creation
-// are mapped to the device memory with the ggml_metal_add_buffer() function. This mapping is
-// used during the graph evaluation to determine the arguments of the compute kernels.
-//
-// Synchronization between device and host memory (for example for input and output tensors)
-// is done with the ggml_metal_set_tensor() and ggml_metal_get_tensor() functions.
-//
-
-#pragma once
-
-#include <stddef.h>
-#include <stdbool.h>
-
-// max memory buffers that can be mapped to the device
-#define GGML_METAL_MAX_BUFFERS 16
-
-struct ggml_tensor;
-struct ggml_cgraph;
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-struct ggml_metal_context;
-
-// number of command buffers to use
-struct ggml_metal_context * ggml_metal_init(int n_cb);
-void ggml_metal_free(struct ggml_metal_context * ctx);
-
-// set the number of command buffers to use
-void ggml_metal_set_n_cb(struct ggml_metal_context * ctx, int n_cb);
-
-// creates a mapping between a host memory buffer and a device memory buffer
-// - make sure to map all buffers used in the graph before calling ggml_metal_graph_compute
-// - the mapping is used during computation to determine the arguments of the compute kernels
-// - you don't need to keep the host memory buffer allocated as it is never accessed by Metal
-// - max_size specifies the maximum size of a tensor and is used to create shared views such
-//   that it is guaranteed that the tensor will fit in at least one of the views
-//
-bool ggml_metal_add_buffer(
-        struct ggml_metal_context * ctx,
-                       const char * name,
-                             void * data,
-                           size_t   size,
-                           size_t   max_size);
-
-// set data from host memory into the device
-void ggml_metal_set_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
-
-// get data from the device into host memory
-void ggml_metal_get_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
-
-// try to find operations that can be run concurrently in the graph
-// you should run it again if the topology of your graph changes
-void ggml_metal_graph_find_concurrency(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);
-
-// if the graph has been optimized for concurrently dispatch
-bool ggml_metal_if_optimized(struct ggml_metal_context * ctx);
-
-// same as ggml_graph_compute but uses Metal
-// creates gf->n_threads command buffers in parallel
-void ggml_metal_graph_compute(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);
-
-#ifdef __cplusplus
-}
-#endif
-

llama/ggml-mpi.c

@@ -1,244 +0,0 @@
-//go:build mpi
-
-/**
- * llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12
- *
- * MIT License
- *
- * Copyright (c) 2023 Georgi Gerganov
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#include "ggml-mpi.h"
-
-#include "ggml.h"
-
-#include <mpi.h>
-
-#include <stdio.h>
-#include <stdlib.h>
-
-#define MIN(a, b) ((a) < (b) ? (a) : (b))
-
-#define UNUSED GGML_UNUSED
-
-struct ggml_mpi_context {
-    int rank;
-    int size;
-};
-
-void ggml_mpi_backend_init(void) {
-    MPI_Init(NULL, NULL);
-}
-
-void ggml_mpi_backend_free(void) {
-    MPI_Finalize();
-}
-
-struct ggml_mpi_context * ggml_mpi_init(void) {
-    struct ggml_mpi_context * ctx = calloc(1, sizeof(struct ggml_mpi_context));
-
-    MPI_Comm_rank(MPI_COMM_WORLD, &ctx->rank);
-    MPI_Comm_size(MPI_COMM_WORLD, &ctx->size);
-
-    return ctx;
-}
-
-void ggml_mpi_free(struct ggml_mpi_context * ctx) {
-    free(ctx);
-}
-
-int ggml_mpi_rank(struct ggml_mpi_context * ctx) {
-    return ctx->rank;
-}
-
-void ggml_mpi_eval_init(
-        struct ggml_mpi_context * ctx_mpi,
-                            int * n_tokens,
-                            int * n_past,
-                            int * n_threads) {
-    UNUSED(ctx_mpi);
-
-    // synchronize the worker node parameters with the root node
-    MPI_Barrier(MPI_COMM_WORLD);
-
-    MPI_Bcast(n_tokens,  1, MPI_INT, 0, MPI_COMM_WORLD);
-    MPI_Bcast(n_past,    1, MPI_INT, 0, MPI_COMM_WORLD);
-    MPI_Bcast(n_threads, 1, MPI_INT, 0, MPI_COMM_WORLD);
-}
-
-static int ggml_graph_get_node_idx(struct ggml_cgraph * gf, const char * name) {
-    struct ggml_tensor * t = ggml_graph_get_tensor(gf, name);
-    if (t == NULL) {
-        fprintf(stderr, "%s: tensor %s not found\n", __func__, name);
-        return -1;
-    }
-
-    for (int i = 0; i < gf->n_nodes; i++) {
-        if (gf->nodes[i] == t) {
-            return i;
-        }
-    }
-
-    fprintf(stderr, "%s: tensor %s not found in graph (should not happen)\n", __func__, name);
-    return -1;
-}
-
-static void ggml_mpi_tensor_send(struct ggml_tensor * t, int mpi_rank_dst) {
-    MPI_Datatype mpi_type;
-
-    switch (t->type) {
-        case GGML_TYPE_I32: mpi_type = MPI_INT32_T; break;
-        case GGML_TYPE_F32: mpi_type = MPI_FLOAT;   break;
-        default: GGML_ASSERT(false && "not implemented");
-    }
-
-    const int retval = MPI_Send(t->data, ggml_nelements(t), mpi_type, mpi_rank_dst, 0, MPI_COMM_WORLD);
-    GGML_ASSERT(retval == MPI_SUCCESS);
-}
-
-static void ggml_mpi_tensor_recv(struct ggml_tensor * t, int mpi_rank_src) {
-    MPI_Datatype mpi_type;
-
-    switch (t->type) {
-        case GGML_TYPE_I32: mpi_type = MPI_INT32_T; break;
-        case GGML_TYPE_F32: mpi_type = MPI_FLOAT;   break;
-        default: GGML_ASSERT(false && "not implemented");
-    }
-
-    MPI_Status status; UNUSED(status);
-
-    const int retval = MPI_Recv(t->data, ggml_nelements(t), mpi_type, mpi_rank_src, MPI_ANY_TAG, MPI_COMM_WORLD, &status);
-    GGML_ASSERT(retval == MPI_SUCCESS);
-}
-
-// TODO: there are many improvements that can be done to this implementation
-void ggml_mpi_graph_compute_pre(
-        struct ggml_mpi_context * ctx_mpi,
-             struct ggml_cgraph * gf,
-                            int   n_layers) {
-    const int mpi_rank = ctx_mpi->rank;
-    const int mpi_size = ctx_mpi->size;
-
-    struct ggml_tensor * inp_tokens = ggml_graph_get_tensor(gf, "inp_tokens");
-    if (inp_tokens == NULL) {
-        fprintf(stderr, "%s: tensor 'inp_tokens' not found\n", __func__);
-        return;
-    }
-
-    struct ggml_tensor * inp0 = ggml_graph_get_tensor(gf, "layer_inp_0");
-    if (inp0 == NULL) {
-        fprintf(stderr, "%s: tensor 'inp0' not found\n", __func__);
-        return;
-    }
-
-    GGML_ASSERT(inp0 == gf->nodes[0]);
-
-    // distribute the compute graph into slices across the MPI nodes
-    //
-    // the main node (0) processes the last layers + the remainder of the compute graph
-    // and is responsible to pass the input tokens to the first node (1)
-    //
-    // node 1:   [(  0) * n_per_node, (  1) * n_per_node)
-    // node 2:   [(  1) * n_per_node, (  2) * n_per_node)
-    // ...
-    // node n-1: [(n-2) * n_per_node, (n-1) * n_per_node)
-    // node 0:   [(n-1) * n_per_node,            n_nodes)
-    //
-    if (mpi_rank > 0) {
-        if (mpi_rank == 1) {
-            // the first node (1) receives the input tokens from the main node (0)
-            ggml_mpi_tensor_recv(inp_tokens, 0);
-        } else {
-            // recv input data for each node into the "inp0" tensor (i.e. the first node in the compute graph)
-            ggml_mpi_tensor_recv(inp0, mpi_rank - 1);
-        }
-    } else if (mpi_size > 1) {
-        // node 0 sends the input tokens to node 1
-        ggml_mpi_tensor_send(inp_tokens, 1);
-
-        // recv the output data from the last node
-        ggml_mpi_tensor_recv(inp0, mpi_size - 1);
-    }
-
-    {
-        const int n_per_node = (n_layers + (mpi_size - 1)) / mpi_size;
-
-        const int mpi_idx = mpi_rank > 0 ? mpi_rank - 1 : mpi_size - 1;
-
-        const int il0 =               (mpi_idx + 0) * n_per_node;
-        const int il1 = MIN(n_layers, (mpi_idx + 1) * n_per_node);
-
-        char name_l0[GGML_MAX_NAME];
-        char name_l1[GGML_MAX_NAME];
-
-        snprintf(name_l0, sizeof(name_l0), "layer_inp_%d", il0);
-        snprintf(name_l1, sizeof(name_l1), "layer_inp_%d", il1);
-
-        const int idx_l0 =                ggml_graph_get_node_idx(gf, name_l0);
-        const int idx_l1 = mpi_rank > 0 ? ggml_graph_get_node_idx(gf, name_l1) + 1 : gf->n_nodes;
-
-        if (idx_l0 < 0 || idx_l1 < 0) {
-            fprintf(stderr, "%s: layer input nodes not found\n", __func__);
-            return;
-        }
-
-        // attach the input data to all nodes that need it
-        // TODO: not great - should be able to do this without modifying the compute graph (see next TODO below)
-        for (int i = idx_l0; i < idx_l1; i++) {
-            if (gf->nodes[i]->src[0] == gf->nodes[idx_l0]) {
-                gf->nodes[i]->src[0] =  inp0;
-            }
-            if (gf->nodes[i]->src[1] == gf->nodes[idx_l0]) {
-                gf->nodes[i]->src[1] =  inp0;
-            }
-        }
-
-        // TODO: instead of rearranging the nodes, we should be able to execute a subset of the compute graph
-        for (int i = 1; i < idx_l1 - idx_l0; i++) {
-            gf->nodes[i] = gf->nodes[idx_l0 + i];
-            gf->grads[i] = gf->grads[idx_l0 + i];
-        }
-
-        // the first node performs the "get_rows" operation, the rest of the nodes get the data from the previous node
-        if (mpi_idx != 0) {
-            gf->nodes[0]->op = GGML_OP_NONE;
-        }
-
-        gf->n_nodes = idx_l1 - idx_l0;
-
-        //fprintf(stderr, "%s: node %d: processing %d nodes [%d, %d)\n", __func__, mpi_rank, gf->n_nodes, il0, il1);
-    }
-}
-
-void ggml_mpi_graph_compute_post(
-        struct ggml_mpi_context * ctx_mpi,
-             struct ggml_cgraph * gf,
-                            int   n_layers) {
-    UNUSED(n_layers);
-
-    const int mpi_rank = ctx_mpi->rank;
-    const int mpi_size = ctx_mpi->size;
-
-    // send the output data to the next node
-    if (mpi_rank > 0) {
-        ggml_mpi_tensor_send(gf->nodes[gf->n_nodes - 1], (mpi_rank + 1) % mpi_size);
-    }
-}

llama/ggml-mpi.h

@@ -1,67 +0,0 @@
-//go:build mpi
-
-/**
- * llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12
- *
- * MIT License
- *
- * Copyright (c) 2023 Georgi Gerganov
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#pragma once
-
-struct ggml_context;
-struct ggml_tensor;
-struct ggml_cgraph;
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-struct ggml_mpi_context;
-
-void ggml_mpi_backend_init(void);
-void ggml_mpi_backend_free(void);
-
-struct ggml_mpi_context * ggml_mpi_init(void);
-void ggml_mpi_free(struct ggml_mpi_context * ctx);
-
-int ggml_mpi_rank(struct ggml_mpi_context * ctx);
-
-void ggml_mpi_eval_init(
-        struct ggml_mpi_context * ctx_mpi,
-                            int * n_tokens,
-                            int * n_past,
-                            int * n_threads);
-
-void ggml_mpi_graph_compute_pre(
-        struct ggml_mpi_context * ctx_mpi,
-             struct ggml_cgraph * gf,
-                            int   n_layers);
-
-void ggml_mpi_graph_compute_post(
-        struct ggml_mpi_context * ctx_mpi,
-             struct ggml_cgraph * gf,
-                            int   n_layers);
-
-#ifdef __cplusplus
-}
-#endif

llama/ggml-opencl.h

@@ -1,53 +0,0 @@
-//go:build opencl
-
-/**
- * llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12
- *
- * MIT License
- *
- * Copyright (c) 2023 Georgi Gerganov
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#pragma once
-
-#include "ggml.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-void ggml_cl_init(void);
-
-void   ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-bool   ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-void   ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
-
-void * ggml_cl_host_malloc(size_t size);
-void   ggml_cl_host_free(void * ptr);
-
-void ggml_cl_free_data(const struct ggml_tensor* tensor);
-
-void ggml_cl_transform_tensor(void * data, struct ggml_tensor * tensor);
-
-#ifdef  __cplusplus
-}
-#endif

llama/k_quants.h

@@ -1,191 +0,0 @@
-/**
- * llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12
- *
- * MIT License
- *
- * Copyright (c) 2023 Georgi Gerganov
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#pragma once
-
-#include "ggml.h"
-
-#include <stdint.h>
-#include <assert.h>
-#include <stddef.h>
-
-// Super-block size
-#ifdef GGML_QKK_64
-#define QK_K 64
-#define K_SCALE_SIZE 4
-#else
-#define QK_K 256
-#define K_SCALE_SIZE 12
-#endif
-
-#ifndef static_assert
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
-#define static_assert(cond, msg) _Static_assert(cond, msg)
-#else
-#define static_assert(cond, msg) struct global_scope_noop_trick
-#endif
-#endif
-
-//
-// Super-block quantization structures
-//
-
-// 2-bit quantization
-// weight is represented as x = a * q + b
-// 16 blocks of 16 elemenets each
-// Effectively 2.5625 bits per weight
-typedef struct {
-    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
-    uint8_t qs[QK_K/4];      // quants
-    ggml_fp16_t d;           // super-block scale for quantized scales
-    ggml_fp16_t dmin;        // super-block scale for quantized mins
-} block_q2_K;
-static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
-
-// 3-bit quantization
-// weight is represented as x = a * q
-// 16 blocks of 16 elemenets each
-// Effectively 3.4375 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    uint8_t hmask[QK_K/8];     // quants - high bit
-    uint8_t qs[QK_K/4];        // quants - low 2 bits
-    uint8_t scales[2];
-    ggml_fp16_t d;             // super-block scale
-} block_q3_K;
-static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding");
-#else
-typedef struct {
-    uint8_t hmask[QK_K/8];     // quants - high bit
-    uint8_t qs[QK_K/4];        // quants - low 2 bits
-    uint8_t scales[12];        // scales, quantized with 6 bits
-    ggml_fp16_t d;             // super-block scale
-} block_q3_K;
-static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
-#endif
-
-// 4-bit quantization
-// 16 blocks of 32 elements each
-// weight is represented as x = a * q + b
-// Effectively 4.5 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    ggml_fp16_t d[2];          // super-block scales/mins
-    uint8_t scales[2];         // 4-bit block scales/mins
-    uint8_t qs[QK_K/2];        // 4--bit quants
-} block_q4_K;
-static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + QK_K/2 + 2, "wrong q4_K block size/padding");
-#else
-typedef struct {
-    ggml_fp16_t d;             // super-block scale for quantized scales
-    ggml_fp16_t dmin;          // super-block scale for quantized mins
-    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
-    uint8_t qs[QK_K/2];        // 4--bit quants
-} block_q4_K;
-static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding");
-#endif
-
-// 5-bit quantization
-// 16 blocks of 32 elements each
-// weight is represented as x = a * q + b
-// Effectively 5.5 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    ggml_fp16_t d;               // super-block scale
-    int8_t  scales[QK_K/16];     // 8-bit block scales
-    uint8_t qh[QK_K/8];          // quants, high bit
-    uint8_t qs[QK_K/2];          // quants, low 4 bits
-} block_q5_K;
-static_assert(sizeof(block_q5_K) == sizeof(ggml_fp16_t) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
-#else
-typedef struct {
-    ggml_fp16_t d;               // super-block scale for quantized scales
-    ggml_fp16_t dmin;            // super-block scale for quantized mins
-    uint8_t scales[K_SCALE_SIZE];   // scales and mins, quantized with 6 bits
-    uint8_t qh[QK_K/8];          // quants, high bit
-    uint8_t qs[QK_K/2];          // quants, low 4 bits
-} block_q5_K;
-static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
-#endif
-
-// 6-bit quantization
-// weight is represented as x = a * q
-// 16 blocks of 16 elemenets each
-// Effectively 6.5625 bits per weight
-typedef struct {
-    uint8_t ql[QK_K/2];      // quants, lower 4 bits
-    uint8_t qh[QK_K/4];      // quants, upper 2 bits
-    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
-    ggml_fp16_t d;           // super-block scale
-} block_q6_K;
-static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + QK_K / 16 + 3*QK_K/4, "wrong q6_K block size/padding");
-
-// This is only used for intermediate quantization and dot products
-typedef struct {
-    float   d;              // delta
-    int8_t  qs[QK_K];       // quants
-    int16_t bsums[QK_K/16]; // sum of quants in groups of 16
-} block_q8_K;
-static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding");
-
-
-// Quantization
-void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict y, int k);
-void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict y, int k);
-void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict y, int k);
-void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict y, int k);
-void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int k);
-void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k);
-
-void quantize_row_q2_K(const float * restrict x, void * restrict y, int k);
-void quantize_row_q3_K(const float * restrict x, void * restrict y, int k);
-void quantize_row_q4_K(const float * restrict x, void * restrict y, int k);
-void quantize_row_q5_K(const float * restrict x, void * restrict y, int k);
-void quantize_row_q6_K(const float * restrict x, void * restrict y, int k);
-void quantize_row_q8_K(const float * restrict x, void * restrict y, int k);
-
-// Dequantization
-void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int k);
-void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k);
-void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int k);
-void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int k);
-void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int k);
-void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k);
-
-// Dot product
-void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-
-// Quantization with histogram collection
-size_t ggml_quantize_q2_K(const float * src, void * dst, int n, int k, int64_t * hist);
-size_t ggml_quantize_q3_K(const float * src, void * dst, int n, int k, int64_t * hist);
-size_t ggml_quantize_q4_K(const float * src, void * dst, int n, int k, int64_t * hist);
-size_t ggml_quantize_q5_K(const float * src, void * dst, int n, int k, int64_t * hist);
-size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist);
-

llama/llama-util.h

@@ -1,530 +0,0 @@
-/**
- * llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12
- *
- * MIT License
- *
- * Copyright (c) 2023 Georgi Gerganov
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-// Internal header to be included only by llama.cpp.
-// Contains wrappers around OS interfaces.
-
-#ifndef LLAMA_UTIL_H
-#define LLAMA_UTIL_H
-
-#include <cstdio>
-#include <cstdint>
-#include <cerrno>
-#include <cstring>
-#include <cstdarg>
-#include <cstdlib>
-#include <climits>
-
-#include <string>
-#include <vector>
-#include <stdexcept>
-
-#ifdef __has_include
-    #if __has_include(<unistd.h>)
-        #include <unistd.h>
-        #if defined(_POSIX_MAPPED_FILES)
-            #include <sys/mman.h>
-        #endif
-        #if defined(_POSIX_MEMLOCK_RANGE)
-            #include <sys/resource.h>
-        #endif
-    #endif
-#endif
-
-#if defined(_WIN32)
-    #define WIN32_LEAN_AND_MEAN
-    #ifndef NOMINMAX
-        #define NOMINMAX
-    #endif
-    #include <windows.h>
-    #include <io.h>
-    #include <stdio.h> // for _fseeki64
-#endif
-
-#define LLAMA_ASSERT(x) \
-    do { \
-        if (!(x)) { \
-            fprintf(stderr, "LLAMA_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
-            abort(); \
-        } \
-    } while (0)
-
-#ifdef __GNUC__
-#ifdef __MINGW32__
-__attribute__((format(gnu_printf, 1, 2)))
-#else
-__attribute__((format(printf, 1, 2)))
-#endif
-#endif
-static std::string format(const char * fmt, ...) {
-    va_list ap, ap2;
-    va_start(ap, fmt);
-    va_copy(ap2, ap);
-    int size = vsnprintf(NULL, 0, fmt, ap);
-    LLAMA_ASSERT(size >= 0 && size < INT_MAX);
-    std::vector<char> buf(size + 1);
-    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
-    LLAMA_ASSERT(size2 == size);
-    va_end(ap2);
-    va_end(ap);
-    return std::string(buf.data(), size);
-}
-
-struct llama_file {
-    // use FILE * so we don't have to re-open the file to mmap
-    FILE * fp;
-    size_t size;
-
-    llama_file(const char * fname, const char * mode) {
-        fp = std::fopen(fname, mode);
-        if (fp == NULL) {
-            throw std::runtime_error(format("failed to open %s: %s", fname, strerror(errno)));
-        }
-        seek(0, SEEK_END);
-        size = tell();
-        seek(0, SEEK_SET);
-    }
-
-    size_t tell() const {
-#ifdef _WIN32
-        __int64 ret = _ftelli64(fp);
-#else
-        long ret = std::ftell(fp);
-#endif
-        LLAMA_ASSERT(ret != -1); // this really shouldn't fail
-        return (size_t) ret;
-    }
-
-    void seek(size_t offset, int whence) {
-#ifdef _WIN32
-        int ret = _fseeki64(fp, (__int64) offset, whence);
-#else
-        int ret = std::fseek(fp, (long) offset, whence);
-#endif
-        LLAMA_ASSERT(ret == 0); // same
-    }
-
-    void read_raw(void * ptr, size_t len) const {
-        if (len == 0) {
-            return;
-        }
-        errno = 0;
-        std::size_t ret = std::fread(ptr, len, 1, fp);
-        if (ferror(fp)) {
-            throw std::runtime_error(format("read error: %s", strerror(errno)));
-        }
-        if (ret != 1) {
-            throw std::runtime_error(std::string("unexpectedly reached end of file"));
-        }
-    }
-
-    std::uint32_t read_u32() {
-        std::uint32_t ret;
-        read_raw(&ret, sizeof(ret));
-        return ret;
-    }
-
-    std::string read_string(std::uint32_t len) {
-        std::vector<char> chars(len);
-        read_raw(chars.data(), len);
-        return std::string(chars.data(), len);
-    }
-
-    void write_raw(const void * ptr, size_t len) const {
-        if (len == 0) {
-            return;
-        }
-        errno = 0;
-        size_t ret = std::fwrite(ptr, len, 1, fp);
-        if (ret != 1) {
-            throw std::runtime_error(format("write error: %s", strerror(errno)));
-        }
-    }
-
-    void write_u32(std::uint32_t val) {
-        write_raw(&val, sizeof(val));
-    }
-
-    ~llama_file() {
-        if (fp) {
-            std::fclose(fp);
-        }
-    }
-};
-
-#if defined(_WIN32)
-static std::string llama_format_win_err(DWORD err) {
-    LPSTR buf;
-    size_t size = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
-                                 NULL, err, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&buf, 0, NULL);
-    if (!size) {
-        return "FormatMessageA failed";
-    }
-    std::string ret(buf, size);
-    LocalFree(buf);
-    return ret;
-}
-#endif
-
-struct llama_mmap {
-    void * addr;
-    size_t size;
-
-    llama_mmap(const llama_mmap &) = delete;
-
-#ifdef _POSIX_MAPPED_FILES
-    static constexpr bool SUPPORTED = true;
-
-    llama_mmap(struct llama_file * file, size_t prefetch = (size_t) -1 /* -1 = max value */, bool numa = false) {
-        size = file->size;
-        int fd = fileno(file->fp);
-        int flags = MAP_SHARED;
-        // prefetch/readahead impairs performance on NUMA systems
-        if (numa) { prefetch = 0; }
-#ifdef __linux__
-        if (prefetch) { flags |= MAP_POPULATE; }
-#endif
-        addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0);
-        if (addr == MAP_FAILED) {
-            throw std::runtime_error(format("mmap failed: %s", strerror(errno)));
-        }
-
-        if (prefetch > 0) {
-            // Advise the kernel to preload the mapped memory
-            if (madvise(addr, std::min(file->size, prefetch), MADV_WILLNEED)) {
-                fprintf(stderr, "warning: madvise(.., MADV_WILLNEED) failed: %s\n",
-                        strerror(errno));
-            }
-        }
-        if (numa) {
-            // advise the kernel not to use readahead
-            // (because the next page might not belong on the same node)
-            if (madvise(addr, file->size, MADV_RANDOM)) {
-                fprintf(stderr, "warning: madvise(.., MADV_RANDOM) failed: %s\n",
-                        strerror(errno));
-            }
-        }
-    }
-
-    ~llama_mmap() {
-        munmap(addr, size);
-    }
-#elif defined(_WIN32)
-    static constexpr bool SUPPORTED = true;
-
-    llama_mmap(struct llama_file * file, bool prefetch = true, bool numa = false) {
-        (void) numa;
-
-        size = file->size;
-
-        HANDLE hFile = (HANDLE) _get_osfhandle(_fileno(file->fp));
-
-        HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
-        DWORD error = GetLastError();
-
-        if (hMapping == NULL) {
-            throw std::runtime_error(format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str()));
-        }
-
-        addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0);
-        error = GetLastError();
-        CloseHandle(hMapping);
-
-        if (addr == NULL) {
-            throw std::runtime_error(format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str()));
-        }
-
-        #if _WIN32_WINNT >= _WIN32_WINNT_WIN8
-        if (prefetch) {
-            // Advise the kernel to preload the mapped memory
-            WIN32_MEMORY_RANGE_ENTRY range;
-            range.VirtualAddress = addr;
-            range.NumberOfBytes = (SIZE_T)size;
-            if (!PrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
-                fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n",
-                        llama_format_win_err(GetLastError()).c_str());
-            }
-        }
-        #else
-        #pragma message("warning: You are building for pre-Windows 8; prefetch not supported")
-        #endif // _WIN32_WINNT >= _WIN32_WINNT_WIN8
-    }
-
-    ~llama_mmap() {
-        if (!UnmapViewOfFile(addr)) {
-            fprintf(stderr, "warning: UnmapViewOfFile failed: %s\n",
-                    llama_format_win_err(GetLastError()).c_str());
-        }
-    }
-#else
-    static constexpr bool SUPPORTED = false;
-
-    llama_mmap(struct llama_file *, bool prefetch = true, bool numa = false) {
-        (void) prefetch;
-        (void) numa;
-
-        throw std::runtime_error(std::string("mmap not supported"));
-    }
-#endif
-};
-
-// Represents some region of memory being locked using mlock or VirtualLock;
-// will automatically unlock on destruction.
-struct llama_mlock {
-    void * addr = NULL;
-    size_t size = 0;
-    bool failed_already = false;
-
-    llama_mlock() {}
-    llama_mlock(const llama_mlock &) = delete;
-
-    ~llama_mlock() {
-        if (size) {
-            raw_unlock(addr, size);
-        }
-    }
-
-    void init(void * ptr) {
-        LLAMA_ASSERT(addr == NULL && size == 0);
-        addr = ptr;
-    }
-
-    void grow_to(size_t target_size) {
-        LLAMA_ASSERT(addr);
-        if (failed_already) {
-            return;
-        }
-        size_t granularity = lock_granularity();
-        target_size = (target_size + granularity - 1) & ~(granularity - 1);
-        if (target_size > size) {
-            if (raw_lock((uint8_t *) addr + size, target_size - size)) {
-                size = target_size;
-            } else {
-                failed_already = true;
-            }
-        }
-    }
-
-#ifdef _POSIX_MEMLOCK_RANGE
-    static constexpr bool SUPPORTED = true;
-
-    size_t lock_granularity() {
-        return (size_t) sysconf(_SC_PAGESIZE);
-    }
-
-    #ifdef __APPLE__
-        #define MLOCK_SUGGESTION \
-            "Try increasing the sysctl values 'vm.user_wire_limit' and 'vm.global_user_wire_limit' and/or " \
-            "decreasing 'vm.global_no_user_wire_amount'.  Also try increasing RLIMIT_MLOCK (ulimit -l).\n"
-    #else
-        #define MLOCK_SUGGESTION \
-            "Try increasing RLIMIT_MLOCK ('ulimit -l' as root).\n"
-    #endif
-
-    bool raw_lock(const void * addr, size_t size) {
-        if (!mlock(addr, size)) {
-            return true;
-        } else {
-            char* errmsg = std::strerror(errno);
-            bool suggest = (errno == ENOMEM);
-
-            // Check if the resource limit is fine after all
-            struct rlimit lock_limit;
-            if (suggest && getrlimit(RLIMIT_MEMLOCK, &lock_limit))
-                suggest = false;
-            if (suggest && (lock_limit.rlim_max > lock_limit.rlim_cur + size))
-                suggest = false;
-
-            fprintf(stderr, "warning: failed to mlock %zu-byte buffer (after previously locking %zu bytes): %s\n%s",
-                    size, this->size, errmsg, suggest ? MLOCK_SUGGESTION : "");
-            return false;
-        }
-    }
-
-    #undef MLOCK_SUGGESTION
-
-    void raw_unlock(void * addr, size_t size) {
-        if (munlock(addr, size)) {
-            fprintf(stderr, "warning: failed to munlock buffer: %s\n", std::strerror(errno));
-        }
-    }
-#elif defined(_WIN32)
-    static constexpr bool SUPPORTED = true;
-
-    size_t lock_granularity() {
-        SYSTEM_INFO si;
-        GetSystemInfo(&si);
-        return (size_t) si.dwPageSize;
-    }
-
-    bool raw_lock(void * ptr, size_t len) {
-        for (int tries = 1; ; tries++) {
-            if (VirtualLock(ptr, len)) {
-                return true;
-            }
-            if (tries == 2) {
-                fprintf(stderr, "warning: failed to VirtualLock %zu-byte buffer (after previously locking %zu bytes): %s\n",
-                    len, size, llama_format_win_err(GetLastError()).c_str());
-                return false;
-            }
-
-            // It failed but this was only the first try; increase the working
-            // set size and try again.
-            SIZE_T min_ws_size, max_ws_size;
-            if (!GetProcessWorkingSetSize(GetCurrentProcess(), &min_ws_size, &max_ws_size)) {
-                fprintf(stderr, "warning: GetProcessWorkingSetSize failed: %s\n",
-                        llama_format_win_err(GetLastError()).c_str());
-                return false;
-            }
-            // Per MSDN: "The maximum number of pages that a process can lock
-            // is equal to the number of pages in its minimum working set minus
-            // a small overhead."
-            // Hopefully a megabyte is enough overhead:
-            size_t increment = len + 1048576;
-            // The minimum must be <= the maximum, so we need to increase both:
-            min_ws_size += increment;
-            max_ws_size += increment;
-            if (!SetProcessWorkingSetSize(GetCurrentProcess(), min_ws_size, max_ws_size)) {
-                fprintf(stderr, "warning: SetProcessWorkingSetSize failed: %s\n",
-                        llama_format_win_err(GetLastError()).c_str());
-                return false;
-            }
-        }
-    }
-
-    void raw_unlock(void * ptr, size_t len) {
-        if (!VirtualUnlock(ptr, len)) {
-            fprintf(stderr, "warning: failed to VirtualUnlock buffer: %s\n",
-                    llama_format_win_err(GetLastError()).c_str());
-        }
-    }
-#else
-    static constexpr bool SUPPORTED = false;
-
-    size_t lock_granularity() {
-        return (size_t) 65536;
-    }
-
-    bool raw_lock(const void * addr, size_t len) {
-        fprintf(stderr, "warning: mlock not supported on this system\n");
-        return false;
-    }
-
-    void raw_unlock(const void * addr, size_t len) {}
-#endif
-};
-
-// Replacement for std::vector<uint8_t> that doesn't require zero-initialization.
-struct llama_buffer {
-    uint8_t * addr = NULL;
-    size_t size = 0;
-
-    llama_buffer() = default;
-
-    void resize(size_t len) {
-#ifdef GGML_USE_METAL
-        free(addr);
-        int result = posix_memalign((void **) &addr, getpagesize(), len);
-        if (result == 0) {
-            memset(addr, 0, len);
-        }
-        else {
-            addr = NULL;
-        }
-#else
-        delete[] addr;
-        addr = new uint8_t[len];
-#endif
-        size = len;
-    }
-
-    ~llama_buffer() {
-#ifdef GGML_USE_METAL
-        free(addr);
-#else
-        delete[] addr;
-#endif
-        addr = NULL;
-    }
-
-    // disable copy and move
-    llama_buffer(const llama_buffer&) = delete;
-    llama_buffer(llama_buffer&&) = delete;
-    llama_buffer& operator=(const llama_buffer&) = delete;
-    llama_buffer& operator=(llama_buffer&&) = delete;
-};
-
-#ifdef GGML_USE_CUBLAS
-#include "ggml-cuda.h"
-struct llama_ctx_buffer {
-    uint8_t * addr = NULL;
-    bool is_cuda;
-    size_t size = 0;
-
-    llama_ctx_buffer() = default;
-
-    void resize(size_t size) {
-        free();
-
-        addr = (uint8_t *) ggml_cuda_host_malloc(size);
-        if (addr) {
-            is_cuda = true;
-        }
-        else {
-            // fall back to pageable memory
-            addr = new uint8_t[size];
-            is_cuda = false;
-        }
-        this->size = size;
-    }
-
-    void free() {
-        if (addr) {
-            if (is_cuda) {
-                ggml_cuda_host_free(addr);
-            }
-            else {
-                delete[] addr;
-            }
-        }
-        addr = NULL;
-    }
-
-    ~llama_ctx_buffer() {
-        free();
-    }
-
-    // disable copy and move
-    llama_ctx_buffer(const llama_ctx_buffer&) = delete;
-    llama_ctx_buffer(llama_ctx_buffer&&) = delete;
-    llama_ctx_buffer& operator=(const llama_ctx_buffer&) = delete;
-    llama_ctx_buffer& operator=(llama_ctx_buffer&&) = delete;
-};
-#else
-typedef llama_buffer llama_ctx_buffer;
-#endif
-
-#endif

llama/llama.go

@@ -1,439 +0,0 @@
-package llama
-
-/*
-#cgo CPPFLAGS: -O3 -Wall -Wextra -Wno-unused-function -Wno-unused-variable -DNDEBUG -DGGML_USE_K_QUANTS
-#cgo CXXFLAGS: -std=gnu++11
-#cgo darwin CPPFLAGS:  -DGGML_USE_ACCELERATE
-#cgo darwin,arm64 CPPFLAGS: -DGGML_USE_METAL -DGGML_METAL_NDEBUG
-#cgo darwin LDFLAGS: -framework Accelerate -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders
-#include <stdlib.h>
-#include "llama.h"
-
-struct llama_sample_options
-{
-	float repeat_penalty;
-	float frequency_penalty;
-	float presence_penalty;
-	float temperature;
-	int32_t top_k;
-	float top_p;
-	float tfs_z;
-	float typical_p;
-	int mirostat;
-	float mirostat_tau;
-	float mirostat_eta;
-	bool penalize_newline;
-};
-
-llama_token llama_sample(
-		struct llama_context *ctx,
-		struct llama_token_data *candidates,
-		size_t n_candidates,
-		const llama_token *last_tokens,
-		size_t n_last_tokens,
-		struct llama_sample_options *opts)
-{
-	llama_token_data_array candidates_p = {
-		candidates,
-		n_candidates,
-		false,
-	};
-
-	struct llama_token_data newline = candidates_p.data[llama_token_nl()];
-
-	llama_sample_repetition_penalty(
-		ctx, &candidates_p,
-		last_tokens, n_last_tokens,
-		opts->repeat_penalty);
-
-	llama_sample_frequency_and_presence_penalties(
-		ctx, &candidates_p,
-		last_tokens, n_last_tokens,
-		opts->frequency_penalty, opts->presence_penalty);
-
-	if (!opts->penalize_newline) {
-		candidates_p.data[llama_token_nl()] = newline;
-	}
-
-	if (opts->temperature <= 0) {
-		return llama_sample_token_greedy(ctx, &candidates_p);
-	}
-
-	if (opts->mirostat == 1) {
-		int mirostat_m = 100;
-		float mirostat_mu = 2.0f * opts->mirostat_tau;
-		llama_sample_temperature(ctx, &candidates_p, opts->temperature);
-		return llama_sample_token_mirostat(
-			ctx, &candidates_p,
-			opts->mirostat_tau, opts->mirostat_eta,
-			mirostat_m, &mirostat_mu);
-	} else if (opts->mirostat == 2) {
-		float mirostat_mu = 2.0f * opts->mirostat_tau;
-		llama_sample_temperature(ctx, &candidates_p, opts->temperature);
-		return llama_sample_token_mirostat_v2(
-			ctx, &candidates_p,
-			opts->mirostat_tau, opts->mirostat_eta,
-			&mirostat_mu);
-	} else {
-		llama_sample_top_k(ctx, &candidates_p, opts->top_k, 1);
-		llama_sample_tail_free(ctx, &candidates_p, opts->tfs_z, 1);
-		llama_sample_typical(ctx, &candidates_p, opts->typical_p, 1);
-		llama_sample_top_p(ctx, &candidates_p, opts->top_p, 1);
-		llama_sample_temperature(ctx, &candidates_p, opts->temperature);
-		return llama_sample_token(ctx, &candidates_p);
-	}
-}
-*/
-import "C"
-
-import (
-	"bytes"
-	"embed"
-	"errors"
-	"fmt"
-	"io"
-	"log"
-	"os"
-	"strings"
-	"sync"
-	"unicode/utf8"
-	"unsafe"
-
-	"github.com/jmorganca/ollama/api"
-)
-
-//go:embed ggml-metal.metal
-var fs embed.FS
-
-type LLM struct {
-	params *C.struct_llama_context_params
-	model  *C.struct_llama_model
-	ctx    *C.struct_llama_context
-
-	last   []C.llama_token
-	embd   []C.llama_token
-	cursor int
-
-	mu sync.Mutex
-	gc bool
-
-	api.Options
-}
-
-func New(model string, opts api.Options) (*LLM, error) {
-	if _, err := os.Stat(model); err != nil {
-		return nil, err
-	}
-
-	llm := LLM{Options: opts}
-
-	C.llama_backend_init(C.bool(llm.UseNUMA))
-
-	params := C.llama_context_default_params()
-	params.seed = C.uint(llm.Seed)
-	params.n_ctx = C.int(llm.NumCtx)
-	params.n_batch = C.int(llm.NumBatch)
-	params.n_gqa = C.int(llm.NumGQA)
-	params.n_gpu_layers = C.int(llm.NumGPU)
-	params.main_gpu = C.int(llm.MainGPU)
-	params.low_vram = C.bool(llm.LowVRAM)
-	params.f16_kv = C.bool(llm.F16KV)
-	params.logits_all = C.bool(llm.LogitsAll)
-	params.vocab_only = C.bool(llm.VocabOnly)
-	params.use_mmap = C.bool(llm.UseMMap)
-	params.use_mlock = C.bool(llm.UseMLock)
-	params.embedding = C.bool(llm.EmbeddingOnly)
-	params.rope_freq_base = C.float(llm.RopeFrequencyBase)
-	params.rope_freq_scale = C.float(llm.RopeFrequencyScale)
-	llm.params = &params
-
-	cModel := C.CString(model)
-	defer C.free(unsafe.Pointer(cModel))
-
-	llm.model = C.llama_load_model_from_file(cModel, params)
-	if llm.model == nil {
-		return nil, errors.New("failed to load model")
-	}
-
-	llm.ctx = C.llama_new_context_with_model(llm.model, params)
-	if llm.ctx == nil {
-		return nil, errors.New("failed to create context")
-	}
-
-	// warm up the model
-	bos := []C.llama_token{C.llama_token_bos()}
-	C.llama_eval(llm.ctx, unsafe.SliceData(bos), C.int(len(bos)), 0, C.int(opts.NumThread))
-	C.llama_reset_timings(llm.ctx)
-
-	return &llm, nil
-}
-
-func (llm *LLM) Close() {
-	llm.gc = true
-
-	llm.mu.Lock()
-	defer llm.mu.Unlock()
-
-	defer C.llama_free_model(llm.model)
-	defer C.llama_free(llm.ctx)
-
-	C.llama_print_timings(llm.ctx)
-}
-
-var errNeedMoreData = errors.New("need more data")
-
-func (llm *LLM) Predict(ctx []int, prompt string, fn func(api.GenerateResponse)) error {
-	C.llama_reset_timings(llm.ctx)
-
-	tokens := make([]C.llama_token, len(ctx))
-	for i := range tokens {
-		tokens[i] = C.llama_token(ctx[i])
-	}
-
-	llm.marshalPrompt(tokens, prompt)
-
-	C.llama_set_rng_seed(llm.ctx, C.uint(llm.Seed))
-
-	var b bytes.Buffer
-	for {
-		token, err := llm.next()
-		if llm.gc {
-			return nil
-		} else if errors.Is(err, io.EOF) {
-			break
-		} else if err != nil {
-			return err
-		}
-
-		b.WriteString(llm.Decode(token))
-
-		if err := llm.checkStopConditions(b); err != nil {
-			if errors.Is(err, io.EOF) {
-				break
-			} else if errors.Is(err, errNeedMoreData) {
-				continue
-			}
-
-			return err
-		}
-
-		if utf8.Valid(b.Bytes()) || b.Len() >= utf8.UTFMax {
-			fn(api.GenerateResponse{Response: b.String()})
-			b.Reset()
-		}
-	}
-
-	embd := make([]int, len(llm.embd))
-	for i := range llm.embd {
-		embd[i] = int(llm.embd[i])
-	}
-
-	timings := C.llama_get_timings(llm.ctx)
-	fn(api.GenerateResponse{
-		Done:               true,
-		Context:            embd,
-		SampleCount:        int(timings.n_sample),
-		SampleDuration:     parseDurationMs(float64(timings.t_sample_ms)),
-		PromptEvalCount:    int(timings.n_p_eval),
-		PromptEvalDuration: parseDurationMs(float64(timings.t_p_eval_ms)),
-		EvalCount:          int(timings.n_eval),
-		EvalDuration:       parseDurationMs(float64(timings.t_eval_ms)),
-	})
-
-	return nil
-}
-
-func (llm *LLM) checkStopConditions(b bytes.Buffer) error {
-	for _, stopCondition := range llm.Stop {
-		if stopCondition == strings.TrimSpace(b.String()) {
-			return io.EOF
-		} else if strings.HasPrefix(stopCondition, strings.TrimSpace(b.String())) {
-			return errNeedMoreData
-		}
-	}
-
-	return nil
-}
-
-func (llm *LLM) marshalPrompt(ctx []C.llama_token, prompt string) []C.llama_token {
-	tokens := append(ctx, llm.Encode(prompt)...)
-	if llm.NumKeep < 0 {
-		llm.NumKeep = len(tokens)
-	}
-
-	// min(llm.NumCtx - 4, llm.NumKeep)
-	if llm.NumCtx-4 < llm.NumKeep {
-		llm.NumKeep = llm.NumCtx - 4
-	}
-
-	if len(tokens) >= llm.NumCtx {
-		// truncate input
-		numLeft := (llm.NumCtx - llm.NumKeep) / 2
-		truncated := tokens[:llm.NumKeep]
-		erasedBlocks := (len(tokens) - llm.NumKeep - numLeft - 1) / numLeft
-		truncated = append(truncated, tokens[llm.NumKeep+erasedBlocks*numLeft:]...)
-		copy(llm.last, tokens[len(tokens)-llm.NumCtx:])
-
-		tokens = truncated
-		log.Printf("input truncated: num_ctx=%d num_keep=%d num_left=%d num_tokens=%d", llm.NumCtx, llm.NumKeep, numLeft, len(truncated))
-	} else {
-		llm.last = make([]C.llama_token, llm.NumCtx-len(tokens))
-		llm.last = append(llm.last, tokens...)
-	}
-
-	var i int
-	for i = 0; i < len(llm.embd) && i < len(tokens) && llm.embd[i] == tokens[i]; i++ {
-		// noop
-	}
-
-	llm.embd = tokens
-	if i == len(tokens) {
-		// evaluate at least one token to generate logits
-		i--
-	}
-
-	llm.cursor = i
-
-	log.Printf("prompt: num_past=%d cached=%v eval=%v", i, len(llm.embd[:i]), len(llm.embd[i:]))
-	return tokens
-}
-
-func (llm *LLM) Encode(prompt string) []C.llama_token {
-	cPrompt := C.CString(prompt)
-	defer C.free(unsafe.Pointer(cPrompt))
-
-	tokens := make([]C.llama_token, len(prompt)+1)
-	if n := C.llama_tokenize(llm.ctx, cPrompt, unsafe.SliceData(tokens), C.int(len(tokens)), true); n > 0 {
-		return tokens[:n]
-	}
-
-	return nil
-}
-
-func (llm *LLM) Decode(tokens ...C.llama_token) string {
-	var sb strings.Builder
-	for _, token := range tokens {
-		sb.WriteString(C.GoString(C.llama_token_to_str(llm.ctx, token)))
-	}
-
-	return sb.String()
-}
-
-func (llm *LLM) next() (C.llama_token, error) {
-	llm.mu.Lock()
-	defer llm.mu.Unlock()
-
-	if len(llm.embd) >= llm.NumCtx {
-		numLeft := (llm.NumCtx - llm.NumKeep) / 2
-		truncated := llm.embd[:llm.NumKeep]
-		truncated = append(truncated, llm.embd[len(llm.embd)-numLeft:]...)
-
-		llm.embd = truncated
-		llm.cursor = llm.NumKeep
-		log.Printf("input truncated: num_ctx=%d num_keep=%d num_left=%d num_tokens=%d cursor=%d", llm.NumCtx, llm.NumKeep, numLeft, len(truncated), llm.cursor)
-	}
-
-	for {
-		if llm.gc {
-			return 0, io.EOF
-		}
-
-		if llm.cursor >= len(llm.embd) {
-			break
-		}
-
-		numEval := len(llm.embd) - llm.cursor
-		if numEval > llm.NumBatch {
-			numEval = llm.NumBatch
-		}
-
-		if retval := C.llama_eval(llm.ctx, unsafe.SliceData(llm.embd[llm.cursor:]), C.int(numEval), C.int(llm.cursor), C.int(llm.NumThread)); retval != 0 {
-			return 0, fmt.Errorf("llama_eval: %d", retval)
-		}
-
-		llm.cursor += numEval
-	}
-
-	var sampleOpts C.struct_llama_sample_options
-	sampleOpts.repeat_penalty = C.float(llm.RepeatPenalty)
-	sampleOpts.frequency_penalty = C.float(llm.FrequencyPenalty)
-	sampleOpts.presence_penalty = C.float(llm.PresencePenalty)
-	sampleOpts.temperature = C.float(llm.Temperature)
-	sampleOpts.top_k = C.int(llm.TopK)
-	sampleOpts.top_p = C.float(llm.TopP)
-	sampleOpts.tfs_z = C.float(llm.TFSZ)
-	sampleOpts.typical_p = C.float(llm.TypicalP)
-	sampleOpts.mirostat = C.int(llm.Mirostat)
-	sampleOpts.mirostat_tau = C.float(llm.MirostatTau)
-	sampleOpts.mirostat_eta = C.float(llm.MirostatEta)
-	sampleOpts.penalize_newline = C.bool(llm.PenalizeNewline)
-
-	numVocab := C.llama_n_vocab(llm.ctx)
-	logits := unsafe.Slice(C.llama_get_logits(llm.ctx), numVocab)
-
-	// TODO: logit bias
-
-	candidates := make([]C.llama_token_data, numVocab)
-	for i := range logits {
-		candidates[i] = C.llama_token_data{
-			id:    C.int(i),
-			logit: logits[i],
-			p:     0,
-		}
-	}
-
-	repeatLastN := llm.RepeatLastN
-	if len(llm.last) < repeatLastN {
-		repeatLastN = len(llm.last)
-	}
-
-	if llm.NumCtx < repeatLastN {
-		repeatLastN = llm.NumCtx
-	}
-
-	lastN := llm.last[len(llm.last)-repeatLastN:]
-
-	token := C.llama_sample(
-		llm.ctx,
-		unsafe.SliceData(candidates), C.size_t(len(candidates)),
-		unsafe.SliceData(lastN), C.size_t(len(lastN)),
-		&sampleOpts,
-	)
-
-	llm.last = append(llm.last, token)
-	llm.embd = append(llm.embd, token)
-
-	if token == C.llama_token_eos() {
-		return 0, io.EOF
-	}
-
-	return token, nil
-}
-
-func (llm *LLM) Embedding(input string) ([]float64, error) {
-	if !llm.EmbeddingOnly {
-		return nil, errors.New("llama: embedding not enabled")
-	}
-
-	tokens := llm.Encode(input)
-	if tokens == nil {
-		return nil, errors.New("llama: tokenize embedding")
-	}
-
-	retval := C.llama_eval(llm.ctx, unsafe.SliceData(tokens), C.int(len(tokens)), 0, C.int(llm.NumThread))
-	if retval != 0 {
-		return nil, errors.New("llama: eval")
-	}
-
-	n := C.llama_n_embd(llm.ctx)
-	if n <= 0 {
-		return nil, errors.New("llama: no embeddings generated")
-	}
-	cEmbeddings := unsafe.Slice(C.llama_get_embeddings(llm.ctx), n)
-
-	embeddings := make([]float64, len(cEmbeddings))
-	for i, v := range cEmbeddings {
-		embeddings[i] = float64(v)
-	}
-	return embeddings, nil
-}

llama/llama.h

@@ -1,495 +0,0 @@
-/**
- * llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12
- *
- * MIT License
- *
- * Copyright (c) 2023 Georgi Gerganov
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#ifndef LLAMA_H
-#define LLAMA_H
-
-#include "ggml.h"
-#ifdef GGML_USE_CUBLAS
-#include "ggml-cuda.h"
-#define LLAMA_MAX_DEVICES GGML_CUDA_MAX_DEVICES
-#else
-#define LLAMA_MAX_DEVICES 1
-#endif // GGML_USE_CUBLAS
-#include <stddef.h>
-#include <stdint.h>
-#include <stdbool.h>
-
-#ifdef LLAMA_SHARED
-#    if defined(_WIN32) && !defined(__MINGW32__)
-#        ifdef LLAMA_BUILD
-#            define LLAMA_API __declspec(dllexport)
-#        else
-#            define LLAMA_API __declspec(dllimport)
-#        endif
-#    else
-#        define LLAMA_API __attribute__ ((visibility ("default")))
-#    endif
-#else
-#    define LLAMA_API
-#endif
-
-#ifdef __GNUC__
-#    define DEPRECATED(func, hint) func __attribute__((deprecated(hint)))
-#elif defined(_MSC_VER)
-#    define DEPRECATED(func, hint) __declspec(deprecated(hint)) func
-#else
-#    define DEPRECATED(func, hint) func
-#endif
-
-#define LLAMA_FILE_MAGIC_GGJT        0x67676a74u // 'ggjt'
-#define LLAMA_FILE_MAGIC_GGLA        0x67676c61u // 'ggla'
-#define LLAMA_FILE_MAGIC_GGMF        0x67676d66u // 'ggmf'
-#define LLAMA_FILE_MAGIC_GGML        0x67676d6cu // 'ggml'
-#define LLAMA_FILE_MAGIC_GGSN        0x6767736eu // 'ggsn'
-
-#define LLAMA_FILE_VERSION           3
-#define LLAMA_FILE_MAGIC             LLAMA_FILE_MAGIC_GGJT
-#define LLAMA_FILE_MAGIC_UNVERSIONED LLAMA_FILE_MAGIC_GGML
-#define LLAMA_SESSION_MAGIC          LLAMA_FILE_MAGIC_GGSN
-#define LLAMA_SESSION_VERSION        1
-
-#define LLAMA_DEFAULT_SEED           0xFFFFFFFF
-
-#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL)
-// Defined when llama.cpp is compiled with support for offloading model layers to GPU.
-#define LLAMA_SUPPORTS_GPU_OFFLOAD
-#endif
-
-#ifndef LLAMA_DEFAULT_RMS_EPS
-#define LLAMA_DEFAULT_RMS_EPS 5e-6f
-#endif
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-    //
-    // C interface
-    //
-    // TODO: show sample usage
-    //
-
-    struct llama_model;
-    struct llama_context;
-
-    typedef int llama_token;
-
-    typedef struct llama_token_data {
-        llama_token id; // token id
-        float logit;    // log-odds of the token
-        float p;        // probability of the token
-    } llama_token_data;
-
-    typedef struct llama_token_data_array {
-        llama_token_data * data;
-        size_t size;
-        bool sorted;
-    } llama_token_data_array;
-
-    typedef void (*llama_progress_callback)(float progress, void *ctx);
-
-   struct llama_context_params {
-        uint32_t seed;         // RNG seed, -1 for random
-        int32_t  n_ctx;        // text context
-        int32_t  n_batch;      // prompt processing batch size
-        int32_t  n_gqa;        // grouped-query attention (TEMP - will be moved to model hparams)
-        float    rms_norm_eps; // rms norm epsilon (TEMP - will be moved to model hparams)
-        int32_t  n_gpu_layers; // number of layers to store in VRAM
-        int32_t  main_gpu;     // the GPU that is used for scratch and small tensors
-
-        const float * tensor_split; // how to split layers across multiple GPUs (size: LLAMA_MAX_DEVICES)
-
-        // ref: https://github.com/ggerganov/llama.cpp/pull/2054
-        float    rope_freq_base;  // RoPE base frequency
-        float    rope_freq_scale; // RoPE frequency scaling factor
-
-        // called with a progress value between 0 and 1, pass NULL to disable
-        llama_progress_callback progress_callback;
-        // context pointer passed to the progress callback
-        void * progress_callback_user_data;
-
-        // Keep the booleans together to avoid misalignment during copy-by-value.
-        bool low_vram;   // if true, reduce VRAM usage at the cost of performance
-        bool mul_mat_q;  // if true, use experimental mul_mat_q kernels
-        bool f16_kv;     // use fp16 for KV cache
-        bool logits_all; // the llama_eval() call computes all logits, not just the last one
-        bool vocab_only; // only load the vocabulary, no weights
-        bool use_mmap;   // use mmap if possible
-        bool use_mlock;  // force system to keep model in RAM
-        bool embedding;  // embedding mode only
-    };
-    // model file types
-    enum llama_ftype {
-        LLAMA_FTYPE_ALL_F32              = 0,
-        LLAMA_FTYPE_MOSTLY_F16           = 1, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_0          = 2, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_1          = 3, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
-        // LLAMA_FTYPE_MOSTLY_Q4_2       = 5, // support has been removed
-        // LLAMA_FTYPE_MOSTLY_Q4_3       = 6, // support has been removed
-        LLAMA_FTYPE_MOSTLY_Q8_0          = 7, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_0          = 8, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_1          = 9, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q2_K          = 10,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q3_K_S        = 11,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q3_K_M        = 12,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q3_K_L        = 13,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_K_S        = 14,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_K_M        = 15,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_K_S        = 16,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_K_M        = 17,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q6_K          = 18,// except 1d tensors
-    };
-
-    // model quantization parameters
-    typedef struct llama_model_quantize_params {
-        int nthread;                 // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
-        enum llama_ftype   ftype;    // quantize to this llama_ftype
-        bool allow_requantize;       // allow quantizing non-f32/f16 tensors
-        bool quantize_output_tensor; // quantize output.weight
-    } llama_model_quantize_params;
-
-    // grammar types
-    struct llama_grammar;
-
-    // grammar element type
-    enum llama_gretype {
-        // end of rule definition
-        LLAMA_GRETYPE_END            = 0,
-
-        // start of alternate definition for rule
-        LLAMA_GRETYPE_ALT            = 1,
-
-        // non-terminal element: reference to rule
-        LLAMA_GRETYPE_RULE_REF       = 2,
-
-        // terminal element: character (code point)
-        LLAMA_GRETYPE_CHAR           = 3,
-
-        // inverse char(s) ([^a], [^a-b] [^abc])
-        LLAMA_GRETYPE_CHAR_NOT       = 4,
-
-        // modifies a preceding LLAMA_GRETYPE_CHAR or LLAMA_GRETYPE_CHAR_ALT to
-        // be an inclusive range ([a-z])
-        LLAMA_GRETYPE_CHAR_RNG_UPPER = 5,
-
-        // modifies a preceding LLAMA_GRETYPE_CHAR or
-        // LLAMA_GRETYPE_CHAR_RNG_UPPER to add an alternate char to match ([ab], [a-zA])
-        LLAMA_GRETYPE_CHAR_ALT       = 6,
-    };
-
-    typedef struct llama_grammar_element {
-        enum llama_gretype type;
-        uint32_t           value; // Unicode code point or rule ID
-    } llama_grammar_element;
-
-    // performance timing information
-    struct llama_timings {
-        double t_start_ms;
-        double t_end_ms;
-        double t_load_ms;
-        double t_sample_ms;
-        double t_p_eval_ms;
-        double t_eval_ms;
-
-        int32_t n_sample;
-        int32_t n_p_eval;
-        int32_t n_eval;
-    };
-
-    LLAMA_API int llama_max_devices();
-
-    LLAMA_API struct llama_context_params llama_context_default_params();
-    LLAMA_API struct llama_model_quantize_params llama_model_quantize_default_params();
-
-    LLAMA_API bool llama_mmap_supported();
-    LLAMA_API bool llama_mlock_supported();
-
-    // TODO: not great API - very likely to change
-    // Initialize the llama + ggml backend
-    // If numa is true, use NUMA optimizations
-    // Call once at the start of the program
-    LLAMA_API void llama_backend_init(bool numa);
-    // Call once at the end of the program - currently only used for MPI
-    LLAMA_API void llama_backend_free();
-
-    LLAMA_API int64_t llama_time_us();
-
-    LLAMA_API struct llama_model * llama_load_model_from_file(
-                             const char * path_model,
-            struct llama_context_params   params);
-
-    LLAMA_API void llama_free_model(struct llama_model * model);
-
-    LLAMA_API struct llama_context * llama_new_context_with_model(
-                     struct llama_model * model,
-            struct llama_context_params   params);
-
-    // Various functions for loading a ggml llama model.
-    // Allocate (almost) all memory needed for the model.
-    // Return NULL on failure
-    LLAMA_API DEPRECATED(struct llama_context * llama_init_from_file(
-                             const char * path_model,
-            struct llama_context_params   params),
-            "please use llama_load_model_from_file combined with llama_new_context_with_model instead");
-
-    // Frees all allocated memory
-    LLAMA_API void llama_free(struct llama_context * ctx);
-
-    // Returns 0 on success
-    LLAMA_API int llama_model_quantize(
-            const char * fname_inp,
-            const char * fname_out,
-            const llama_model_quantize_params * params);
-
-    // Apply a LoRA adapter to a loaded model
-    // path_base_model is the path to a higher quality model to use as a base for
-    // the layers modified by the adapter. Can be NULL to use the current loaded model.
-    // The model needs to be reloaded before applying a new adapter, otherwise the adapter
-    // will be applied on top of the previous one
-    // Returns 0 on success
-    LLAMA_API DEPRECATED(int llama_apply_lora_from_file(
-            struct llama_context * ctx,
-                      const char * path_lora,
-                      const char * path_base_model,
-                             int   n_threads),
-            "please use llama_model_apply_lora_from_file instead");
-
-    LLAMA_API int llama_model_apply_lora_from_file(
-            const struct llama_model * model,
-                      const char * path_lora,
-                      const char * path_base_model,
-                             int   n_threads);
-
-    // Returns the number of tokens in the KV cache
-    LLAMA_API int llama_get_kv_cache_token_count(const struct llama_context * ctx);
-
-    // Sets the current rng seed.
-    LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed);
-
-    // Returns the maximum size in bytes of the state (rng, logits, embedding
-    // and kv_cache) - will often be smaller after compacting tokens
-    LLAMA_API size_t llama_get_state_size(const struct llama_context * ctx);
-
-    // Copies the state to the specified destination address.
-    // Destination needs to have allocated enough memory.
-    // Returns the number of bytes copied
-    LLAMA_API size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst);
-
-    // Set the state reading from the specified address
-    // Returns the number of bytes read
-    LLAMA_API size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src);
-
-    // Save/load session file
-    LLAMA_API bool llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out);
-    LLAMA_API bool llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count);
-
-    // Run the llama inference to obtain the logits and probabilities for the next token.
-    // tokens + n_tokens is the provided batch of new tokens to process
-    // n_past is the number of tokens to use from previous eval calls
-    // Returns 0 on success
-    LLAMA_API int llama_eval(
-            struct llama_context * ctx,
-               const llama_token * tokens,
-                             int   n_tokens,
-                             int   n_past,
-                             int   n_threads);
-
-    // Same as llama_eval, but use float matrix input directly.
-    LLAMA_API int llama_eval_embd(
-            struct llama_context * ctx,
-                     const float * embd,
-                             int   n_tokens,
-                             int   n_past,
-                             int   n_threads);
-
-    // Export a static computation graph for context of 511 and batch size of 1
-    // NOTE: since this functionality is mostly for debugging and demonstration purposes, we hardcode these
-    //       parameters here to keep things simple
-    // IMPORTANT: do not use for anything else other than debugging and testing!
-    LLAMA_API int llama_eval_export(struct llama_context * ctx, const char * fname);
-
-    // Convert the provided text into tokens.
-    // The tokens pointer must be large enough to hold the resulting tokens.
-    // Returns the number of tokens on success, no more than n_max_tokens
-    // Returns a negative number on failure - the number of tokens that would have been returned
-    // TODO: not sure if correct
-    LLAMA_API int llama_tokenize(
-            struct llama_context * ctx,
-                      const char * text,
-                     llama_token * tokens,
-                             int   n_max_tokens,
-                            bool   add_bos);
-
-    LLAMA_API int llama_tokenize_with_model(
-        const struct llama_model * model,
-                      const char * text,
-                     llama_token * tokens,
-                             int   n_max_tokens,
-                            bool   add_bos);
-
-    LLAMA_API int llama_n_vocab(const struct llama_context * ctx);
-    LLAMA_API int llama_n_ctx  (const struct llama_context * ctx);
-    LLAMA_API int llama_n_embd (const struct llama_context * ctx);
-
-    LLAMA_API int llama_n_vocab_from_model(const struct llama_model * model);
-    LLAMA_API int llama_n_ctx_from_model  (const struct llama_model * model);
-    LLAMA_API int llama_n_embd_from_model (const struct llama_model * model);
-
-    // Get the vocabulary as output parameters.
-    // Returns number of results.
-    LLAMA_API int llama_get_vocab(
-            const struct llama_context * ctx,
-                          const char * * strings,
-                                 float * scores,
-                                   int   capacity);
-
-    LLAMA_API int llama_get_vocab_from_model(
-              const struct llama_model * model,
-                          const char * * strings,
-                                 float * scores,
-                                   int   capacity);
-
-    // Token logits obtained from the last call to llama_eval()
-    // The logits for the last token are stored in the last row
-    // Can be mutated in order to change the probabilities of the next token
-    // Rows: n_tokens
-    // Cols: n_vocab
-    LLAMA_API float * llama_get_logits(struct llama_context * ctx);
-
-    // Get the embeddings for the input
-    // shape: [n_embd] (1-dimensional)
-    LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
-
-    // Token Id -> String. Uses the vocabulary in the provided context
-    LLAMA_API const char * llama_token_to_str(
-            const struct llama_context * ctx,
-                           llama_token   token);
-
-    LLAMA_API const char * llama_token_to_str_with_model(
-              const struct llama_model * model,
-                           llama_token   token);
-
-    // Special tokens
-    LLAMA_API llama_token llama_token_bos();  // beginning-of-sentence
-    LLAMA_API llama_token llama_token_eos();  // end-of-sentence
-    LLAMA_API llama_token llama_token_nl();   // next-line
-
-    // Grammar
-    //
-    LLAMA_API struct llama_grammar * llama_grammar_init(
-            const llama_grammar_element ** rules,
-                                 size_t    n_rules,
-                                 size_t    start_rule_index);
-
-    LLAMA_API void llama_grammar_free(struct llama_grammar * grammar);
-
-    // Sampling functions
-
-    /// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix.
-    LLAMA_API void llama_sample_repetition_penalty(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float penalty);
-
-    /// @details Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details.
-    LLAMA_API void llama_sample_frequency_and_presence_penalties(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float alpha_frequency, float alpha_presence);
-
-    /// @details Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806
-    /// @param candidates A vector of `llama_token_data` containing the candidate tokens, the logits must be directly extracted from the original generation context without being sorted.
-    /// @params guidance_ctx A separate context from the same model. Other than a negative prompt at the beginning, it should have all generated and user input tokens copied from the main context.
-    /// @params scale Guidance strength. 1.0f means no guidance. Higher values mean stronger guidance.
-    LLAMA_API void llama_sample_classifier_free_guidance(
-              struct llama_context * ctx,
-            llama_token_data_array * candidates,
-              struct llama_context * guidance_ctx,
-                             float   scale);
-
-    /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
-    LLAMA_API void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * candidates);
-
-    /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
-    LLAMA_API void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * candidates, int k, size_t min_keep);
-
-    /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
-    LLAMA_API void llama_sample_top_p(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
-
-    /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
-    LLAMA_API void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array * candidates, float z, size_t min_keep);
-
-    /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
-    LLAMA_API void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
-    LLAMA_API void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array * candidates, float temp);
-
-    /// @details Apply constraints from grammar
-    LLAMA_API void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * candidates, const struct llama_grammar * grammar);
-
-    /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
-    /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
-    /// @param tau  The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
-    /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
-    /// @param m The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm.
-    /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
-    LLAMA_API llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, int m, float * mu);
-
-    /// @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
-    /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
-    /// @param tau  The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
-    /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
-    /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
-    LLAMA_API llama_token llama_sample_token_mirostat_v2(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, float * mu);
-
-    /// @details Selects the token with the highest probability.
-    LLAMA_API llama_token llama_sample_token_greedy(struct llama_context * ctx, llama_token_data_array * candidates);
-
-    /// @details Randomly selects a token from the candidates based on their probabilities.
-    LLAMA_API llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_array * candidates);
-
-    /// @details Accepts the sampled token into the grammar
-    LLAMA_API void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar * grammar, llama_token token);
-
-    // Performance information
-    LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx);
-    LLAMA_API void llama_print_timings(struct llama_context * ctx);
-    LLAMA_API void llama_reset_timings(struct llama_context * ctx);
-
-    // Print system information
-    LLAMA_API const char * llama_print_system_info(void);
-
-#ifdef __cplusplus
-}
-#endif
-
-// Internal API to be implemented by llama.cpp and used by tests/benchmarks only
-#ifdef LLAMA_API_INTERNAL
-
-#include <vector>
-#include <string>
-struct ggml_tensor;
-
-const std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx);
-
-#endif
-
-#endif // LLAMA_H

llama/llama_darwin.go

@@ -1,80 +0,0 @@
-package llama
-
-import (
-	"bytes"
-	"crypto/sha256"
-	"errors"
-	"io"
-	"log"
-	"os"
-	"path/filepath"
-)
-
-func init() {
-	if err := initBackend(); err != nil {
-		log.Printf("WARNING: GPU could not be initialized correctly: %v", err)
-		log.Printf("WARNING: falling back to CPU")
-	}
-}
-
-func initBackend() error {
-	exec, err := os.Executable()
-	if err != nil {
-		return err
-	}
-
-	exec, err = filepath.EvalSymlinks(exec)
-	if err != nil {
-		return err
-	}
-
-	metal := filepath.Join(filepath.Dir(exec), "ggml-metal.metal")
-	fi, err := os.Stat(metal)
-	if err != nil && !errors.Is(err, os.ErrNotExist) {
-		return err
-	}
-
-	if fi != nil {
-		actual, err := os.Open(metal)
-		if err != nil {
-			return err
-		}
-
-		actualSum := sha256.New()
-		if _, err := io.Copy(actualSum, actual); err != nil {
-			return err
-		}
-
-		expect, err := fs.Open("ggml-metal.metal")
-		if err != nil {
-			return err
-		}
-
-		expectSum := sha256.New()
-		if _, err := io.Copy(expectSum, expect); err != nil {
-			return err
-		}
-
-		if bytes.Equal(actualSum.Sum(nil), expectSum.Sum(nil)) {
-			return nil
-		}
-	}
-
-	dst, err := os.Create(filepath.Join(filepath.Dir(exec), "ggml-metal.metal"))
-	if err != nil {
-		return err
-	}
-	defer dst.Close()
-
-	src, err := fs.Open("ggml-metal.metal")
-	if err != nil {
-		return err
-	}
-	defer src.Close()
-
-	if _, err := io.Copy(dst, src); err != nil {
-		return err
-	}
-
-	return nil
-}

llama/update-llama-cpp.sh

@@ -1,70 +0,0 @@
-#!/bin/sh
-
-set -eu
-
-
-status() { echo >&2 ">>> $*"; }
-error() { status "ERROR $*"; }
-usage() {
-    echo "usage: $(basename $0) /path/to/repo"
-    exit 1
-}
-
-OUT=$(dirname $0)
-while getopts "hC:" OPTION; do
-    case $OPTION in
-        C) OUT=$OPTARG ;;
-        *) usage ;;
-    esac
-done
-
-shift $(( $OPTIND - 1 ))
-[ $# -eq 1 ] || usage
-
-status "updating source..."
-cp -a "$1"/*.{c,h,cpp,m,metal,cu} "$OUT"
-
-status "removing incompatible files..."
-rm -f "$OUT"/build-info.h
-
-SHA1=$(git -C $1 rev-parse @)
-
-LICENSE=$(mktemp)
-cleanup() {
-    rm -f $LICENSE
-}
-trap cleanup 0
-
-cat <<EOF | sed 's/ *$//' >$LICENSE
-/**
- * llama.cpp - git $SHA1
- *
-$(sed 's/^/ * /' <$1/LICENSE)
- */
-
-EOF
-
-for IN in $OUT/*.{c,h,cpp,m,metal,cu}; do
-    TMP=$(mktemp)
-    status "updating license $IN"
-    cat $LICENSE $IN >$TMP
-    mv $TMP $IN
-done
-
-touchup() {
-    local CONSTRAINT=$1 && shift
-
-    for IN in $*; do
-        status "touching up $IN..."
-        TMP=$(mktemp)
-        {
-            echo "//go:build $CONSTRAINT"
-            echo
-        } | cat - $IN >$TMP
-        mv $TMP $IN
-    done
-}
-
-touchup darwin $OUT/ggml-metal.*
-touchup mpi $OUT/ggml-mpi.*
-touchup opencl $OUT/ggml-opencl.*

llm/ggml.go

@@ -0,0 +1,185 @@
+package llm
+
+import (
+	"encoding/binary"
+	"errors"
+	"fmt"
+	"io"
+)
+
+type ModelFamily string
+
+type ModelType uint32
+
+const (
+	ModelType3B  ModelType = 26
+	ModelType7B  ModelType = 32
+	ModelType13B ModelType = 40
+	ModelType34B ModelType = 48
+	ModelType30B ModelType = 60
+	ModelType65B ModelType = 80
+)
+
+func (mt ModelType) String() string {
+	switch mt {
+	case ModelType3B:
+		return "3B"
+	case ModelType7B:
+		return "7B"
+	case ModelType13B:
+		return "13B"
+	case ModelType34B:
+		return "34B"
+	case ModelType30B:
+		return "30B"
+	case ModelType65B:
+		return "65B"
+	default:
+		return "Unknown"
+	}
+}
+
+type FileType interface {
+	String() string
+}
+
+type GGML struct {
+	magic uint32
+	container
+	model
+}
+
+type model interface {
+	ModelFamily() ModelFamily
+	ModelType() ModelType
+	FileType() FileType
+}
+
+type container interface {
+	Name() string
+	Decode(io.Reader) error
+}
+
+type containerGGML struct {
+}
+
+func (c *containerGGML) Name() string {
+	return "ggml"
+}
+
+func (c *containerGGML) Decode(r io.Reader) error {
+	return nil
+}
+
+type containerGGMF struct {
+	version uint32
+}
+
+func (c *containerGGMF) Name() string {
+	return "ggmf"
+}
+
+func (c *containerGGMF) Decode(r io.Reader) error {
+	var version uint32
+	binary.Read(r, binary.LittleEndian, &version)
+
+	switch version {
+	case 1:
+	default:
+		return errors.New("invalid version")
+	}
+
+	c.version = version
+	return nil
+}
+
+type containerGGJT struct {
+	version uint32
+}
+
+func (c *containerGGJT) Name() string {
+	return "ggjt"
+}
+
+func (c *containerGGJT) Decode(r io.Reader) error {
+	var version uint32
+	binary.Read(r, binary.LittleEndian, &version)
+
+	switch version {
+	case 1, 2, 3:
+	default:
+		return errors.New("invalid version")
+	}
+
+	c.version = version
+	return nil
+}
+
+type containerLORA struct {
+	version uint32
+}
+
+func (c *containerLORA) Name() string {
+	return "ggla"
+}
+
+func (c *containerLORA) Decode(r io.Reader) error {
+	var version uint32
+	binary.Read(r, binary.LittleEndian, &version)
+
+	switch version {
+	case 1:
+	default:
+		return errors.New("invalid version")
+	}
+
+	c.version = version
+	return nil
+}
+
+const (
+	// / Magic constant for `ggml` files (unversioned).
+	FILE_MAGIC_GGML = 0x67676d6c
+	// / Magic constant for `ggml` files (versioned, ggmf).
+	FILE_MAGIC_GGMF = 0x67676d66
+	// / Magic constant for `ggml` files (versioned, ggjt).
+	FILE_MAGIC_GGJT = 0x67676a74
+	// / Magic constant for `ggla` files (LoRA adapter).
+	FILE_MAGIC_GGLA = 0x67676C61
+)
+
+func DecodeGGML(r io.ReadSeeker, hint ModelFamily) (*GGML, error) {
+	var ggml GGML
+	binary.Read(r, binary.LittleEndian, &ggml.magic)
+
+	switch ggml.magic {
+	case FILE_MAGIC_GGML:
+		ggml.container = &containerGGML{}
+	case FILE_MAGIC_GGMF:
+		ggml.container = &containerGGMF{}
+	case FILE_MAGIC_GGJT:
+		ggml.container = &containerGGJT{}
+	case FILE_MAGIC_GGLA:
+		ggml.container = &containerLORA{}
+	default:
+		return nil, errors.New("invalid file magic")
+	}
+
+	if err := ggml.Decode(r); err != nil {
+		return nil, err
+	}
+
+	// different model types may have different layouts for hyperparameters
+	switch hint {
+	case ModelFamilyLlama:
+		var llama llamaModel
+		binary.Read(r, binary.LittleEndian, &llama.hyperparameters)
+		ggml.model = &llama
+		// TODO: sanity check hyperparameters
+	default:
+		return nil, fmt.Errorf("unsupported model type: %s", hint)
+	}
+
+	// final model type
+	return &ggml, nil
+}

llm/ggml_llama.go

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

llm/llama.cpp/generate.go

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

llm/llama.cpp/generate_darwin_amd64.go

@@ -0,0 +1,10 @@
+package llm
+
+//go:generate git submodule init
+//go:generate git submodule update --force ggml
+//go:generate git -C ggml apply ../ggml_patch/0001-add-detokenize-endpoint.patch
+//go:generate git -C ggml apply ../ggml_patch/0002-34B-model-support.patch
+//go:generate git -C ggml apply ../ggml_patch/0003-metal-fix-synchronization-in-new-matrix-multiplicati.patch
+//go:generate git -C ggml apply ../ggml_patch/0004-metal-add-missing-barriers-for-mul-mat-2699.patch
+//go:generate cmake --fresh -S ggml -B ggml/build/cpu -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=x86_64 -DCMAKE_OSX_ARCHITECTURES=x86_64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
+//go:generate cmake --build ggml/build/cpu --target server --config Release

llm/llama.cpp/generate_darwin_arm64.go

@@ -0,0 +1,10 @@
+package llm
+
+//go:generate git submodule init
+//go:generate git submodule update --force ggml
+//go:generate git -C ggml apply ../ggml_patch/0001-add-detokenize-endpoint.patch
+//go:generate git -C ggml apply ../ggml_patch/0002-34B-model-support.patch
+//go:generate git -C ggml apply ../ggml_patch/0003-metal-fix-synchronization-in-new-matrix-multiplicati.patch
+//go:generate git -C ggml apply ../ggml_patch/0004-metal-add-missing-barriers-for-mul-mat-2699.patch
+//go:generate cmake --fresh -S ggml -B ggml/build/gpu -DLLAMA_METAL=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=arm64 -DCMAKE_OSX_ARCHITECTURES=arm64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
+//go:generate cmake --build ggml/build/gpu --target server --config Release

llm/llama.cpp/ggml_patch/0001-add-detokenize-endpoint.patch

@@ -0,0 +1,51 @@
+From 032ef7ff2423f5117bb59d42fb71be9cebf0a2de Mon Sep 17 00:00:00 2001
+From: Bruce MacDonald <brucewmacdonald@gmail.com>
+Date: Mon, 28 Aug 2023 18:08:12 -0400
+Subject: [PATCH] add detokenize endpoint
+
+---
+ examples/server/server.cpp | 21 +++++++++++++++++++++
+ 1 file changed, 21 insertions(+)
+
+diff --git a/examples/server/server.cpp b/examples/server/server.cpp
+index 9966045..5014691 100644
+--- a/examples/server/server.cpp
++++ b/examples/server/server.cpp
+@@ -1075,6 +1075,12 @@ static json format_tokenizer_response(const std::vector<llama_token> &tokens)
+         {"tokens", tokens}};
+ }
+ 
++static json format_detokenized_response(std::string content)
++{
++    return json{
++        {"content", content}};
++}
++
+ static void parse_options_completion(const json &body, llama_server_context &llama)
+ {
+     gpt_params default_params;
+@@ -1361,6 +1367,21 @@ int main(int argc, char **argv)
+         const json data = format_tokenizer_response(tokens);
+         return res.set_content(data.dump(), "application/json"); });
+ 
++    svr.Post("/detokenize", [&llama](const Request &req, Response &res)
++             {
++        auto lock = llama.lock();
++
++        const json body = json::parse(req.body);
++        std::string content;
++        if (body.count("tokens") != 0)
++        {
++            const std::vector<llama_token> tokens = body["tokens"];
++            content = tokens_to_str(llama.ctx, tokens.cbegin(), tokens.cend());
++        }
++
++        const json data = format_detokenized_response(content);
++        return res.set_content(data.dump(), "application/json"); });
++
+     svr.Post("/embedding", [&llama](const Request &req, Response &res)
+              {
+         auto lock = llama.lock();
+-- 
+2.39.2 (Apple Git-143)
+

llm/llama.cpp/ggml_patch/0002-34B-model-support.patch

@@ -0,0 +1,89 @@
+From 6145068a6613c37bb43a7408b5496524bdcfc402 Mon Sep 17 00:00:00 2001
+From: Bruce MacDonald <brucewmacdonald@gmail.com>
+Date: Mon, 28 Aug 2023 18:08:53 -0400
+Subject: [PATCH] 34B model support
+
+---
+ llama.cpp | 10 ++++++++++
+ 1 file changed, 10 insertions(+)
+
+diff --git a/llama.cpp b/llama.cpp
+index f2cbe76..62c5cdf 100644
+--- a/llama.cpp
++++ b/llama.cpp
+@@ -79,6 +79,7 @@ enum e_model {
+     MODEL_7B,
+     MODEL_13B,
+     MODEL_30B,
++    MODEL_34B,
+     MODEL_65B,
+     MODEL_70B,
+ };
+@@ -122,6 +123,7 @@ static std::map<e_model, size_t> MEM_REQ_SCRATCH0(int n_ctx)
+         { MODEL_7B,   ((size_t) n_ctx / 16ull + 100ull) * MB },
+         { MODEL_13B,  ((size_t) n_ctx / 12ull + 120ull) * MB },
+         { MODEL_30B,  ((size_t) n_ctx /  9ull + 160ull) * MB },
++        { MODEL_34B,  ((size_t) n_ctx / 9ull + 160ull) * MB },
+         { MODEL_65B,  ((size_t) n_ctx /  6ull + 256ull) * MB }, // guess
+         { MODEL_70B,  ((size_t) n_ctx /  7ull + 164ull) * MB },
+     };
+@@ -135,6 +137,7 @@ static const std::map<e_model, size_t> & MEM_REQ_SCRATCH1()
+         { MODEL_7B,  160ull * MB },
+         { MODEL_13B, 192ull * MB },
+         { MODEL_30B, 256ull * MB },
++        { MODEL_34B, 256ull * MB },
+         { MODEL_65B, 384ull * MB }, // guess
+         { MODEL_70B, 304ull * MB },
+     };
+@@ -149,6 +152,7 @@ static const std::map<e_model, size_t> & MEM_REQ_EVAL()
+         { MODEL_7B,  10ull * MB },
+         { MODEL_13B, 12ull * MB },
+         { MODEL_30B, 16ull * MB },
++        { MODEL_34B, 16ull * MB },
+         { MODEL_65B, 24ull * MB }, // guess
+         { MODEL_70B, 24ull * MB },
+     };
+@@ -164,6 +168,7 @@ static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_BASE()
+         { MODEL_7B,   512ull * kB },
+         { MODEL_13B,  640ull * kB },
+         { MODEL_30B,  768ull * kB },
++        { MODEL_34B,  768ull * kB },
+         { MODEL_65B, 1280ull * kB },
+         { MODEL_70B, 1280ull * kB },
+     };
+@@ -179,6 +184,7 @@ static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_PER_CONTEXT()
+         { MODEL_7B,  128ull },
+         { MODEL_13B, 160ull },
+         { MODEL_30B, 208ull },
++        { MODEL_34B, 208ull },
+         { MODEL_65B, 256ull },
+         { MODEL_70B, 256ull },
+     };
+@@ -1027,6 +1033,7 @@ static const char * llama_model_type_name(e_model type) {
+         case MODEL_7B: return "7B";
+         case MODEL_13B: return "13B";
+         case MODEL_30B: return "30B";
++        case MODEL_34B: return "34B";
+         case MODEL_65B: return "65B";
+         case MODEL_70B: return "70B";
+         default: LLAMA_ASSERT(false);
+@@ -1074,6 +1081,7 @@ static void llama_model_load_internal(
+             case 26: model.type = e_model::MODEL_3B; break;
+             case 32: model.type = e_model::MODEL_7B; break;
+             case 40: model.type = e_model::MODEL_13B; break;
++            case 48: model.type = e_model::MODEL_34B; break;
+             case 60: model.type = e_model::MODEL_30B; break;
+             case 80: model.type = e_model::MODEL_65B; break;
+             default:
+@@ -1094,6 +1102,8 @@ static void llama_model_load_internal(
+             LLAMA_LOG_WARN("%s: warning: assuming 70B model based on GQA == %d\n", __func__, n_gqa);
+             model.type = e_model::MODEL_70B;
+             hparams.f_ffn_mult = 1.3f; // from the params.json of the 70B model
++        } else if (model.type == e_model::MODEL_34B && n_gqa == 8) {
++            hparams.f_ffn_mult = 1.0f; // from the params.json of the 34B model
+         }
+ 
+         hparams.rope_freq_base  = rope_freq_base;
+-- 
+2.39.2 (Apple Git-143)
+

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

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

llm/llama.cpp/ggml_patch/0003-metal-fix-synchronization-in-new-matrix-multiplicati.patch

@@ -0,0 +1,30 @@
+From dadbed99e65252d79f81101a392d0d6497b86caa Mon Sep 17 00:00:00 2001
+From: Shouzheng Liu <lshzh.hi@gmail.com>
+Date: Mon, 21 Aug 2023 06:59:29 -0400
+Subject: [PATCH] metal : fix synchronization in new matrix multiplication
+ kernel (#2686)
+
+---
+ ggml-metal.metal | 3 ++-
+ 1 file changed, 2 insertions(+), 1 deletion(-)
+
+diff --git a/ggml-metal.metal b/ggml-metal.metal
+index 3f31252..88d48f6 100644
+--- a/ggml-metal.metal
++++ b/ggml-metal.metal
+@@ -1898,10 +1898,11 @@ kernel void kernel_mul_mm(device const  uchar * src0,
+         threadgroup float *temp_str = ((threadgroup float *)shared_memory) \
+                                       + 32 * (sgitg&1) + (16 * (sgitg>>1)) * BLOCK_SIZE_M;
+         for (int i = 0; i < 8; i++) {
++            threadgroup_barrier(mem_flags::mem_device);
+             simdgroup_store(c_res[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
+         }
+ 
+-        threadgroup_barrier(mem_flags::mem_threadgroup);
++        threadgroup_barrier(mem_flags::mem_device);
+         device float *C = dst + BLOCK_SIZE_M * r0 + (BLOCK_SIZE_N * r1) * ne0 + im*ne1*ne0;
+         if (sgitg==0) {
+             for (int i = 0; i < n_rows; i++) {
+-- 
+2.41.0
+

llm/llama.cpp/ggml_patch/0004-metal-add-missing-barriers-for-mul-mat-2699.patch

@@ -0,0 +1,41 @@
+From 14b1d7e6f720dee41ce5a826376df738096d9033 Mon Sep 17 00:00:00 2001
+From: Shouzheng Liu <lshzh.hi@gmail.com>
+Date: Tue, 22 Aug 2023 02:18:40 -0400
+Subject: [PATCH] metal : add missing barriers for mul-mat (#2699)
+
+---
+ ggml-metal.metal | 5 +++--
+ 1 file changed, 3 insertions(+), 2 deletions(-)
+
+diff --git a/ggml-metal.metal b/ggml-metal.metal
+index 88d48f6..ce3541f 100644
+--- a/ggml-metal.metal
++++ b/ggml-metal.metal
+@@ -1850,6 +1850,7 @@ kernel void kernel_mul_mm(device const  uchar * src0,
+         //load data and store to threadgroup memory
+         half4x4 temp_a;
+         dequantize_func(x, il, temp_a);
++        threadgroup_barrier(mem_flags::mem_threadgroup);
+         #pragma unroll(16)
+         for (int i = 0; i < 16; i++) {
+             *(sa + SG_MAT_SIZE * ((tiitg / THREAD_PER_ROW / 8) \
+@@ -1895,14 +1896,14 @@ kernel void kernel_mul_mm(device const  uchar * src0,
+         }
+     } else {
+         // block is smaller than 64x32, we should avoid writing data outside of the matrix
++        threadgroup_barrier(mem_flags::mem_threadgroup);
+         threadgroup float *temp_str = ((threadgroup float *)shared_memory) \
+                                       + 32 * (sgitg&1) + (16 * (sgitg>>1)) * BLOCK_SIZE_M;
+         for (int i = 0; i < 8; i++) {
+-            threadgroup_barrier(mem_flags::mem_device);
+             simdgroup_store(c_res[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
+         }
+ 
+-        threadgroup_barrier(mem_flags::mem_device);
++        threadgroup_barrier(mem_flags::mem_threadgroup);
+         device float *C = dst + BLOCK_SIZE_M * r0 + (BLOCK_SIZE_N * r1) * ne0 + im*ne1*ne0;
+         if (sgitg==0) {
+             for (int i = 0; i < n_rows; i++) {
+-- 
+2.41.0
+

llm/llm.go

@@ -0,0 +1,84 @@
+package llm
+
+import (
+	"context"
+	"fmt"
+	"log"
+	"os"
+
+	"github.com/pbnjay/memory"
+
+	"github.com/jmorganca/ollama/api"
+)
+
+type LLM interface {
+	Predict(context.Context, []int, string, func(api.GenerateResponse)) error
+	Embedding(context.Context, string) ([]float64, error)
+	Encode(context.Context, string) ([]int, error)
+	Decode(context.Context, []int) (string, error)
+	SetOptions(api.Options)
+	Close()
+	Ping(context.Context) error
+}
+
+func New(model string, adapters []string, opts api.Options) (LLM, error) {
+	if _, err := os.Stat(model); err != nil {
+		return nil, err
+	}
+
+	f, err := os.Open(model)
+	if err != nil {
+		return nil, err
+	}
+	defer f.Close()
+
+	ggml, err := DecodeGGML(f, ModelFamilyLlama)
+	if err != nil {
+		return nil, err
+	}
+
+	switch ggml.FileType().String() {
+	case "F32", "Q5_0", "Q5_1", "Q8_0":
+		if opts.NumGPU != 0 {
+			// F32, F16, Q5_0, Q5_1, and Q8_0 do not support Metal API and will
+			// cause the runner to segmentation fault so disable GPU
+			log.Printf("WARNING: GPU disabled for F32, Q5_0, Q5_1, and Q8_0")
+			opts.NumGPU = 0
+		}
+	}
+
+	totalResidentMemory := memory.TotalMemory()
+	switch ggml.ModelType() {
+	case ModelType3B, ModelType7B:
+		if ggml.FileType().String() == "F16" && totalResidentMemory < 16*1024*1024 {
+			return nil, fmt.Errorf("F16 model requires at least 16GB of memory")
+		} else if totalResidentMemory < 8*1024*1024 {
+			return nil, fmt.Errorf("model requires at least 8GB of memory")
+		}
+	case ModelType13B:
+		if ggml.FileType().String() == "F16" && totalResidentMemory < 32*1024*1024 {
+			return nil, fmt.Errorf("F16 model requires at least 32GB of memory")
+		} else if totalResidentMemory < 16*1024*1024 {
+			return nil, fmt.Errorf("model requires at least 16GB of memory")
+		}
+	case ModelType30B, ModelType34B:
+		if ggml.FileType().String() == "F16" && totalResidentMemory < 64*1024*1024 {
+			return nil, fmt.Errorf("F16 model requires at least 64GB of memory")
+		} else if totalResidentMemory < 32*1024*1024 {
+			return nil, fmt.Errorf("model requires at least 32GB of memory")
+		}
+	case ModelType65B:
+		if ggml.FileType().String() == "F16" && totalResidentMemory < 128*1024*1024 {
+			return nil, fmt.Errorf("F16 model requires at least 128GB of memory")
+		} else if totalResidentMemory < 64*1024*1024 {
+			return nil, fmt.Errorf("model requires at least 64GB of memory")
+		}
+	}
+
+	switch ggml.ModelFamily() {
+	case ModelFamilyLlama:
+		return newLlama(model, adapters, ggmlRunner(), opts)
+	default:
+		return nil, fmt.Errorf("unknown ggml type: %s", ggml.ModelFamily())
+	}
+}

llm/utils.go

@@ -1,4 +1,4 @@
-package llama
+package llm
 
 import (
 	"fmt"

main.go

@@ -4,8 +4,9 @@ import (
 	"context"
 
 	"github.com/jmorganca/ollama/cmd"
+	"github.com/spf13/cobra"
 )
 
 func main() {
-	cmd.NewCLI().ExecuteContext(context.Background())
+	cobra.CheckErr(cmd.NewCLI().ExecuteContext(context.Background()))
 }

parser/parser.go

@@ -40,16 +40,22 @@ func Parse(reader io.Reader) ([]Command, error) {
 			command.Args = string(fields[1])
 			// copy command for validation
 			modelCommand = command
-		case "LICENSE", "TEMPLATE", "SYSTEM", "PROMPT", "EMBED":
+		case "LICENSE", "TEMPLATE", "SYSTEM", "PROMPT", "EMBED", "ADAPTER":
 			command.Name = string(bytes.ToLower(fields[0]))
 			command.Args = string(fields[1])
 		case "PARAMETER":
 			fields = bytes.SplitN(fields[1], []byte(" "), 2)
+			if len(fields) < 2 {
+				return nil, fmt.Errorf("missing value for %s", fields)
+			}
+
 			command.Name = string(fields[0])
 			command.Args = string(fields[1])
 		default:
-			// log a warning for unknown commands
-			log.Printf("WARNING: Unknown command: %s", fields[0])
+			if !bytes.HasPrefix(fields[0], []byte("#")) {
+				// log a warning for unknown commands
+				log.Printf("WARNING: Unknown command: %s", fields[0])
+			}
 			continue
 		}
 

scripts/build_darwin.sh

@@ -2,9 +2,15 @@
 
 mkdir -p dist
 
+GO_LDFLAGS="-X github.com/jmorganca/ollama/version.Version=$VERSION"
+GO_LDFLAGS="$GO_LDFLAGS -X github.com/jmorganca/ollama/server.mode=release"
+
 # build universal binary
-CGO_ENABLED=1 GOARCH=arm64 go build -o dist/ollama-darwin-arm64
-CGO_ENABLED=1 GOARCH=amd64 go build -o dist/ollama-darwin-amd64
+GOARCH=arm64 go generate ./...
+GOARCH=arm64 go build -ldflags "$GO_LDFLAGS" -o dist/ollama-darwin-arm64
+rm -rf llm/llama.cpp/ggml/build/*/bin
+GOARCH=amd64 go generate ./...
+GOARCH=amd64 go build -ldflags "$GO_LDFLAGS" -o dist/ollama-darwin-amd64
 lipo -create -output dist/ollama dist/ollama-darwin-arm64 dist/ollama-darwin-amd64
 rm dist/ollama-darwin-amd64 dist/ollama-darwin-arm64
 codesign --deep --force --options=runtime --sign "$APPLE_IDENTITY" --timestamp dist/ollama

server/auth.go

@@ -2,6 +2,7 @@ package server
 
 import (
 	"bytes"
+	"context"
 	"crypto/rand"
 	"crypto/sha256"
 	"encoding/base64"
@@ -9,11 +10,12 @@ import (
 	"encoding/json"
 	"fmt"
 	"io"
-	"io/ioutil"
 	"log"
 	"net/http"
+	"net/url"
 	"os"
 	"path"
+	"strconv"
 	"strings"
 	"time"
 
@@ -43,16 +45,34 @@ func generateNonce(length int) (string, error) {
 	return base64.RawURLEncoding.EncodeToString(nonce), nil
 }
 
-func (r AuthRedirect) URL() (string, error) {
+func (r AuthRedirect) URL() (*url.URL, error) {
+	redirectURL, err := url.Parse(r.Realm)
+	if err != nil {
+		return nil, err
+	}
+
+	values := redirectURL.Query()
+
+	values.Add("service", r.Service)
+
+	for _, s := range strings.Split(r.Scope, " ") {
+		values.Add("scope", s)
+	}
+
+	values.Add("ts", strconv.FormatInt(time.Now().Unix(), 10))
+
 	nonce, err := generateNonce(16)
 	if err != nil {
-		return "", err
+		return nil, err
 	}
-	return fmt.Sprintf("%s?service=%s&scope=%s&ts=%d&nonce=%s", r.Realm, r.Service, r.Scope, time.Now().Unix(), nonce), nil
+	values.Add("nonce", nonce)
+
+	redirectURL.RawQuery = values.Encode()
+	return redirectURL, nil
 }
 
-func getAuthToken(redirData AuthRedirect, regOpts *RegistryOptions) (string, error) {
-	url, err := redirData.URL()
+func getAuthToken(ctx context.Context, redirData AuthRedirect, regOpts *RegistryOptions) (string, error) {
+	redirectURL, err := redirData.URL()
 	if err != nil {
 		return "", err
 	}
@@ -62,9 +82,9 @@ func getAuthToken(redirData AuthRedirect, regOpts *RegistryOptions) (string, err
 		return "", err
 	}
 
-	keyPath := path.Join(home, ".ollama/id_ed25519")
+	keyPath := path.Join(home, ".ollama", "id_ed25519")
 
-	rawKey, err := ioutil.ReadFile(keyPath)
+	rawKey, err := os.ReadFile(keyPath)
 	if err != nil {
 		log.Printf("Failed to load private key: %v", err)
 		return "", err
@@ -72,34 +92,24 @@ func getAuthToken(redirData AuthRedirect, regOpts *RegistryOptions) (string, err
 
 	s := SignatureData{
 		Method: "GET",
-		Path:   url,
+		Path:   redirectURL.String(),
 		Data:   nil,
 	}
 
-	if !strings.HasPrefix(s.Path, "http") {
-		if regOpts.Insecure {
-			s.Path = "http://" + url
-		} else {
-			s.Path = "https://" + url
-		}
-	}
-
 	sig, err := s.Sign(rawKey)
 	if err != nil {
 		return "", err
 	}
 
-	headers := map[string]string{
-		"Authorization": sig,
-	}
-
-	resp, err := makeRequest("GET", url, headers, nil, regOpts)
+	headers := make(http.Header)
+	headers.Set("Authorization", sig)
+	resp, err := makeRequest(ctx, "GET", redirectURL, headers, nil, regOpts)
 	if err != nil {
 		log.Printf("couldn't get token: %q", err)
 	}
 	defer resp.Body.Close()
 
-	if resp.StatusCode != http.StatusOK {
+	if resp.StatusCode >= http.StatusBadRequest {
 		body, _ := io.ReadAll(resp.Body)
 		return "", fmt.Errorf("on pull registry responded with code %d: %s", resp.StatusCode, body)
 	}

server/download.go

@@ -25,17 +25,27 @@ type FileDownload struct {
 
 var inProgress sync.Map // map of digests currently being downloaded to their current download progress
 
+type downloadOpts struct {
+	mp      ModelPath
+	digest  string
+	regOpts *RegistryOptions
+	fn      func(api.ProgressResponse)
+	retry   int // track the number of retries on this download
+}
+
+const maxRetry = 3
+
 // downloadBlob downloads a blob from the registry and stores it in the blobs directory
-func downloadBlob(ctx context.Context, mp ModelPath, digest string, regOpts *RegistryOptions, fn func(api.ProgressResponse)) error {
-	fp, err := GetBlobsPath(digest)
+func downloadBlob(ctx context.Context, opts downloadOpts) error {
+	fp, err := GetBlobsPath(opts.digest)
 	if err != nil {
 		return err
 	}
 
 	if fi, _ := os.Stat(fp); fi != nil {
 		// we already have the file, so return
-		fn(api.ProgressResponse{
-			Digest:    digest,
+		opts.fn(api.ProgressResponse{
+			Digest:    opts.digest,
 			Total:     int(fi.Size()),
 			Completed: int(fi.Size()),
 		})
@@ -44,24 +54,33 @@ func downloadBlob(ctx context.Context, mp ModelPath, digest string, regOpts *Reg
 	}
 
 	fileDownload := &FileDownload{
-		Digest:    digest,
+		Digest:    opts.digest,
 		FilePath:  fp,
 		Total:     1, // dummy value to indicate that we don't know the total size yet
 		Completed: 0,
 	}
 
-	_, downloading := inProgress.LoadOrStore(digest, fileDownload)
+	_, downloading := inProgress.LoadOrStore(opts.digest, fileDownload)
 	if downloading {
 		// this is another client requesting the server to download the same blob concurrently
-		return monitorDownload(ctx, mp, regOpts, fileDownload, fn)
+		return monitorDownload(ctx, opts, fileDownload)
 	}
-	return doDownload(ctx, mp, regOpts, fileDownload, fn)
+	if err := doDownload(ctx, opts, fileDownload); err != nil {
+		if errors.Is(err, errDownload) && opts.retry < maxRetry {
+			opts.retry++
+			log.Print(err)
+			log.Printf("retrying download of %s", opts.digest)
+			return downloadBlob(ctx, opts)
+		}
+		return err
+	}
+	return nil
 }
 
 var downloadMu sync.Mutex // mutex to check to resume a download while monitoring
 
 // monitorDownload monitors the download progress of a blob and resumes it if it is interrupted
-func monitorDownload(ctx context.Context, mp ModelPath, regOpts *RegistryOptions, f *FileDownload, fn func(api.ProgressResponse)) error {
+func monitorDownload(ctx context.Context, opts downloadOpts, f *FileDownload) error {
 	tick := time.NewTicker(time.Second)
 	for range tick.C {
 		done, resume, err := func() (bool, bool, error) {
@@ -72,7 +91,7 @@ func monitorDownload(ctx context.Context, mp ModelPath, regOpts *RegistryOptions
 				// check once again if the download is complete
 				if fi, _ := os.Stat(f.FilePath); fi != nil {
 					// successful download while monitoring
-					fn(api.ProgressResponse{
+					opts.fn(api.ProgressResponse{
 						Digest:    f.Digest,
 						Total:     int(fi.Size()),
 						Completed: int(fi.Size()),
@@ -87,7 +106,7 @@ func monitorDownload(ctx context.Context, mp ModelPath, regOpts *RegistryOptions
 			if !ok {
 				return false, false, fmt.Errorf("invalid type for in progress download: %T", val)
 			}
-			fn(api.ProgressResponse{
+			opts.fn(api.ProgressResponse{
 				Status:    fmt.Sprintf("downloading %s", f.Digest),
 				Digest:    f.Digest,
 				Total:     int(f.Total),
@@ -103,16 +122,20 @@ func monitorDownload(ctx context.Context, mp ModelPath, regOpts *RegistryOptions
 			return nil
 		}
 		if resume {
-			return doDownload(ctx, mp, regOpts, f, fn)
+			return doDownload(ctx, opts, f)
 		}
 	}
 	return nil
 }
 
-var chunkSize = 1024 * 1024 // 1 MiB in bytes
+var (
+	chunkSize   = 1024 * 1024 // 1 MiB in bytes
+	errDownload = fmt.Errorf("download failed")
+)
 
 // doDownload downloads a blob from the registry and stores it in the blobs directory
-func doDownload(ctx context.Context, mp ModelPath, regOpts *RegistryOptions, f *FileDownload, fn func(api.ProgressResponse)) error {
+func doDownload(ctx context.Context, opts downloadOpts, f *FileDownload) error {
+	defer inProgress.Delete(f.Digest)
 	var size int64
 
 	fi, err := os.Stat(f.FilePath + "-partial")
@@ -132,21 +155,22 @@ func doDownload(ctx context.Context, mp ModelPath, regOpts *RegistryOptions, f *
 		}
 	}
 
-	url := fmt.Sprintf("%s/v2/%s/blobs/%s", mp.Registry, mp.GetNamespaceRepository(), f.Digest)
-	headers := map[string]string{
-		"Range": fmt.Sprintf("bytes=%d-", size),
-	}
+	requestURL := opts.mp.BaseURL()
+	requestURL = requestURL.JoinPath("v2", opts.mp.GetNamespaceRepository(), "blobs", f.Digest)
 
-	resp, err := makeRequest("GET", url, headers, nil, regOpts)
+	headers := make(http.Header)
+	headers.Set("Range", fmt.Sprintf("bytes=%d-", size))
+
+	resp, err := makeRequest(ctx, "GET", requestURL, headers, nil, opts.regOpts)
 	if err != nil {
 		log.Printf("couldn't download blob: %v", err)
-		return err
+		return fmt.Errorf("%w: %w", errDownload, err)
 	}
 	defer resp.Body.Close()
 
-	if resp.StatusCode != http.StatusOK && resp.StatusCode != http.StatusPartialContent {
+	if resp.StatusCode >= http.StatusBadRequest {
 		body, _ := io.ReadAll(resp.Body)
-		return fmt.Errorf("on download registry responded with code %d: %v", resp.StatusCode, string(body))
+		return fmt.Errorf("%w: on download registry responded with code %d: %v", errDownload, resp.StatusCode, string(body))
 	}
 
 	err = os.MkdirAll(path.Dir(f.FilePath), 0o700)
@@ -173,7 +197,7 @@ outerLoop:
 			inProgress.Delete(f.Digest)
 			return nil
 		default:
-			fn(api.ProgressResponse{
+			opts.fn(api.ProgressResponse{
 				Status:    fmt.Sprintf("downloading %s", f.Digest),
 				Digest:    f.Digest,
 				Total:     int(f.Total),
@@ -186,7 +210,7 @@ outerLoop:
 				}
 
 				if err := os.Rename(f.FilePath+"-partial", f.FilePath); err != nil {
-					fn(api.ProgressResponse{
+					opts.fn(api.ProgressResponse{
 						Status:    fmt.Sprintf("error renaming file: %v", err),
 						Digest:    f.Digest,
 						Total:     int(f.Total),
@@ -201,15 +225,13 @@ outerLoop:
 
 		n, err := io.CopyN(out, resp.Body, int64(chunkSize))
 		if err != nil && !errors.Is(err, io.EOF) {
-			return err
+			return fmt.Errorf("%w: %w", errDownload, err)
 		}
 		f.Completed += n
 
 		inProgress.Store(f.Digest, f)
 	}
 
-	inProgress.Delete(f.Digest)
-
 	log.Printf("success getting %s\n", f.Digest)
 	return nil
 }

server/images_test.go

@@ -0,0 +1,23 @@
+package server
+
+import (
+	"testing"
+
+	"github.com/jmorganca/ollama/api"
+)
+
+func TestModelPrompt(t *testing.T) {
+	var m Model
+	req := api.GenerateRequest{
+		Template: "a{{ .Prompt }}b",
+		Prompt:   "<h1>",
+	}
+	s, err := m.Prompt(req, "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	want := "a<h1>b"
+	if s != want {
+		t.Errorf("got %q, want %q", s, want)
+	}
+}

server/modelpath.go

@@ -1,7 +1,9 @@
 package server
 
 import (
+	"errors"
 	"fmt"
+	"net/url"
 	"os"
 	"path/filepath"
 	"runtime"
@@ -23,42 +25,46 @@ const (
 	DefaultProtocolScheme = "https"
 )
 
+var (
+	ErrInvalidImageFormat = errors.New("invalid image format")
+	ErrInvalidProtocol    = errors.New("invalid protocol scheme")
+	ErrInsecureProtocol   = errors.New("insecure protocol http")
+)
+
 func ParseModelPath(name string) ModelPath {
-	slashParts := strings.Split(name, "/")
-	var registry, namespace, repository, tag string
-
-	switch len(slashParts) {
-	case 3:
-		registry = slashParts[0]
-		namespace = slashParts[1]
-		repository = strings.Split(slashParts[2], ":")[0]
-	case 2:
-		registry = DefaultRegistry
-		namespace = slashParts[0]
-		repository = strings.Split(slashParts[1], ":")[0]
-	case 1:
-		registry = DefaultRegistry
-		namespace = DefaultNamespace
-		repository = strings.Split(slashParts[0], ":")[0]
-	default:
-		fmt.Println("Invalid image format.")
-		return ModelPath{}
-	}
-
-	colonParts := strings.Split(slashParts[len(slashParts)-1], ":")
-	if len(colonParts) == 2 {
-		tag = colonParts[1]
-	} else {
-		tag = DefaultTag
-	}
-
-	return ModelPath{
+	mp := ModelPath{
 		ProtocolScheme: DefaultProtocolScheme,
-		Registry:       registry,
-		Namespace:      namespace,
-		Repository:     repository,
-		Tag:            tag,
+		Registry:       DefaultRegistry,
+		Namespace:      DefaultNamespace,
+		Repository:     "",
+		Tag:            DefaultTag,
 	}
+
+	before, after, found := strings.Cut(name, "://")
+	if found {
+		mp.ProtocolScheme = before
+		name = after
+	}
+
+	parts := strings.Split(name, string(os.PathSeparator))
+	switch len(parts) {
+	case 3:
+		mp.Registry = parts[0]
+		mp.Namespace = parts[1]
+		mp.Repository = parts[2]
+	case 2:
+		mp.Namespace = parts[0]
+		mp.Repository = parts[1]
+	case 1:
+		mp.Repository = parts[0]
+	}
+
+	if repo, tag, found := strings.Cut(mp.Repository, ":"); found {
+		mp.Repository = repo
+		mp.Tag = tag
+	}
+
+	return mp
 }
 
 func (mp ModelPath) GetNamespaceRepository() string {
@@ -95,13 +101,25 @@ func (mp ModelPath) GetManifestPath(createDir bool) (string, error) {
 	return path, nil
 }
 
+func (mp ModelPath) BaseURL() *url.URL {
+	return &url.URL{
+		Scheme: mp.ProtocolScheme,
+		Host:   mp.Registry,
+	}
+}
+
 func GetManifestPath() (string, error) {
 	home, err := os.UserHomeDir()
 	if err != nil {
 		return "", err
 	}
 
-	return filepath.Join(home, ".ollama", "models", "manifests"), nil
+	path := filepath.Join(home, ".ollama", "models", "manifests")
+	if err := os.MkdirAll(path, 0o755); err != nil {
+		return "", err
+	}
+
+	return path, nil
 }
 
 func GetBlobsPath(digest string) (string, error) {

server/modelpath_test.go

@@ -0,0 +1,88 @@
+package server
+
+import "testing"
+
+func TestParseModelPath(t *testing.T) {
+	tests := []struct {
+		name    string
+		arg    string
+		want    ModelPath
+	}{
+		{
+			"full path https",
+			"https://example.com/ns/repo:tag",
+			ModelPath{
+				ProtocolScheme: "https",
+				Registry:       "example.com",
+				Namespace:      "ns",
+				Repository:     "repo",
+				Tag:            "tag",
+			},
+		},
+		{
+			"full path http",
+			"http://example.com/ns/repo:tag",
+			ModelPath{
+				ProtocolScheme: "http",
+				Registry:       "example.com",
+				Namespace:      "ns",
+				Repository:     "repo",
+				Tag:            "tag",
+			},
+		},
+		{
+			"no protocol",
+			"example.com/ns/repo:tag",
+			ModelPath{
+				ProtocolScheme: "https",
+				Registry:       "example.com",
+				Namespace:      "ns",
+				Repository:     "repo",
+				Tag:            "tag",
+			},
+		},
+		{
+			"no registry",
+			"ns/repo:tag",
+			ModelPath{
+				ProtocolScheme: "https",
+				Registry:       DefaultRegistry,
+				Namespace:      "ns",
+				Repository:     "repo",
+				Tag:            "tag",
+			},
+		},
+		{
+			"no namespace",
+			"repo:tag",
+			ModelPath{
+				ProtocolScheme: "https",
+				Registry:       DefaultRegistry,
+				Namespace:      DefaultNamespace,
+				Repository:     "repo",
+				Tag:            "tag",
+			},
+		},
+		{
+			"no tag",
+			"repo",
+			ModelPath{
+				ProtocolScheme: "https",
+				Registry:       DefaultRegistry,
+				Namespace:      DefaultNamespace,
+				Repository:     "repo",
+				Tag:            DefaultTag,
+			},
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			got := ParseModelPath(tc.arg)
+
+			if got != tc.want {
+				t.Errorf("got: %q want: %q", got, tc.want)
+			}
+		})
+	}
+}

server/routes.go

@@ -3,17 +3,19 @@ package server
 import (
 	"context"
 	"encoding/json"
-	"errors"
 	"fmt"
 	"io"
 	"log"
 	"net"
 	"net/http"
 	"os"
+	"os/signal"
 	"path/filepath"
 	"reflect"
+	"strconv"
 	"strings"
 	"sync"
+	"syscall"
 	"time"
 
 	"github.com/gin-contrib/cors"
@@ -21,14 +23,28 @@ import (
 	"gonum.org/v1/gonum/mat"
 
 	"github.com/jmorganca/ollama/api"
-	"github.com/jmorganca/ollama/llama"
+	"github.com/jmorganca/ollama/llm"
 	"github.com/jmorganca/ollama/vector"
 )
 
+var mode string = gin.DebugMode
+
+func init() {
+	switch mode {
+	case gin.DebugMode:
+	case gin.ReleaseMode:
+	case gin.TestMode:
+	default:
+		mode = gin.DebugMode
+	}
+
+	gin.SetMode(mode)
+}
+
 var loaded struct {
 	mu sync.Mutex
 
-	llm        *llama.LLM
+	llm        llm.LLM
 	Embeddings []vector.Embedding
 
 	expireAt    time.Time
@@ -38,8 +54,10 @@ var loaded struct {
 	options api.Options
 }
 
+var defaultSessionDuration = 5 * time.Minute
+
 // load a model into memory if it is not already loaded, it is up to the caller to lock loaded.mu before calling this function
-func load(model *Model, reqOpts map[string]interface{}, sessionDuration time.Duration) error {
+func load(ctx context.Context, model *Model, reqOpts map[string]interface{}, sessionDuration time.Duration) error {
 	opts := api.DefaultOptions()
 	if err := opts.FromMap(model.Options); err != nil {
 		log.Printf("could not load model options: %v", err)
@@ -51,8 +69,20 @@ func load(model *Model, reqOpts map[string]interface{}, sessionDuration time.Dur
 		return err
 	}
 
+	// 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 {
+			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 = ""
+		}
+	}
+
 	if model.Digest != loaded.digest || !reflect.DeepEqual(loaded.options, opts) {
 		if loaded.llm != nil {
+			log.Println("changing loaded model")
 			loaded.llm.Close()
 			loaded.llm = nil
 			loaded.digest = ""
@@ -63,11 +93,16 @@ func load(model *Model, reqOpts map[string]interface{}, sessionDuration time.Dur
 			loaded.Embeddings = model.Embeddings
 		}
 
-		llm, err := llama.New(model.ModelPath, opts)
+		llmModel, err := llm.New(model.ModelPath, model.AdapterPaths, opts)
 		if err != nil {
 			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 {
@@ -79,15 +114,20 @@ func load(model *Model, reqOpts map[string]interface{}, sessionDuration time.Dur
 				return err
 			}
 
-			tokensWithSystem := llm.Encode(promptWithSystem)
-			tokensNoSystem := llm.Encode(promptNoSystem)
+			tokensWithSystem, err := llmModel.Encode(ctx, promptWithSystem)
+			if err != nil {
+				return err
+			}
 
-			llm.NumKeep = len(tokensWithSystem) - len(tokensNoSystem) + 1
+			tokensNoSystem, err := llmModel.Encode(ctx, promptNoSystem)
+			if err != nil {
+				return err
+			}
+
+			opts.NumKeep = len(tokensWithSystem) - len(tokensNoSystem)
+
+			llmModel.SetOptions(opts)
 		}
-
-		loaded.llm = llm
-		loaded.digest = model.Digest
-		loaded.options = opts
 	}
 	loaded.expireAt = time.Now().Add(sessionDuration)
 
@@ -131,8 +171,8 @@ func GenerateHandler(c *gin.Context) {
 		return
 	}
 
-	sessionDuration := 5 * time.Minute
-	if err := load(model, req.Options, sessionDuration); err != nil {
+	sessionDuration := defaultSessionDuration // TODO: set this duration from the request if specified
+	if err := load(c.Request.Context(), model, req.Options, sessionDuration); err != nil {
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 		return
 	}
@@ -141,7 +181,7 @@ func GenerateHandler(c *gin.Context) {
 
 	embedding := ""
 	if model.Embeddings != nil && len(model.Embeddings) > 0 {
-		promptEmbed, err := loaded.llm.Embedding(req.Prompt)
+		promptEmbed, err := loaded.llm.Embedding(c.Request.Context(), req.Prompt)
 		if err != nil {
 			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 			return
@@ -177,7 +217,7 @@ func GenerateHandler(c *gin.Context) {
 			ch <- r
 		}
 
-		if err := loaded.llm.Predict(req.Context, prompt, fn); err != nil {
+		if err := loaded.llm.Predict(c.Request.Context(), req.Context, prompt, fn); err != nil {
 			ch <- gin.H{"error": err.Error()}
 		}
 	}()
@@ -200,7 +240,7 @@ func EmbeddingHandler(c *gin.Context) {
 		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
 		return
 	}
-	if err := load(model, req.Options, 5*time.Minute); err != nil {
+	if err := load(c.Request.Context(), model, req.Options, 5*time.Minute); err != nil {
 		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
 		return
 	}
@@ -210,7 +250,7 @@ func EmbeddingHandler(c *gin.Context) {
 		return
 	}
 
-	embedding, err := loaded.llm.Embedding(req.Prompt)
+	embedding, err := loaded.llm.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"})
@@ -274,7 +314,8 @@ func PushModelHandler(c *gin.Context) {
 			Password: req.Password,
 		}
 
-		if err := PushModel(req.Name, regOpts, fn); err != nil {
+		ctx := context.Background()
+		if err := PushModel(ctx, req.Name, regOpts, fn); err != nil {
 			ch <- gin.H{"error": err.Error()}
 		}
 	}()
@@ -324,49 +365,110 @@ 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()})
+		return
+	}
+
+	resp, err := GetModelInfo(req.Name)
+	if err != nil {
+		if os.IsNotExist(err) {
+			c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found", req.Name)})
+		} else {
+			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		}
+		return
+	}
+
+	c.JSON(http.StatusOK, resp)
+}
+
+func GetModelInfo(name string) (*api.ShowResponse, error) {
+	model, err := GetModel(name)
+	if err != nil {
+		return nil, err
+	}
+
+	resp := &api.ShowResponse{
+		License:  strings.Join(model.License, "\n"),
+		System:   model.System,
+		Template: model.Template,
+	}
+
+	mf, err := ShowModelfile(model)
+	if err != nil {
+		return nil, err
+	}
+
+	resp.Modelfile = mf
+
+	var params []string
+	cs := 30
+	for k, v := range model.Options {
+		switch val := v.(type) {
+		case string:
+			params = append(params, fmt.Sprintf("%-*s %s", cs, k, val))
+		case int:
+			params = append(params, fmt.Sprintf("%-*s %s", cs, k, strconv.Itoa(val)))
+		case float64:
+			params = append(params, fmt.Sprintf("%-*s %s", cs, k, strconv.FormatFloat(val, 'f', 0, 64)))
+		case bool:
+			params = append(params, fmt.Sprintf("%-*s %s", cs, k, strconv.FormatBool(val)))
+		case []interface{}:
+			for _, nv := range val {
+				switch nval := nv.(type) {
+				case string:
+					params = append(params, fmt.Sprintf("%-*s %s", cs, k, nval))
+				case int:
+					params = append(params, fmt.Sprintf("%-*s %s", cs, k, strconv.Itoa(nval)))
+				case float64:
+					params = append(params, fmt.Sprintf("%-*s %s", cs, k, strconv.FormatFloat(nval, 'f', 0, 64)))
+				case bool:
+					params = append(params, fmt.Sprintf("%-*s %s", cs, k, strconv.FormatBool(nval)))
+				}
+			}
+		}
+	}
+	resp.Parameters = strings.Join(params, "\n")
+
+	return resp, nil
+}
+
 func ListModelsHandler(c *gin.Context) {
-	var models []api.ListResponseModel
+	var models []api.ModelResponse
 	fp, err := GetManifestPath()
 	if err != nil {
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 		return
 	}
-	err = filepath.Walk(fp, func(path string, info os.FileInfo, err error) error {
-		if err != nil {
-			if errors.Is(err, os.ErrNotExist) {
-				log.Printf("manifest file does not exist: %s", fp)
-				return nil
-			}
-			return err
-		}
+
+	walkFunc := func(path string, info os.FileInfo, _ error) error {
 		if !info.IsDir() {
-			fi, err := os.Stat(path)
-			if err != nil {
-				log.Printf("skipping file: %s", fp)
-				return nil
-			}
-			path := path[len(fp)+1:]
-			slashIndex := strings.LastIndex(path, "/")
-			if slashIndex == -1 {
-				return nil
-			}
-			tag := path[:slashIndex] + ":" + path[slashIndex+1:]
+			dir, file := filepath.Split(path)
+			dir = strings.Trim(strings.TrimPrefix(dir, fp), string(os.PathSeparator))
+			tag := strings.Join([]string{dir, file}, ":")
+
 			mp := ParseModelPath(tag)
-			manifest, err := GetManifest(mp)
+			manifest, digest, err := GetManifest(mp)
 			if err != nil {
 				log.Printf("skipping file: %s", fp)
 				return nil
 			}
-			model := api.ListResponseModel{
+
+			models = append(models, api.ModelResponse{
 				Name:       mp.GetShortTagname(),
 				Size:       manifest.GetTotalSize(),
-				ModifiedAt: fi.ModTime(),
-			}
-			models = append(models, model)
+				Digest:     digest,
+				ModifiedAt: info.ModTime(),
+			})
 		}
+
 		return nil
-	})
-	if err != nil {
+	}
+
+	if err := filepath.Walk(fp, walkFunc); err != nil {
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 		return
 	}
@@ -427,12 +529,24 @@ func Serve(ln net.Listener, origins []string) error {
 	r.POST("/api/copy", CopyModelHandler)
 	r.GET("/api/tags", ListModelsHandler)
 	r.DELETE("/api/delete", DeleteModelHandler)
+	r.POST("/api/show", ShowModelHandler)
 
 	log.Printf("Listening on %s", ln.Addr())
 	s := &http.Server{
 		Handler: r,
 	}
 
+	// listen for a ctrl+c and stop any loaded llm
+	signals := make(chan os.Signal, 1)
+	signal.Notify(signals, syscall.SIGINT)
+	go func() {
+		<-signals
+		if loaded.llm != nil {
+			loaded.llm.Close()
+		}
+		os.Exit(0)
+	}()
+
 	return s.Serve(ln)
 }
 

server/upload.go

@@ -0,0 +1,165 @@
+package server
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"io"
+	"log"
+	"net/http"
+	"net/url"
+	"os"
+	"strconv"
+
+	"github.com/jmorganca/ollama/api"
+)
+
+func startUpload(ctx context.Context, mp ModelPath, layer *Layer, regOpts *RegistryOptions) (*url.URL, error) {
+	requestURL := mp.BaseURL()
+	requestURL = requestURL.JoinPath("v2", mp.GetNamespaceRepository(), "blobs/uploads/")
+	if layer.From != "" {
+		values := requestURL.Query()
+		values.Add("mount", layer.Digest)
+		values.Add("from", layer.From)
+		requestURL.RawQuery = values.Encode()
+	}
+
+	resp, err := makeRequestWithRetry(ctx, "POST", requestURL, nil, nil, regOpts)
+	if err != nil {
+		log.Printf("couldn't start upload: %v", err)
+		return nil, err
+	}
+	defer resp.Body.Close()
+
+	// Extract UUID location from header
+	location := resp.Header.Get("Location")
+	if location == "" {
+		return nil, fmt.Errorf("location header is missing in response")
+	}
+
+	return url.Parse(location)
+}
+
+func uploadBlobChunked(ctx context.Context, requestURL *url.URL, layer *Layer, regOpts *RegistryOptions, fn func(api.ProgressResponse)) error {
+	// TODO allow resumability
+	// TODO allow canceling uploads via DELETE
+
+	fp, err := GetBlobsPath(layer.Digest)
+	if err != nil {
+		return err
+	}
+
+	f, err := os.Open(fp)
+	if err != nil {
+		return err
+	}
+	defer f.Close()
+
+	// 95MB chunk size
+	chunkSize := 95 * 1024 * 1024
+
+	for offset := int64(0); offset < int64(layer.Size); {
+		chunk := int64(layer.Size) - offset
+		if chunk > int64(chunkSize) {
+			chunk = int64(chunkSize)
+		}
+
+		sectionReader := io.NewSectionReader(f, int64(offset), chunk)
+		for try := 0; try < MaxRetries; try++ {
+			r, w := io.Pipe()
+			defer r.Close()
+			go func() {
+				defer w.Close()
+
+				for chunked := int64(0); chunked < chunk; {
+					n, err := io.CopyN(w, sectionReader, 1024*1024)
+					if err != nil && !errors.Is(err, io.EOF) {
+						fn(api.ProgressResponse{
+							Status:    fmt.Sprintf("error reading chunk: %v", err),
+							Digest:    layer.Digest,
+							Total:     layer.Size,
+							Completed: int(offset),
+						})
+
+						return
+					}
+
+					chunked += n
+					fn(api.ProgressResponse{
+						Status:    fmt.Sprintf("uploading %s", layer.Digest),
+						Digest:    layer.Digest,
+						Total:     layer.Size,
+						Completed: int(offset) + int(chunked),
+					})
+				}
+			}()
+
+			headers := make(http.Header)
+			headers.Set("Content-Type", "application/octet-stream")
+			headers.Set("Content-Length", strconv.Itoa(int(chunk)))
+			headers.Set("Content-Range", fmt.Sprintf("%d-%d", offset, offset+sectionReader.Size()-1))
+			resp, err := makeRequest(ctx, "PATCH", requestURL, headers, r, regOpts)
+			if err != nil && !errors.Is(err, io.EOF) {
+				fn(api.ProgressResponse{
+					Status:    fmt.Sprintf("error uploading chunk: %v", err),
+					Digest:    layer.Digest,
+					Total:     layer.Size,
+					Completed: int(offset),
+				})
+
+				return err
+			}
+			defer resp.Body.Close()
+
+			switch {
+			case resp.StatusCode == http.StatusUnauthorized:
+				auth := resp.Header.Get("www-authenticate")
+				authRedir := ParseAuthRedirectString(auth)
+				token, err := getAuthToken(ctx, authRedir, regOpts)
+				if err != nil {
+					return err
+				}
+
+				regOpts.Token = token
+				if _, err := sectionReader.Seek(0, io.SeekStart); err != nil {
+					return err
+				}
+
+				continue
+			case resp.StatusCode >= http.StatusBadRequest:
+				body, _ := io.ReadAll(resp.Body)
+				return fmt.Errorf("on upload registry responded with code %d: %s", resp.StatusCode, body)
+			}
+
+			offset += sectionReader.Size()
+			requestURL, err = url.Parse(resp.Header.Get("Location"))
+			if err != nil {
+				return err
+			}
+
+			break
+		}
+	}
+
+	values := requestURL.Query()
+	values.Add("digest", layer.Digest)
+	requestURL.RawQuery = values.Encode()
+
+	headers := make(http.Header)
+	headers.Set("Content-Type", "application/octet-stream")
+	headers.Set("Content-Length", "0")
+
+	// finish the upload
+	resp, err := makeRequest(ctx, "PUT", requestURL, headers, nil, regOpts)
+	if err != nil {
+		log.Printf("couldn't finish upload: %v", err)
+		return err
+	}
+	defer resp.Body.Close()
+
+	if resp.StatusCode >= http.StatusBadRequest {
+		body, _ := io.ReadAll(resp.Body)
+		return fmt.Errorf("on finish upload registry responded with code %d: %v", resp.StatusCode, string(body))
+	}
+	return nil
+}

version/version.go

@@ -0,0 +1,3 @@
+package version
+
+var Version string = "0.0.0"