restore model load duration on generate response (#1524)

* restore model load duration on generate response

- set model load duration on generate and chat done response
- calculate createAt time when response created

* remove checkpoints predict opts

* Update routes.go

.dockerignore

@@ -6,3 +6,4 @@ scripts
 llm/llama.cpp/ggml
 llm/llama.cpp/gguf
 .env
+.cache

.gitignore

@@ -6,3 +6,6 @@
 dist
 ollama
 ggml-metal.metal
+.cache
+*.exe
+.idea

Dockerfile

@@ -19,5 +19,11 @@ RUN apt-get update && apt-get install -y ca-certificates
 COPY --from=0 /go/src/github.com/jmorganca/ollama/ollama /bin/ollama
 EXPOSE 11434
 ENV OLLAMA_HOST 0.0.0.0
+
+# set some environment variable for better NVIDIA compatibility
+ENV PATH=/usr/local/nvidia/bin:/usr/local/cuda/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
+ENV LD_LIBRARY_PATH=/usr/local/nvidia/lib:/usr/local/nvidia/lib64
+ENV NVIDIA_DRIVER_CAPABILITIES=compute,utility
+
 ENTRYPOINT ["/bin/ollama"]
 CMD ["serve"]

README.md

@@ -47,6 +47,8 @@ Here are some example open-source models that can be downloaded:
 
 | Model              | Parameters | Size  | Download                       |
 | ------------------ | ---------- | ----- | ------------------------------ |
+| Neural Chat        | 7B         | 4.1GB | `ollama run neural-chat`       |
+| Starling           | 7B         | 4.1GB | `ollama run starling-lm`       |
 | Mistral            | 7B         | 4.1GB | `ollama run mistral`           |
 | Llama 2            | 7B         | 3.8GB | `ollama run llama2`            |
 | Code Llama         | 7B         | 3.8GB | `ollama run codellama`         |
@@ -55,6 +57,7 @@ Here are some example open-source models that can be downloaded:
 | Llama 2 70B        | 70B        | 39GB  | `ollama run llama2:70b`        |
 | Orca Mini          | 3B         | 1.9GB | `ollama run orca-mini`         |
 | Vicuna             | 7B         | 3.8GB | `ollama run vicuna`            |
+| LLaVA              | 7B         | 4.5GB | `ollama run llava`             |
 
 > 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.
 
@@ -102,7 +105,7 @@ FROM llama2
 # set the temperature to 1 [higher is more creative, lower is more coherent]
 PARAMETER temperature 1
 
-# set the system prompt
+# set the system message
 SYSTEM """
 You are Mario from Super Mario Bros. Answer as Mario, the assistant, only.
 """
@@ -156,6 +159,13 @@ For multiline input, you can wrap text with `"""`:
 I'm a basic program that prints the famous "Hello, world!" message to the console.
 ```
 
+### Multimodal models
+
+```
+>>> What's in this image? /Users/jmorgan/Desktop/smile.png
+The image features a yellow smiley face, which is likely the central focus of the picture.
+```
+
 ### Pass in prompt as arguments
 
 ```
@@ -203,15 +213,27 @@ Finally, in a separate shell, run a model:
 ## REST API
 
 Ollama has a REST API for running and managing models.
-For example, to generate text from a model:
+
+### Generate a response
 
 ```
-curl -X POST http://localhost:11434/api/generate -d '{
+curl http://localhost:11434/api/generate -d '{
   "model": "llama2",
   "prompt":"Why is the sky blue?"
 }'
 ```
 
+### Chat with a model
+
+```
+curl http://localhost:11434/api/chat -d '{
+  "model": "mistral",
+  "messages": [
+    { "role": "user", "content": "why is the sky blue?" }
+  ]
+}'
+```
+
 See the [API documentation](./docs/api.md) for all endpoints.
 
 ## Community Integrations
@@ -225,6 +247,9 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Web UI](https://github.com/ollama-webui/ollama-webui)
 - [Ollamac](https://github.com/kevinhermawan/Ollamac)
 - [big-AGI](https://github.com/enricoros/big-agi/blob/main/docs/config-ollama.md)
+- [Cheshire Cat assistant framework](https://github.com/cheshire-cat-ai/core)
+- [Amica](https://github.com/semperai/amica)
+- [chatd](https://github.com/BruceMacD/chatd)
 
 ### Terminal
 
@@ -233,11 +258,22 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Emacs client](https://github.com/zweifisch/ollama)
 - [gen.nvim](https://github.com/David-Kunz/gen.nvim)
 - [ollama.nvim](https://github.com/nomnivore/ollama.nvim)
+- [ogpt.nvim](https://github.com/huynle/ogpt.nvim)
 - [gptel Emacs client](https://github.com/karthink/gptel)
+- [Oatmeal](https://github.com/dustinblackman/oatmeal)
+
+### Database
+
+- [MindsDB](https://github.com/mindsdb/mindsdb/blob/staging/mindsdb/integrations/handlers/ollama_handler/README.md)
+
+### Package managers
+
+- [Pacman](https://archlinux.org/packages/extra/x86_64/ollama/)
 
 ### Libraries
 
 - [LangChain](https://python.langchain.com/docs/integrations/llms/ollama) and [LangChain.js](https://js.langchain.com/docs/modules/model_io/models/llms/integrations/ollama) with [example](https://js.langchain.com/docs/use_cases/question_answering/local_retrieval_qa)
+- [LangChainGo](https://github.com/tmc/langchaingo/) with [example](https://github.com/tmc/langchaingo/tree/main/examples/ollama-completion-example)
 - [LlamaIndex](https://gpt-index.readthedocs.io/en/stable/examples/llm/ollama.html)
 - [LiteLLM](https://github.com/BerriAI/litellm)
 - [OllamaSharp for .NET](https://github.com/awaescher/OllamaSharp)
@@ -246,6 +282,11 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [ModelFusion Typescript Library](https://modelfusion.dev/integration/model-provider/ollama)
 - [OllamaKit for Swift](https://github.com/kevinhermawan/OllamaKit)
 - [Ollama for Dart](https://github.com/breitburg/dart-ollama)
+- [Ollama for Laravel](https://github.com/cloudstudio/ollama-laravel)
+
+### Mobile
+
+- [Maid](https://github.com/danemadsen/Maid) (Mobile Artificial Intelligence Distribution)
 
 ### Extensions & Plugins
 
@@ -256,4 +297,8 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Logseq Ollama plugin](https://github.com/omagdy7/ollama-logseq)
 - [Dagger Chatbot](https://github.com/samalba/dagger-chatbot)
 - [Discord AI Bot](https://github.com/mekb-turtle/discord-ai-bot)
+- [Ollama Telegram Bot](https://github.com/ruecat/ollama-telegram)
 - [Hass Ollama Conversation](https://github.com/ej52/hass-ollama-conversation)
+- [Rivet plugin](https://github.com/abrenneke/rivet-plugin-ollama)
+- [Llama Coder](https://github.com/ex3ndr/llama-coder) (Copilot alternative using Ollama)
+- [Obsidian BMO Chatbot plugin](https://github.com/longy2k/obsidian-bmo-chatbot)

api/client.go

@@ -5,6 +5,7 @@ import (
 	"bytes"
 	"context"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"io"
 	"net"
@@ -95,11 +96,19 @@ func (c *Client) do(ctx context.Context, method, path string, reqData, respData
 	var reqBody io.Reader
 	var data []byte
 	var err error
-	if reqData != nil {
+
+	switch reqData := reqData.(type) {
+	case io.Reader:
+		// reqData is already an io.Reader
+		reqBody = reqData
+	case nil:
+		// noop
+	default:
 		data, err = json.Marshal(reqData)
 		if err != nil {
 			return err
 		}
+
 		reqBody = bytes.NewReader(data)
 	}
 
@@ -212,6 +221,19 @@ func (c *Client) Generate(ctx context.Context, req *GenerateRequest, fn Generate
 	})
 }
 
+type ChatResponseFunc func(ChatResponse) error
+
+func (c *Client) Chat(ctx context.Context, req *ChatRequest, fn ChatResponseFunc) error {
+	return c.stream(ctx, http.MethodPost, "/api/chat", req, func(bts []byte) error {
+		var resp ChatResponse
+		if err := json.Unmarshal(bts, &resp); err != nil {
+			return err
+		}
+
+		return fn(resp)
+	})
+}
+
 type PullProgressFunc func(ProgressResponse) error
 
 func (c *Client) Pull(ctx context.Context, req *PullRequest, fn PullProgressFunc) error {
@@ -287,3 +309,30 @@ func (c *Client) Heartbeat(ctx context.Context) error {
 	}
 	return nil
 }
+
+func (c *Client) CreateBlob(ctx context.Context, digest string, r io.Reader) error {
+	if err := c.do(ctx, http.MethodHead, fmt.Sprintf("/api/blobs/%s", digest), nil, nil); err != nil {
+		var statusError StatusError
+		if !errors.As(err, &statusError) || statusError.StatusCode != http.StatusNotFound {
+			return err
+		}
+
+		if err := c.do(ctx, http.MethodPost, fmt.Sprintf("/api/blobs/%s", digest), r, nil); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (c *Client) Version(ctx context.Context) (string, error) {
+	var version struct {
+		Version string `json:"version"`
+	}
+
+	if err := c.do(ctx, http.MethodGet, "/api/version", nil, &version); err != nil {
+		return "", err
+	}
+
+	return version.Version, nil
+}

api/client.py

@@ -1,6 +1,10 @@
 import os
 import json
 import requests
+import os
+import hashlib
+import json
+from pathlib import Path
 
 BASE_URL = os.environ.get('OLLAMA_HOST', 'http://localhost:11434')
 
@@ -57,30 +61,86 @@ def generate(model_name, prompt, system=None, template=None, format="", context=
     except requests.exceptions.RequestException as e:
         print(f"An error occurred: {e}")
         return None, None
+    
+
+# Create a blob file on the server if it doesn't exist.
+def create_blob(digest, file_path):
+    url = f"{BASE_URL}/api/blobs/{digest}"
+
+    # Check if the blob exists
+    response = requests.head(url)
+    if response.status_code != 404:
+        return  # Blob already exists, no need to upload
+    response.raise_for_status()
+
+    # Upload the blob
+    with open(file_path, 'rb') as file_data:
+        requests.post(url, data=file_data)
+
 
 # Create a model from a Modelfile. Use the callback function to override the default handler.
-def create(model_name, model_path, callback=None):
+def create(model_name, filename, callback=None):
     try:
+        file_path = Path(filename).expanduser().resolve()
+        processed_lines = []
+
+        # Read and process the modelfile
+        with open(file_path, 'r') as f:
+            for line in f:            
+                # Skip empty or whitespace-only lines
+                if not line.strip():
+                    continue
+            
+                command, args = line.split(maxsplit=1)
+
+                if command.upper() in ["FROM", "ADAPTER"]:
+                    path = Path(args.strip()).expanduser()
+
+                    # Check if path is relative and resolve it
+                    if not path.is_absolute():
+                        path = (file_path.parent / path)
+
+                    # Skip if file does not exist for "model", this is handled by the server
+                    if not path.exists():
+                        processed_lines.append(line)
+                        continue
+
+                    # Calculate SHA-256 hash
+                    with open(path, 'rb') as bin_file:
+                        hash = hashlib.sha256()
+                        hash.update(bin_file.read())
+                        blob = f"sha256:{hash.hexdigest()}"
+                
+                    # Add the file to the remote server
+                    create_blob(blob, path)
+
+                    # Replace path with digest in the line
+                    line = f"{command} @{blob}\n"
+
+                processed_lines.append(line)
+
+        # Combine processed lines back into a single string
+        modelfile_content = '\n'.join(processed_lines)
+
         url = f"{BASE_URL}/api/create"
-        payload = {"name": model_name, "path": model_path}
-        
+        payload = {"name": model_name, "modelfile": modelfile_content}
+
         # Making a POST request with the stream parameter set to True to handle streaming responses
         with requests.post(url, json=payload, stream=True) as response:
             response.raise_for_status()
-
             # Iterating over the response line by line and displaying the status
             for line in response.iter_lines():
                 if line:
-                    # Parsing each line (JSON chunk) and extracting the status
                     chunk = json.loads(line)
-
                     if callback:
                         callback(chunk)
                     else:
                         print(f"Status: {chunk.get('status')}")
-    except requests.exceptions.RequestException as e:
+
+    except Exception as e:
         print(f"An error occurred: {e}")
 
+
 # Pull a model from a the model registry. Cancelled pulls are resumed from where they left off, and multiple
 # calls to will share the same download progress. Use the callback function to override the default handler.
 def pull(model_name, insecure=False, callback=None):
@@ -222,5 +282,3 @@ def heartbeat():
     except requests.exceptions.RequestException as e:
         print(f"An error occurred: {e}")
         return "Ollama is not running"
-
-

api/types.go

@@ -6,6 +6,7 @@ import (
 	"math"
 	"os"
 	"reflect"
+	"strconv"
 	"strings"
 	"time"
 )
@@ -30,19 +31,56 @@ func (e StatusError) Error() string {
 	}
 }
 
+type ImageData []byte
+
 type GenerateRequest struct {
-	Model    string `json:"model"`
-	Prompt   string `json:"prompt"`
-	System   string `json:"system"`
-	Template string `json:"template"`
-	Context  []int  `json:"context,omitempty"`
-	Stream   *bool  `json:"stream,omitempty"`
-	Raw      bool   `json:"raw,omitempty"`
-	Format   string `json:"format"`
+	Model    string      `json:"model"`
+	Prompt   string      `json:"prompt"`
+	System   string      `json:"system"`
+	Template string      `json:"template"`
+	Context  []int       `json:"context,omitempty"`
+	Stream   *bool       `json:"stream,omitempty"`
+	Raw      bool        `json:"raw,omitempty"`
+	Format   string      `json:"format"`
+	Images   []ImageData `json:"images,omitempty"`
 
 	Options map[string]interface{} `json:"options"`
 }
 
+type ChatRequest struct {
+	Model    string    `json:"model"`
+	Messages []Message `json:"messages"`
+	Stream   *bool     `json:"stream,omitempty"`
+	Format   string    `json:"format"`
+
+	Options map[string]interface{} `json:"options"`
+}
+
+type Message struct {
+	Role    string      `json:"role"` // one of ["system", "user", "assistant"]
+	Content string      `json:"content"`
+	Images  []ImageData `json:"images, omitempty"`
+}
+
+type ChatResponse struct {
+	Model     string    `json:"model"`
+	CreatedAt time.Time `json:"created_at"`
+	Message   *Message  `json:"message,omitempty"`
+
+	Done bool `json:"done"`
+
+	Metrics
+}
+
+type Metrics struct {
+	TotalDuration      time.Duration `json:"total_duration,omitempty"`
+	LoadDuration       time.Duration `json:"load_duration,omitempty"`
+	PromptEvalCount    int           `json:"prompt_eval_count,omitempty"`
+	PromptEvalDuration time.Duration `json:"prompt_eval_duration,omitempty"`
+	EvalCount          int           `json:"eval_count,omitempty"`
+	EvalDuration       time.Duration `json:"eval_duration,omitempty"`
+}
+
 // Options specfied in GenerateRequest, if you add a new option here add it to the API docs also
 type Options struct {
 	Runner
@@ -99,9 +137,10 @@ type EmbeddingResponse struct {
 }
 
 type CreateRequest struct {
-	Name   string `json:"name"`
-	Path   string `json:"path"`
-	Stream *bool  `json:"stream,omitempty"`
+	Name      string `json:"name"`
+	Path      string `json:"path"`
+	Modelfile string `json:"modelfile"`
+	Stream    *bool  `json:"stream,omitempty"`
 }
 
 type DeleteRequest struct {
@@ -113,11 +152,12 @@ type ShowRequest struct {
 }
 
 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"`
+	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"`
+	Details    ModelDetails `json:"details,omitempty"`
 }
 
 type CopyRequest struct {
@@ -153,10 +193,11 @@ type ListResponse struct {
 }
 
 type ModelResponse struct {
-	Name       string    `json:"name"`
-	ModifiedAt time.Time `json:"modified_at"`
-	Size       int64     `json:"size"`
-	Digest     string    `json:"digest"`
+	Name       string       `json:"name"`
+	ModifiedAt time.Time    `json:"modified_at"`
+	Size       int64        `json:"size"`
+	Digest     string       `json:"digest"`
+	Details    ModelDetails `json:"details,omitempty"`
 }
 
 type TokenResponse struct {
@@ -171,39 +212,42 @@ type GenerateResponse struct {
 	Done    bool  `json:"done"`
 	Context []int `json:"context,omitempty"`
 
-	TotalDuration      time.Duration `json:"total_duration,omitempty"`
-	LoadDuration       time.Duration `json:"load_duration,omitempty"`
-	PromptEvalCount    int           `json:"prompt_eval_count,omitempty"`
-	PromptEvalDuration time.Duration `json:"prompt_eval_duration,omitempty"`
-	EvalCount          int           `json:"eval_count,omitempty"`
-	EvalDuration       time.Duration `json:"eval_duration,omitempty"`
+	Metrics
 }
 
-func (r *GenerateResponse) Summary() {
-	if r.TotalDuration > 0 {
-		fmt.Fprintf(os.Stderr, "total duration:       %v\n", r.TotalDuration)
+type ModelDetails struct {
+	Format            string   `json:"format"`
+	Family            string   `json:"family"`
+	Families          []string `json:"families"`
+	ParameterSize     string   `json:"parameter_size"`
+	QuantizationLevel string   `json:"quantization_level"`
+}
+
+func (m *Metrics) Summary() {
+	if m.TotalDuration > 0 {
+		fmt.Fprintf(os.Stderr, "total duration:       %v\n", m.TotalDuration)
 	}
 
-	if r.LoadDuration > 0 {
-		fmt.Fprintf(os.Stderr, "load duration:        %v\n", r.LoadDuration)
+	if m.LoadDuration > 0 {
+		fmt.Fprintf(os.Stderr, "load duration:        %v\n", m.LoadDuration)
 	}
 
-	if r.PromptEvalCount > 0 {
-		fmt.Fprintf(os.Stderr, "prompt eval count:    %d token(s)\n", r.PromptEvalCount)
+	if m.PromptEvalCount > 0 {
+		fmt.Fprintf(os.Stderr, "prompt eval count:    %d token(s)\n", m.PromptEvalCount)
 	}
 
-	if r.PromptEvalDuration > 0 {
-		fmt.Fprintf(os.Stderr, "prompt eval duration: %s\n", r.PromptEvalDuration)
-		fmt.Fprintf(os.Stderr, "prompt eval rate:     %.2f tokens/s\n", float64(r.PromptEvalCount)/r.PromptEvalDuration.Seconds())
+	if m.PromptEvalDuration > 0 {
+		fmt.Fprintf(os.Stderr, "prompt eval duration: %s\n", m.PromptEvalDuration)
+		fmt.Fprintf(os.Stderr, "prompt eval rate:     %.2f tokens/s\n", float64(m.PromptEvalCount)/m.PromptEvalDuration.Seconds())
 	}
 
-	if r.EvalCount > 0 {
-		fmt.Fprintf(os.Stderr, "eval count:           %d token(s)\n", r.EvalCount)
+	if m.EvalCount > 0 {
+		fmt.Fprintf(os.Stderr, "eval count:           %d token(s)\n", m.EvalCount)
 	}
 
-	if r.EvalDuration > 0 {
-		fmt.Fprintf(os.Stderr, "eval duration:        %s\n", r.EvalDuration)
-		fmt.Fprintf(os.Stderr, "eval rate:            %.2f tokens/s\n", float64(r.EvalCount)/r.EvalDuration.Seconds())
+	if m.EvalDuration > 0 {
+		fmt.Fprintf(os.Stderr, "eval duration:        %s\n", m.EvalDuration)
+		fmt.Fprintf(os.Stderr, "eval rate:            %.2f tokens/s\n", float64(m.EvalCount)/m.EvalDuration.Seconds())
 	}
 }
 
@@ -359,3 +403,63 @@ func (d *Duration) UnmarshalJSON(b []byte) (err error) {
 
 	return nil
 }
+
+// FormatParams converts specified parameter options to their correct types
+func FormatParams(params map[string][]string) (map[string]interface{}, error) {
+	opts := Options{}
+	valueOpts := reflect.ValueOf(&opts).Elem() // names of the fields in the options struct
+	typeOpts := reflect.TypeOf(opts)           // types of the fields in the options struct
+
+	// build map of json struct tags to their types
+	jsonOpts := make(map[string]reflect.StructField)
+	for _, field := range reflect.VisibleFields(typeOpts) {
+		jsonTag := strings.Split(field.Tag.Get("json"), ",")[0]
+		if jsonTag != "" {
+			jsonOpts[jsonTag] = field
+		}
+	}
+
+	out := make(map[string]interface{})
+	// iterate params and set values based on json struct tags
+	for key, vals := range params {
+		if opt, ok := jsonOpts[key]; !ok {
+			return nil, fmt.Errorf("unknown parameter '%s'", key)
+		} else {
+			field := valueOpts.FieldByName(opt.Name)
+			if field.IsValid() && field.CanSet() {
+				switch field.Kind() {
+				case reflect.Float32:
+					floatVal, err := strconv.ParseFloat(vals[0], 32)
+					if err != nil {
+						return nil, fmt.Errorf("invalid float value %s", vals)
+					}
+
+					out[key] = float32(floatVal)
+				case reflect.Int:
+					intVal, err := strconv.ParseInt(vals[0], 10, 64)
+					if err != nil {
+						return nil, fmt.Errorf("invalid int value %s", vals)
+					}
+
+					out[key] = intVal
+				case reflect.Bool:
+					boolVal, err := strconv.ParseBool(vals[0])
+					if err != nil {
+						return nil, fmt.Errorf("invalid bool value %s", vals)
+					}
+
+					out[key] = boolVal
+				case reflect.String:
+					out[key] = vals[0]
+				case reflect.Slice:
+					// TODO: only string slices are supported right now
+					out[key] = vals
+				default:
+					return nil, fmt.Errorf("unknown type %s for %s", field.Kind(), key)
+				}
+			}
+		}
+	}
+
+	return out, nil
+}

cmd/spinner.go

@@ -1,44 +0,0 @@
-package cmd
-
-import (
-	"fmt"
-	"os"
-	"time"
-
-	"github.com/jmorganca/ollama/progressbar"
-)
-
-type Spinner struct {
-	description string
-	*progressbar.ProgressBar
-}
-
-func NewSpinner(description string) *Spinner {
-	return &Spinner{
-		description: description,
-		ProgressBar: progressbar.NewOptions(-1,
-			progressbar.OptionSetWriter(os.Stderr),
-			progressbar.OptionThrottle(60*time.Millisecond),
-			progressbar.OptionSpinnerType(14),
-			progressbar.OptionSetRenderBlankState(true),
-			progressbar.OptionSetElapsedTime(false),
-			progressbar.OptionClearOnFinish(),
-			progressbar.OptionSetDescription(description),
-		),
-	}
-}
-
-func (s *Spinner) Spin(tick time.Duration) {
-	for range time.Tick(tick) {
-		if s.IsFinished() {
-			break
-		}
-
-		s.Add(1)
-	}
-}
-
-func (s *Spinner) Stop() {
-	s.Finish()
-	fmt.Println(s.description)
-}

docs/api.md

@@ -3,6 +3,7 @@
 ## Endpoints
 
 - [Generate a completion](#generate-a-completion)
+- [Generate a chat completion](#generate-a-chat-completion)
 - [Create a Model](#create-a-model)
 - [List Local Models](#list-local-models)
 - [Show Model Information](#show-model-information)
@@ -24,7 +25,7 @@ All durations are returned in nanoseconds.
 
 ### Streaming responses
 
-Certain endpoints stream responses as JSON objects delineated with the newline (`\n`) character.
+Certain endpoints stream responses as JSON objects.
 
 ## Generate a completion
 
@@ -32,33 +33,36 @@ Certain endpoints stream responses as JSON objects delineated with the newline (
 POST /api/generate
 ```
 
-Generate a response for a given prompt with a provided model. This is a streaming endpoint, so will be a series of responses. The final response object will include statistics and additional data from the request.
+Generate a response for a given prompt with a provided model. This is a streaming endpoint, so there will be a series of responses. The final response object will include statistics and additional data from the request.
 
 ### Parameters
 
 - `model`: (required) the [model name](#model-names)
 - `prompt`: the prompt to generate a response for
+- `images`: a list of base64-encoded images (for multimodal models such as `llava`)
 
 Advanced parameters (optional):
 
 - `format`: the format to return a response in. Currently the only accepted value is `json`
 - `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`)
+- `system`: system message to (overrides what is defined in the `Modelfile`)
 - `template`: the full prompt or prompt template (overrides what is defined in the `Modelfile`)
 - `context`: the context parameter returned from a previous request to `/generate`, this can be used to keep a short conversational memory
 - `stream`: if `false` the response will be returned as a single response object, rather than a stream of objects
-- `raw`: if `true` no formatting will be applied to the prompt and no context will be returned. You may choose to use the `raw` parameter if you are specifying a full templated prompt in your request to the API, and are managing history yourself.
+- `raw`: if `true` no formatting will be applied to the prompt. You may choose to use the `raw` parameter if you are specifying a full templated prompt in your request to the API.
 
 ### JSON mode
 
-Enable JSON mode by setting the `format` parameter to `json` and specifying the model should use JSON in the `prompt`. This will structure the response as valid JSON. See the JSON mode [example](#request-json-mode) below.
+Enable JSON mode by setting the `format` parameter to `json`. This will structure the response as valid JSON. See the JSON mode [example](#request-json-mode) below.
+
+> Note: it's important to instruct the model to use JSON in the `prompt`. Otherwise, the model may generate large amounts whitespace.
 
 ### Examples
 
 #### Request
 
 ```shell
-curl -X POST http://localhost:11434/api/generate -d '{
+curl http://localhost:11434/api/generate -d '{
   "model": "llama2",
   "prompt": "Why is the sky blue?"
 }'
@@ -112,9 +116,11 @@ To calculate how fast the response is generated in tokens per second (token/s),
 
 #### Request (No streaming)
 
+A response can be recieved in one reply when streaming is off.
+
 ```shell
-curl -X POST http://localhost:11434/api/generate -d '{
-  "model": "llama2:7b",
+curl http://localhost:11434/api/generate -d '{
+  "model": "llama2",
   "prompt": "Why is the sky blue?",
   "stream": false
 }'
@@ -126,7 +132,7 @@ If `stream` is set to `false`, the response will be a single JSON object:
 
 ```json
 {
-  "model": "llama2:7b",
+  "model": "llama2",
   "created_at": "2023-08-04T19:22:45.499127Z",
   "response": "The sky is blue because it is the color of the sky.",
   "context": [1, 2, 3],
@@ -142,12 +148,43 @@ If `stream` is set to `false`, the response will be a single JSON object:
 }
 ```
 
-#### Request (Raw mode)
+#### Request (with images)
 
-In some cases you may wish to bypass the templating system and provide a full prompt. In this case, you can use the `raw` parameter to disable formatting and context.
+To submit images to multimodal models such as `llava` or `bakllava`, provide a list of base64-encoded `images`:
 
 ```shell
-curl -X POST http://localhost:11434/api/generate -d '{
+curl http://localhost:11434/api/generate -d '{
+  "model": "llava",
+  "prompt":"What is in this picture?",
+  "stream": false,
+  "images": ["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"]
+}'
+```
+
+#### Response
+
+```
+{
+  "model": "llava",
+  "created_at": "2023-11-03T15:36:02.583064Z",
+  "response": "A happy cartoon character, which is cute and cheerful.",
+  "context": [1, 2, 3],
+  "done": true,
+  "total_duration": 14648695333,
+  "load_duration": 3302671417,
+  "prompt_eval_count": 14,
+  "prompt_eval_duration": 286243000,
+  "eval_count": 129,
+  "eval_duration": 10931424000
+}
+```
+
+#### Request (Raw Mode)
+
+In some cases you may wish to bypass the templating system and provide a full prompt. In this case, you can use the `raw` parameter to disable formatting.
+
+```shell
+curl http://localhost:11434/api/generate -d '{
   "model": "mistral",
   "prompt": "[INST] why is the sky blue? [/INST]",
   "raw": true,
@@ -162,6 +199,7 @@ curl -X POST http://localhost:11434/api/generate -d '{
   "model": "mistral",
   "created_at": "2023-11-03T15:36:02.583064Z",
   "response": " The sky appears blue because of a phenomenon called Rayleigh scattering.",
+  "context": [1, 2, 3],
   "done": true,
   "total_duration": 14648695333,
   "load_duration": 3302671417,
@@ -175,7 +213,7 @@ curl -X POST http://localhost:11434/api/generate -d '{
 #### Request (JSON mode)
 
 ```shell
-curl -X POST http://localhost:11434/api/generate -d '{
+curl http://localhost:11434/api/generate -d '{
   "model": "llama2",
   "prompt": "What color is the sky at different times of the day? Respond using JSON",
   "format": "json",
@@ -224,8 +262,8 @@ The value of `response` will be a string containing JSON similar to:
 If you want to set custom options for the model at runtime rather than in the Modelfile, you can do so with the `options` parameter. This example sets every available option, but you can set any of them individually and omit the ones you do not want to override.
 
 ```shell
-curl -X POST http://localhost:11434/api/generate -d '{
-  "model": "llama2:7b",
+curl http://localhost:11434/api/generate -d '{
+  "model": "llama2",
   "prompt": "Why is the sky blue?",
   "stream": false,
   "options": {
@@ -247,7 +285,7 @@ curl -X POST http://localhost:11434/api/generate -d '{
     "penalize_newline": true,
     "stop": ["\n", "user:"],
     "numa": false,
-    "num_ctx": 4,
+    "num_ctx": 1024,
     "num_batch": 2,
     "num_gqa": 1,
     "num_gpu": 1,
@@ -262,7 +300,7 @@ curl -X POST http://localhost:11434/api/generate -d '{
     "rope_frequency_base": 1.1,
     "rope_frequency_scale": 0.8,
     "num_thread": 8
-    }
+  }
 }'
 ```
 
@@ -270,10 +308,9 @@ curl -X POST http://localhost:11434/api/generate -d '{
 
 ```json
 {
-  "model": "llama2:7b",
+  "model": "llama2",
   "created_at": "2023-08-04T19:22:45.499127Z",
   "response": "The sky is blue because it is the color of the sky.",
-  "context": [1, 2, 3],
   "done": true,
   "total_duration": 5589157167,
   "load_duration": 3013701500,
@@ -286,28 +323,182 @@ curl -X POST http://localhost:11434/api/generate -d '{
 }
 ```
 
+## Generate a chat completion
+
+```shell
+POST /api/chat
+```
+
+Generate the next message in a chat with a provided model. This is a streaming endpoint, so there will be a series of responses. The final response object will include statistics and additional data from the request.
+
+### Parameters
+
+- `model`: (required) the [model name](#model-names)
+- `messages`: the messages of the chat, this can be used to keep a chat memory
+
+The `message` object has the following fields:
+
+- `role`: the role of the message, either `system`, `user` or `assistant`
+- `content`: the content of the message
+- `images` (optional): a list of images to include in the message (for multimodal models such as `llava`)
+
+Advanced parameters (optional):
+
+- `format`: the format to return a response in. Currently the only accepted value is `json`
+- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
+- `template`: the full prompt or prompt template (overrides what is defined in the `Modelfile`)
+- `stream`: if `false` the response will be returned as a single response object, rather than a stream of objects
+
+### Examples
+
+#### Request
+
+Send a chat message with a streaming response.
+
+```shell
+curl http://localhost:11434/api/chat -d '{
+  "model": "llama2",
+  "messages": [
+    {
+      "role": "user",
+      "content": "why is the sky blue?"
+    }
+  ]
+}'
+```
+
+#### Response
+
+A stream of JSON objects is returned:
+
+```json
+{
+  "model": "llama2",
+  "created_at": "2023-08-04T08:52:19.385406455-07:00",
+  "message": {
+    "role": "assisant",
+    "content": "The"
+  },
+  "done": false
+}
+```
+
+Final response:
+
+```json
+{
+  "model": "llama2",
+  "created_at": "2023-08-04T19:22:45.499127Z",
+  "done": true,
+  "total_duration": 5589157167,
+  "load_duration": 3013701500,
+  "sample_count": 114,
+  "sample_duration": 81442000,
+  "prompt_eval_count": 46,
+  "prompt_eval_duration": 1160282000,
+  "eval_count": 113,
+  "eval_duration": 1325948000
+}
+```
+
+#### Request (With History)
+
+Send a chat message with a conversation history.
+
+```shell
+curl http://localhost:11434/api/chat -d '{
+  "model": "llama2",
+  "messages": [
+    {
+      "role": "user",
+      "content": "why is the sky blue?"
+    },
+    {
+      "role": "assistant",
+      "content": "due to rayleigh scattering."
+    },
+    {
+      "role": "user",
+      "content": "how is that different than mie scattering?"
+    }
+  ]
+}'
+```
+
+#### Response
+
+A stream of JSON objects is returned:
+
+```json
+{
+  "model": "llama2",
+  "created_at": "2023-08-04T08:52:19.385406455-07:00",
+  "message": {
+    "role": "assisant",
+    "content": "The"
+  },
+  "done": false
+}
+```
+
+Final response:
+
+```json
+{
+  "model": "llama2",
+  "created_at": "2023-08-04T19:22:45.499127Z",
+  "done": true,
+  "total_duration": 5589157167,
+  "load_duration": 3013701500,
+  "sample_count": 114,
+  "sample_duration": 81442000,
+  "prompt_eval_count": 46,
+  "prompt_eval_duration": 1160282000,
+  "eval_count": 113,
+  "eval_duration": 1325948000
+}
+```
+
+#### Request (with images)
+
+Send a chat message with a conversation history.
+
+```shell
+curl http://localhost:11434/api/chat -d '{
+  "model": "llama2",
+  "messages": [
+    {
+      "role": "user",
+      "content": "what is in this image?",
+      "images": ["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"]
+    },
+  ]
+}'
+```
+
 ## Create a Model
 
 ```shell
 POST /api/create
 ```
 
-Create a model from a [`Modelfile`](./modelfile.md)
+Create a model from a [`Modelfile`](./modelfile.md). It is recommended to set `modelfile` to the content of the Modelfile rather than just set `path`. This is a requirement for remote create. Remote model creation should also create any file blobs, fields such as `FROM` and `ADAPTER`, explicitly with the server using [Create a Blob](#create-a-blob) and the value to the path indicated in the response.
 
 ### Parameters
 
 - `name`: name of the model to create
-- `path`: path to the Modelfile
+- `modelfile` (optional): contents of the Modelfile
 - `stream`: (optional) if `false` the response will be returned as a single response object, rather than a stream of objects
+- `path` (optional): path to the Modelfile
 
 ### Examples
 
 #### Request
 
 ```shell
-curl -X POST http://localhost:11434/api/create -d '{
+curl http://localhost:11434/api/create -d '{
   "name": "mario",
-  "path": "~/Modelfile"
+  "modelfile": "FROM llama2\nSYSTEM You are mario from Super Mario Bros."
 }'
 ```
 
@@ -321,6 +512,54 @@ A stream of JSON objects. When finished, `status` is `success`.
 }
 ```
 
+### Check if a Blob Exists
+
+```shell
+HEAD /api/blobs/:digest
+```
+
+Check if a blob is known to the server.
+
+#### Query Parameters
+
+- `digest`: the SHA256 digest of the blob
+
+#### Examples
+
+##### Request
+
+```shell
+curl -I http://localhost:11434/api/blobs/sha256:29fdb92e57cf0827ded04ae6461b5931d01fa595843f55d36f5b275a52087dd2
+```
+
+##### Response
+
+Return 200 OK if the blob exists, 404 Not Found if it does not.
+
+### Create a Blob
+
+```shell
+POST /api/blobs/:digest
+```
+
+Create a blob from a file. Returns the server file path.
+
+#### Query Parameters
+
+- `digest`: the expected SHA256 digest of the file
+
+#### Examples
+
+##### Request
+
+```shell
+curl -T model.bin -X POST http://localhost:11434/api/blobs/sha256:29fdb92e57cf0827ded04ae6461b5931d01fa595843f55d36f5b275a52087dd2
+```
+
+##### Response
+
+Return 201 Created if the blob was successfully created.
+
 ## List Local Models
 
 ```shell
@@ -345,7 +584,7 @@ A single JSON object will be returned.
 {
   "models": [
     {
-      "name": "llama2:7b",
+      "name": "llama2",
       "modified_at": "2023-08-02T17:02:23.713454393-07:00",
       "size": 3791730596
     },
@@ -364,7 +603,7 @@ A single JSON object will be returned.
 POST /api/show
 ```
 
-Show details about a model including modelfile, template, parameters, license, and system prompt.
+Show information about a model including details, modelfile, template, parameters, license, and system prompt.
 
 ### Parameters
 
@@ -376,7 +615,7 @@ Show details about a model including modelfile, template, parameters, license, a
 
 ```shell
 curl http://localhost:11434/api/show -d '{
-  "name": "llama2:7b"
+  "name": "llama2"
 }'
 ```
 
@@ -384,10 +623,16 @@ curl http://localhost:11434/api/show -d '{
 
 ```json
 {
-  "license": "<contents of license block>",
-  "modelfile": "# Modelfile generated by \"ollama show\"\n# To build a new Modelfile based on this one, replace the FROM line with:\n# FROM llama2:latest\n\nFROM /Users/username/.ollama/models/blobs/sha256:8daa9615cce30c259a9555b1cc250d461d1bc69980a274b44d7eda0be78076d8\nTEMPLATE \"\"\"[INST] {{ if and .First .System }}<<SYS>>{{ .System }}<</SYS>>\n\n{{ end }}{{ .Prompt }} [/INST] \"\"\"\nSYSTEM \"\"\"\"\"\"\nPARAMETER stop [INST]\nPARAMETER stop [/INST]\nPARAMETER stop <<SYS>>\nPARAMETER stop <</SYS>>\n",
-  "parameters": "stop                           [INST]\nstop                           [/INST]\nstop                           <<SYS>>\nstop                           <</SYS>>",
-  "template": "[INST] {{ if and .First .System }}<<SYS>>{{ .System }}<</SYS>>\n\n{{ end }}{{ .Prompt }} [/INST] "
+  "modelfile": "# Modelfile generated by \"ollama show\"\n# To build a new Modelfile based on this one, replace the FROM line with:\n# FROM llava:latest\n\nFROM mike/llava:latest\nTEMPLATE \"\"\"\nUSER:{{ .Prompt }}\nASSISTANT:\n\"\"\"\nPARAMETER num_ctx 4096",
+  "parameters": "num_ctx                        4096",
+  "template": "\nUSER:{{ .Prompt }}\nASSISTANT:\n",
+  "license:": "<license>",
+  "details": {
+    "format": "gguf",
+    "families": ["llama", "clip"],
+    "parameter_size": "7B",
+    "quantization_level": "Q4_0"
+  }
 }
 ```
 
@@ -405,7 +650,7 @@ Copy a model. Creates a model with another name from an existing model.
 
 ```shell
 curl http://localhost:11434/api/copy -d '{
-  "source": "llama2:7b",
+  "source": "llama2",
   "destination": "llama2-backup"
 }'
 ```
@@ -459,8 +704,8 @@ Download a model from the ollama library. Cancelled pulls are resumed from where
 #### Request
 
 ```shell
-curl -X POST http://localhost:11434/api/pull -d '{
-  "name": "llama2:7b"
+curl http://localhost:11434/api/pull -d '{
+  "name": "llama2"
 }'
 ```
 
@@ -531,7 +776,7 @@ Upload a model to a model library. Requires registering for ollama.ai and adding
 #### Request
 
 ```shell
-curl -X POST http://localhost:11434/api/push -d '{
+curl http://localhost:11434/api/push -d '{
   "name": "mattw/pygmalion:latest"
 }'
 ```
@@ -599,8 +844,8 @@ Advanced parameters:
 #### Request
 
 ```shell
-curl -X POST http://localhost:11434/api/embeddings -d '{
-  "model": "llama2:7b",
+curl http://localhost:11434/api/embeddings -d '{
+  "model": "llama2",
   "prompt": "Here is an article about llamas..."
 }'
 ```

docs/faq.md

@@ -23,7 +23,7 @@ Ollama binds to 127.0.0.1 port 11434 by default. Change the bind address with th
 On macOS:
 
 ```bash
-OLLAMA_HOST=0.0.0.0:11435 ollama serve
+OLLAMA_HOST=0.0.0.0:11434 ollama serve
 ```
 
 On Linux:
@@ -32,11 +32,11 @@ Create a `systemd` drop-in directory and set `Environment=OLLAMA_HOST`
 
 ```bash
 mkdir -p /etc/systemd/system/ollama.service.d
-echo "[Service]" >>/etc/systemd/system/ollama.service.d/environment.conf
+echo '[Service]' >>/etc/systemd/system/ollama.service.d/environment.conf
 ```
 
 ```bash
-echo "Environment=OLLAMA_HOST=0.0.0.0:11434" >>/etc/systemd/system/ollama.service.d/environment.conf
+echo 'Environment="OLLAMA_HOST=0.0.0.0:11434"' >>/etc/systemd/system/ollama.service.d/environment.conf
 ```
 
 Reload `systemd` and restart Ollama:
@@ -59,7 +59,7 @@ OLLAMA_ORIGINS=http://192.168.1.1:*,https://example.com ollama serve
 On Linux:
 
 ```bash
-echo "Environment=OLLAMA_ORIGINS=http://129.168.1.1:*,https://example.com" >>/etc/systemd/system/ollama.service.d/environment.conf
+echo 'Environment="OLLAMA_ORIGINS=http://192.168.1.1:*,https://example.com"' >>/etc/systemd/system/ollama.service.d/environment.conf
 ```
 
 Reload `systemd` and restart Ollama:
@@ -74,8 +74,6 @@ systemctl restart ollama
 - macOS: Raw model data is stored under `~/.ollama/models`.
 - Linux: Raw model data is stored under `/usr/share/ollama/.ollama/models`
 
-
-
 Below the models directory you will find a structure similar to the following:
 
 ```shell
@@ -96,3 +94,63 @@ The manifest lists all the layers used in this model. You will see a `media type
 ### How can I change where Ollama stores models?
 
 To modify where models are stored, you can use the `OLLAMA_MODELS` environment variable. Note that on Linux this means defining `OLLAMA_MODELS` in a drop-in `/etc/systemd/system/ollama.service.d` service file, reloading systemd, and restarting the ollama service.
+
+## Does Ollama send my prompts and answers back to Ollama.ai to use in any way?
+
+No. Anything you do with Ollama, such as generate a response from the model, stays with you. We don't collect any data about how you use the model. You are always in control of your own data.
+
+## How can I use Ollama in Visual Studio Code?
+
+There is already a large collection of plugins available for VSCode as well as other editors that leverage Ollama. You can see the list of [extensions & plugins](https://github.com/jmorganca/ollama#extensions--plugins) at the bottom of the main repository readme.
+
+## How do I use Ollama behind a proxy?
+
+Ollama is compatible with proxy servers if `HTTP_PROXY` or `HTTPS_PROXY` are configured. When using either variables, ensure it is set where `ollama serve` can access the values.
+
+When using `HTTPS_PROXY`, ensure the proxy certificate is installed as a system certificate.
+
+On macOS:
+
+```bash
+HTTPS_PROXY=http://proxy.example.com ollama serve
+```
+
+On Linux:
+
+```bash
+echo 'Environment="HTTPS_PROXY=https://proxy.example.com"' >>/etc/systemd/system/ollama.service.d/environment.conf
+```
+
+Reload `systemd` and restart Ollama:
+
+```bash
+systemctl daemon-reload
+systemctl restart ollama
+```
+
+### How do I use Ollama behind a proxy in Docker?
+
+The Ollama Docker container image can be configured to use a proxy by passing `-e HTTPS_PROXY=https://proxy.example.com` when starting the container.
+
+Alternatively, Docker daemon can be configured to use a proxy. Instructions are available for Docker Desktop on [macOS](https://docs.docker.com/desktop/settings/mac/#proxies), [Windows](https://docs.docker.com/desktop/settings/windows/#proxies), and [Linux](https://docs.docker.com/desktop/settings/linux/#proxies), and Docker [daemon with systemd](https://docs.docker.com/config/daemon/systemd/#httphttps-proxy).
+
+Ensure the certificate is installed as a system certificate when using HTTPS. This may require a new Docker image when using a self-signed certificate.
+
+```dockerfile
+FROM ollama/ollama
+COPY my-ca.pem /usr/local/share/ca-certificates/my-ca.crt
+RUN update-ca-certificates
+```
+
+Build and run this image:
+
+```shell
+docker build -t ollama-with-ca .
+docker run -d -e HTTPS_PROXY=https://my.proxy.example.com -p 11434:11434 ollama-with-ca
+```
+
+## How do I use Ollama with GPU acceleration in Docker?
+
+The Ollama Docker container can be configured with GPU acceleration in Linux or Windows (with WSL2). This requires the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia-container-toolkit). See [ollama/ollama](https://hub.docker.com/r/ollama/ollama) for more details.
+
+GPU acceleration is not available for Docker Desktop in macOS due to the lack of GPU passthrough and emulation.

docs/import.md

@@ -43,7 +43,6 @@ Ollama supports a set of model architectures, with support for more coming soon:
 
 - Llama & Mistral
 - Falcon & RW
-- GPT-NeoX
 - BigCode
 
 To view a model's architecture, check the `config.json` file in its HuggingFace repo. You should see an entry under `architectures` (e.g. `LlamaForCausalLM`).
@@ -184,9 +183,6 @@ python convert.py <path to model directory>
 # FalconForCausalLM
 python convert-falcon-hf-to-gguf.py <path to model directory>
 
-# GPTNeoXForCausalLM
-python convert-gptneox-hf-to-gguf.py <path to model directory>
-
 # GPTBigCodeForCausalLM
 python convert-starcoder-hf-to-gguf.py <path to model directory>
 ```

docs/modelfile.md

@@ -30,17 +30,19 @@ The format of the `Modelfile`:
 INSTRUCTION arguments
 ```
 
-| Instruction                         | Description                                                   |
-| ----------------------------------- | ------------------------------------------------------------- |
-| [`FROM`](#from-required) (required) | Defines the base model to use.                                |
-| [`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.                                  |
+| Instruction                         | Description                                                    |
+| ----------------------------------- | -------------------------------------------------------------- |
+| [`FROM`](#from-required) (required) | Defines the base model to use.                                 |
+| [`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 message 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
 
+### Basic `Modelfile`
+
 An example of a `Modelfile` creating a mario blueprint:
 
 ```modelfile
@@ -50,7 +52,7 @@ PARAMETER temperature 1
 # sets the context window size to 4096, this controls how many tokens the LLM can use as context to generate the next token
 PARAMETER num_ctx 4096
 
-# sets a custom system prompt to specify the behavior of the chat assistant
+# sets a custom system message to specify the behavior of the chat assistant
 SYSTEM You are Mario from super mario bros, acting as an assistant.
 ```
 
@@ -63,6 +65,35 @@ To use this:
 
 More examples are available in the [examples directory](../examples).
 
+### `Modelfile`s in [ollama.ai/library][1]
+
+There are two ways to view `Modelfile`s underlying the models in [ollama.ai/library][1]:
+
+- Option 1: view a details page from a model's tags page:
+  1.  Go to a particular model's tags (e.g. https://ollama.ai/library/llama2/tags)
+  2.  Click on a tag (e.g. https://ollama.ai/library/llama2:13b)
+  3.  Scroll down to "Layers"
+      - Note: if the [`FROM` instruction](#from-required) is not present,
+        it means the model was created from a local file
+- Option 2: use `ollama show` to print the `Modelfile` like so:
+
+  ```bash
+  > ollama show --modelfile llama2:13b
+  # Modelfile generated by "ollama show"
+  # To build a new Modelfile based on this one, replace the FROM line with:
+  # FROM llama2:13b
+
+  FROM /root/.ollama/models/blobs/sha256:123abc
+  TEMPLATE """[INST] {{ if and .First .System }}<<SYS>>{{ .System }}<</SYS>>
+
+  {{ end }}{{ .Prompt }} [/INST] """
+  SYSTEM """"""
+  PARAMETER stop [INST]
+  PARAMETER stop [/INST]
+  PARAMETER stop <<SYS>>
+  PARAMETER stop <</SYS>>
+  ```
+
 ## Instructions
 
 ### FROM (Required)
@@ -121,15 +152,15 @@ PARAMETER <parameter> <parametervalue>
 
 ### TEMPLATE
 
-`TEMPLATE` of the full prompt template to be passed into the model. It may include (optionally) a system prompt and a user's prompt. This is used to create a full custom prompt, and syntax may be model specific. You can usually find the template for a given model in the readme for that model.
+`TEMPLATE` of the full prompt template to be passed into the model. It may include (optionally) a system message and a user's prompt. This is used to create a full custom prompt, and syntax may be model specific. You can usually find the template for a given model in the readme for that model.
 
 #### Template Variables
 
-| Variable        | Description                                                                                                  |
-| --------------- | ------------------------------------------------------------------------------------------------------------ |
-| `{{ .System }}` | The system prompt used to specify custom behavior, this must also be set in the Modelfile as an instruction. |
-| `{{ .Prompt }}` | The incoming prompt, this is not specified in the model file and will be set based on input.                 |
-| `{{ .First }}`  | A boolean value used to render specific template information for the first generation of a session.          |
+| Variable        | Description                                                                                                   |
+| --------------- | ------------------------------------------------------------------------------------------------------------- |
+| `{{ .System }}` | The system message used to specify custom behavior, this must also be set in the Modelfile as an instruction. |
+| `{{ .Prompt }}` | The incoming prompt, this is not specified in the model file and will be set based on input.                  |
+| `{{ .First }}`  | A boolean value used to render specific template information for the first generation of a session.           |
 
 ```modelfile
 TEMPLATE """
@@ -149,7 +180,7 @@ SYSTEM """<system message>"""
 
 ### SYSTEM
 
-The `SYSTEM` instruction specifies the system prompt to be used in the template, if applicable.
+The `SYSTEM` instruction specifies the system message to be used in the template, if applicable.
 
 ```modelfile
 SYSTEM """<system message>"""
@@ -177,3 +208,5 @@ LICENSE """
 
 - 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.
+
+[1]: https://ollama.ai/library

docs/tutorials/fly-gpu.md

@@ -0,0 +1,83 @@
+# Running Ollama on Fly.io GPU Instances
+
+Ollama runs with little to no configuration on [Fly.io GPU instances](https://fly.io/docs/gpus/gpu-quickstart/). If you don't have access to GPUs yet, you'll need to [apply for access](https://fly.io/gpu/) on the waitlist. Once you're accepted, you'll get an email with instructions on how to get started.
+
+Create a new app with `fly apps create`:
+
+```bash
+fly apps create
+```
+
+Then create a `fly.toml` file in a new folder that looks like this:
+
+```toml
+app = "sparkling-violet-709"
+primary_region = "ord"
+vm.size = "a100-40gb" # see https://fly.io/docs/gpus/gpu-quickstart/ for more info
+
+[build]
+  image = "ollama/ollama"
+
+[http_service]
+  internal_port = 11434
+  force_https = false
+  auto_stop_machines = true
+  auto_start_machines = true
+  min_machines_running = 0
+  processes = ["app"]
+
+[mounts]
+  source = "models"
+  destination = "/root/.ollama"
+  initial_size = "100gb"
+```
+
+Then create a [new private IPv6 address](https://fly.io/docs/reference/private-networking/#flycast-private-load-balancing) for your app:
+
+```bash
+fly ips allocate-v6 --private
+```
+
+Then deploy your app:
+
+```bash
+fly deploy
+```
+
+And finally you can access it interactively with a new Fly.io Machine:
+
+```
+fly machine run -e OLLAMA_HOST=http://your-app-name.flycast --shell ollama/ollama
+```
+
+```bash
+$ ollama run openchat:7b-v3.5-fp16
+>>> How do I bake chocolate chip cookies?
+ To bake chocolate chip cookies, follow these steps:
+
+1. Preheat the oven to 375°F (190°C) and line a baking sheet with parchment paper or silicone baking mat.
+
+2. In a large bowl, mix together 1 cup of unsalted butter (softened), 3/4 cup granulated sugar, and 3/4
+cup packed brown sugar until light and fluffy.
+
+3. Add 2 large eggs, one at a time, to the butter mixture, beating well after each addition. Stir in 1
+teaspoon of pure vanilla extract.
+
+4. In a separate bowl, whisk together 2 cups all-purpose flour, 1/2 teaspoon baking soda, and 1/2 teaspoon
+salt. Gradually add the dry ingredients to the wet ingredients, stirring until just combined.
+
+5. Fold in 2 cups of chocolate chips (or chunks) into the dough.
+
+6. Drop rounded tablespoons of dough onto the prepared baking sheet, spacing them about 2 inches apart.
+
+7. Bake for 10-12 minutes, or until the edges are golden brown. The centers should still be slightly soft.
+
+8. Allow the cookies to cool on the baking sheet for a few minutes before transferring them to a wire rack
+to cool completely.
+
+Enjoy your homemade chocolate chip cookies!
+```
+
+When you set it up like this, it will automatically turn off when you're done using it. Then when you access it again, it will automatically turn back on. This is a great way to save money on GPU instances when you're not using them. If you want a persistent wake-on-use connection to your Ollama instance, you can set up a [connection to your Fly network using WireGuard](https://fly.io/docs/reference/private-networking/#discovering-apps-through-dns-on-a-wireguard-connection). Then you can access your Ollama instance at `http://your-app-name.flycast`.
+
+And that's it!

docs/tutorials/langchainpy.md

@@ -42,12 +42,13 @@ 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`
+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. We can use Ollama directly to instantiate an embedding model. We will use ChromaDB in this example for a vector database. `pip install GPT4All chromadb`
 
 ```python
-from langchain.embeddings import GPT4AllEmbeddings
+from langchain.embeddings import OllamaEmbeddings
 from langchain.vectorstores import Chroma
-vectorstore = Chroma.from_documents(documents=all_splits, embedding=GPT4AllEmbeddings())
+oembed = OllamaEmbeddings(base_url="http://localhost:11434", model="llama2")
+vectorstore = Chroma.from_documents(documents=all_splits, embedding=oembed)
 ```
 
 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.

examples/jupyter-notebook/README.md

@@ -0,0 +1,5 @@
+# Ollama Jupyter Notebook
+
+This example downloads and installs Ollama in a Jupyter instance such as Google Colab. It will start the Ollama service and expose an endpoint using `ngrok` which can be used to communicate with the Ollama instance remotely.
+
+For best results, use an instance with GPU accelerator.

examples/jupyter-notebook/ollama.ipynb

@@ -0,0 +1,102 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "93f59dcb-c588-41b8-a792-55d88ade739c",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Download and run the Ollama Linux install script\n",
+    "!curl https://ollama.ai/install.sh | sh\n",
+    "!command -v systemctl >/dev/null && sudo systemctl stop ollama"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "658c147e-c7f8-490e-910e-62b80f577dda",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "!pip install aiohttp pyngrok\n",
+    "\n",
+    "import os\n",
+    "import asyncio\n",
+    "from aiohttp import ClientSession\n",
+    "\n",
+    "# Set LD_LIBRARY_PATH so the system NVIDIA library becomes preferred\n",
+    "# over the built-in library. This is particularly important for \n",
+    "# Google Colab which installs older drivers\n",
+    "os.environ.update({'LD_LIBRARY_PATH': '/usr/lib64-nvidia'})\n",
+    "\n",
+    "async def run(cmd):\n",
+    "  '''\n",
+    "  run is a helper function to run subcommands asynchronously.\n",
+    "  '''\n",
+    "  print('>>> starting', *cmd)\n",
+    "  p = await asyncio.subprocess.create_subprocess_exec(\n",
+    "      *cmd,\n",
+    "      stdout=asyncio.subprocess.PIPE,\n",
+    "      stderr=asyncio.subprocess.PIPE,\n",
+    "  )\n",
+    "\n",
+    "  async def pipe(lines):\n",
+    "    async for line in lines:\n",
+    "      print(line.strip().decode('utf-8'))\n",
+    "\n",
+    "  await asyncio.gather(\n",
+    "      pipe(p.stdout),\n",
+    "      pipe(p.stderr),\n",
+    "  )\n",
+    "\n",
+    "\n",
+    "await asyncio.gather(\n",
+    "    run(['ollama', 'serve']),\n",
+    "    run(['ngrok', 'http', '--log', 'stderr', '11434']),\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "e7735a55-9aad-4caf-8683-52e2163ba53b",
+   "metadata": {},
+   "source": [
+    "The previous cell starts two processes, `ollama` and `ngrok`. The log output will show a line like the following which describes the external address.\n",
+    "\n",
+    "```\n",
+    "t=2023-11-12T22:55:56+0000 lvl=info msg=\"started tunnel\" obj=tunnels name=command_line addr=http://localhost:11434 url=https://8249-34-125-179-11.ngrok.io\n",
+    "```\n",
+    "\n",
+    "The external address in this case is `https://8249-34-125-179-11.ngrok.io` which can be passed into `OLLAMA_HOST` to access this instance.\n",
+    "\n",
+    "```bash\n",
+    "export OLLAMA_HOST=https://8249-34-125-179-11.ngrok.io\n",
+    "ollama list\n",
+    "ollama run mistral\n",
+    "```"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3 (ipykernel)",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.11.6"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

examples/kubernetes/gpu.yaml

@@ -25,9 +25,11 @@ spec:
         image: ollama/ollama:latest
         env:
         - name: PATH
-          value: /usr/local/nvidia/bin:/usr/local/nvidia/lib64:/usr/bin:/usr/sbin:/bin:/sbin
+          value: /usr/local/nvidia/bin:/usr/local/cuda/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
         - name: LD_LIBRARY_PATH
-          value: /usr/local/nvidia/lib64
+          value: /usr/local/nvidia/lib:/usr/local/nvidia/lib64
+        - name: NVIDIA_DRIVER_CAPABILITIES
+          value: compute,utility
         ports:
         - name: http
           containerPort: 11434

examples/python-json-datagenerator/predefinedschema.py

@@ -0,0 +1,31 @@
+import requests
+import json
+import random
+
+model = "llama2"
+template = {
+  "firstName": "", 
+  "lastName": "", 
+  "address": {
+    "street": "", 
+    "city": "", 
+    "state": "", 
+    "zipCode": ""
+  }, 
+  "phoneNumber": ""
+}
+
+prompt = f"generate one realistically believable sample data set of a persons first name, last name, address in the US, and  phone number. \nUse the following template: {json.dumps(template)}."
+
+data = {
+    "prompt": prompt,
+    "model": model,
+    "format": "json",
+    "stream": False,
+    "options": {"temperature": 2.5, "top_p": 0.99, "top_k": 100},
+}
+
+print(f"Generating a sample user")
+response = requests.post("http://localhost:11434/api/generate", json=data, stream=False)
+json_data = json.loads(response.text)
+print(json.dumps(json.loads(json_data["response"]), indent=2))

examples/python-json-datagenerator/randomaddresses.py

@@ -0,0 +1,31 @@
+import requests
+import json
+import random
+
+countries = [
+    "United States",
+    "United Kingdom",
+    "the Netherlands",
+    "Germany",
+    "Mexico",
+    "Canada",
+    "France",
+]
+country = random.choice(countries)
+model = "llama2"
+
+prompt = f"generate one realistically believable sample data set of a persons first name, last name, address in {country}, and phone number. Do not use common names. Respond using JSON. Key names should have no backslashes, values should use plain ascii with no special characters."
+
+data = {
+    "prompt": prompt,
+    "model": model,
+    "format": "json",
+    "stream": False,
+    "options": {"temperature": 2.5, "top_p": 0.99, "top_k": 100},
+}
+
+print(f"Generating a sample user in {country}")
+response = requests.post("http://localhost:11434/api/generate", json=data, stream=False)
+json_data = json.loads(response.text)
+
+print(json.dumps(json.loads(json_data["response"]), indent=2))

examples/python-json-datagenerator/readme.md

@@ -0,0 +1,34 @@
+# JSON Output Example
+
+![llmjson 2023-11-10 15_31_31](https://github.com/jmorganca/ollama/assets/633681/e599d986-9b4a-4118-81a4-4cfe7e22da25)
+
+There are two python scripts in this example. `randomaddresses.py` generates random addresses from different countries. `predefinedschema.py` sets a template for the model to fill in.
+
+## Review the Code
+
+Both programs are basically the same, with a different prompt for each, demonstrating two different ideas. The key part of getting JSON out of a model is to state in the prompt or system prompt that it should respond using JSON, and specifying the `format` as `json` in the data body.
+
+```python
+prompt = f"generate one realistically believable sample data set of a persons first name, last name, address in {country}, and  phone number. Do not use common names. Respond using JSON. Key names should with no backslashes, values should use plain ascii with no special characters."
+
+data = {
+    "prompt": prompt,
+    "model": model,
+    "format": "json",
+    "stream": False,
+    "options": {"temperature": 2.5, "top_p": 0.99, "top_k": 100},
+}
+```
+
+When running `randomaddresses.py` you will see that the schema changes and adapts to the chosen country.
+
+In `predefinedschema.py`, a template has been specified in the prompt as well. It's been defined as JSON and then dumped into the prompt string to make it easier to work with.
+
+Both examples turn streaming off so that we end up with the completed JSON all at once. We need to convert the `response.text` to JSON so that when we output it as a string we can set the indent spacing to make the output easy to read.
+
+```python
+response = requests.post("http://localhost:11434/api/generate", json=data, stream=False)
+json_data = json.loads(response.text)
+
+print(json.dumps(json.loads(json_data["response"]), indent=2))
+```

examples/python-json-datagenerator/requirements.txt

@@ -0,0 +1 @@
+Requests==2.31.0

examples/python-loganalysis/Modelfile

@@ -0,0 +1,8 @@
+FROM codebooga:latest
+
+SYSTEM """
+You are a log file analyzer. You will receive a set of lines from a log file for some software application, find the errors and other interesting aspects of the logs, and explain them so a new user can understand what they mean. If there are any steps they can do to resolve them, list the steps in your answer.
+"""
+
+PARAMETER TEMPERATURE 0.3
+

examples/python-loganalysis/loganalysis.py

@@ -0,0 +1,42 @@
+import sys
+import re
+import requests
+import json
+
+# prelines and postlines represent the number of lines of context to include in the output around the error
+prelines = 10
+postlines = 10
+
+def find_errors_in_log_file():
+  if len(sys.argv) < 2:
+    print("Usage: python loganalysis.py <filename>")
+    return
+
+  log_file_path = sys.argv[1]
+  with open(log_file_path, 'r') as log_file:
+    log_lines = log_file.readlines()
+
+error_logs = []
+    for i, line in enumerate(log_lines):
+        if "error" in line.lower():
+            start_index = max(0, i - prelines)
+            end_index = min(len(log_lines), i + postlines + 1)
+            error_logs.extend(log_lines[start_index:end_index])
+
+  return error_logs
+
+error_logs = find_errors_in_log_file()
+
+data = {
+  "prompt": "\n".join(error_logs), 
+  "model": "mattw/loganalyzer"
+}
+
+
+response = requests.post("http://localhost:11434/api/generate", json=data, stream=True)
+for line in response.iter_lines():
+  if line:
+    json_data = json.loads(line)
+    if json_data['done'] == False:
+      print(json_data['response'], end='', flush=True)
+

examples/python-loganalysis/logtest.logfile

@@ -0,0 +1,32 @@
+2023-11-10 07:17:40 /docker-entrypoint.sh: /docker-entrypoint.d/ is not empty, will attempt to perform configuration
+2023-11-10 07:17:40 /docker-entrypoint.sh: Looking for shell scripts in /docker-entrypoint.d/
+2023-11-10 07:17:40 /docker-entrypoint.sh: Launching /docker-entrypoint.d/10-listen-on-ipv6-by-default.sh
+2023-11-10 07:17:40 10-listen-on-ipv6-by-default.sh: info: Getting the checksum of /etc/nginx/conf.d/default.conf
+2023-11-10 07:17:40 10-listen-on-ipv6-by-default.sh: info: Enabled listen on IPv6 in /etc/nginx/conf.d/default.conf
+2023-11-10 07:17:40 /docker-entrypoint.sh: Sourcing /docker-entrypoint.d/15-local-resolvers.envsh
+2023-11-10 07:17:40 /docker-entrypoint.sh: Launching /docker-entrypoint.d/20-envsubst-on-templates.sh
+2023-11-10 07:17:40 /docker-entrypoint.sh: Launching /docker-entrypoint.d/30-tune-worker-processes.sh
+2023-11-10 07:17:40 /docker-entrypoint.sh: Configuration complete; ready for start up
+2023-11-10 07:17:40 2023/11/10 13:17:40 [notice] 1#1: using the "epoll" event method
+2023-11-10 07:17:40 2023/11/10 13:17:40 [notice] 1#1: nginx/1.25.3
+2023-11-10 07:17:40 2023/11/10 13:17:40 [notice] 1#1: built by gcc 12.2.0 (Debian 12.2.0-14) 
+2023-11-10 07:17:40 2023/11/10 13:17:40 [notice] 1#1: OS: Linux 6.4.16-linuxkit
+2023-11-10 07:17:40 2023/11/10 13:17:40 [notice] 1#1: getrlimit(RLIMIT_NOFILE): 1048576:1048576
+2023-11-10 07:17:40 2023/11/10 13:17:40 [notice] 1#1: start worker processes
+2023-11-10 07:17:40 2023/11/10 13:17:40 [notice] 1#1: start worker process 29
+2023-11-10 07:17:40 2023/11/10 13:17:40 [notice] 1#1: start worker process 30
+2023-11-10 07:17:40 2023/11/10 13:17:40 [notice] 1#1: start worker process 31
+2023-11-10 07:17:40 2023/11/10 13:17:40 [notice] 1#1: start worker process 32
+2023-11-10 07:17:40 2023/11/10 13:17:40 [notice] 1#1: start worker process 33
+2023-11-10 07:17:40 2023/11/10 13:17:40 [notice] 1#1: start worker process 34
+2023-11-10 07:17:40 2023/11/10 13:17:40 [notice] 1#1: start worker process 35
+2023-11-10 07:17:40 2023/11/10 13:17:40 [notice] 1#1: start worker process 36
+2023-11-10 07:17:40 2023/11/10 13:17:40 [notice] 1#1: start worker process 37
+2023-11-10 07:17:40 2023/11/10 13:17:40 [notice] 1#1: start worker process 38
+2023-11-10 07:17:44 192.168.65.1 - - [10/Nov/2023:13:17:43 +0000] "GET / HTTP/1.1" 200 615 "-" "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/119.0.0.0 Safari/537.36" "-"
+2023-11-10 07:17:44 2023/11/10 13:17:44 [error] 29#29: *1 open() "/usr/share/nginx/html/favicon.ico" failed (2: No such file or directory), client: 192.168.65.1, server: localhost, request: "GET /favicon.ico HTTP/1.1", host: "localhost:8080", referrer: "http://localhost:8080/"
+2023-11-10 07:17:44 192.168.65.1 - - [10/Nov/2023:13:17:44 +0000] "GET /favicon.ico HTTP/1.1" 404 555 "http://localhost:8080/" "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/119.0.0.0 Safari/537.36" "-"
+2023-11-10 07:17:50 2023/11/10 13:17:50 [error] 29#29: *1 open() "/usr/share/nginx/html/ahstat" failed (2: No such file or directory), client: 192.168.65.1, server: localhost, request: "GET /ahstat HTTP/1.1", host: "localhost:8080"
+2023-11-10 07:17:50 192.168.65.1 - - [10/Nov/2023:13:17:50 +0000] "GET /ahstat HTTP/1.1" 404 555 "-" "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/119.0.0.0 Safari/537.36" "-"
+2023-11-10 07:18:53 2023/11/10 13:18:53 [error] 29#29: *1 open() "/usr/share/nginx/html/ahstat" failed (2: No such file or directory), client: 192.168.65.1, server: localhost, request: "GET /ahstat HTTP/1.1", host: "localhost:8080"
+2023-11-10 07:18:53 192.168.65.1 - - [10/Nov/2023:13:18:53 +0000] "GET /ahstat HTTP/1.1" 404 555 "-" "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/119.0.0.0 Safari/537.36" "-"

examples/python-loganalysis/readme.md

@@ -0,0 +1,48 @@
+# Log Analysis example
+
+![loganalyzer 2023-11-10 08_53_29](https://github.com/jmorganca/ollama/assets/633681/ad30f1fc-321f-4953-8914-e30e24db9921)
+
+This example shows one possible way to create a log file analyzer. To use it, run:
+
+`python loganalysis.py <logfile>`
+
+You can try this with the `logtest.logfile` file included in this directory.
+
+## Review the code
+
+The first part of this example is a Modelfile that takes `codebooga` and applies a new System Prompt:
+
+```plaintext
+SYSTEM """
+You are a log file analyzer. You will receive a set of lines from a log file for some software application, find the errors and other interesting aspects of the logs, and explain them so a new user can understand what they mean. If there are any steps they can do to resolve them, list the steps in your answer.
+"""
+```
+
+This model is available at https://ollama.ai/mattw/loganalyzer. You can customize it and add to your own namespace using the command `ollama create <namespace/modelname> -f <path-to-modelfile>` then `ollama push <namespace/modelname>`.
+
+Then loganalysis.py scans all the lines in the given log file and searches for the word 'error'. When the word is found, the 10 lines before and after are set as the prompt for a call to the Generate API.
+
+```python
+data = {
+  "prompt": "\n".join(error_logs), 
+  "model": "mattw/loganalyzer"
+}
+```
+
+Finally, the streamed output is parsed and the response field in the output is printed to the line.
+
+```python
+response = requests.post("http://localhost:11434/api/generate", json=data, stream=True)
+for line in response.iter_lines():
+  if line:
+    json_data = json.loads(line)
+    if json_data['done'] == False:
+      print(json_data['response'], end='')
+
+```
+
+## Next Steps
+
+There is a lot more that can be done here. This is a simple way to detect errors, looking for the word error. Perhaps it would be interesting to find anomalous activity in the logs. It could be interesting to create embeddings for each line and compare them, looking for similar lines. Or look into applying Levenshtein Distance algorithms to find similar lines to help identify the anomalous lines.
+
+Also try different models and different prompts to analyze the data. You could consider adding retrieval augmented generation (RAG) to this to help understand newer log formats.

examples/python-loganalysis/requirements.txt

@@ -0,0 +1 @@
+Requests==2.31.0

examples/python-simplechat/client.py

@@ -0,0 +1,46 @@
+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 chat(messages):
+    r = requests.post(
+        "http://0.0.0.0:11434/api/chat",
+        json={"model": model, "messages": messages, "stream": True},
+    )
+    r.raise_for_status()
+    output = ""
+
+    for line in r.iter_lines():
+        body = json.loads(line)
+        if "error" in body:
+            raise Exception(body["error"])
+        if body.get("done") is False:
+            message = body.get("message", "")
+            content = message.get("content", "")
+            output += content
+            # the response streams one token at a time, print that as we receive it
+            print(content, end="", flush=True)
+
+
+        if body.get("done", False):
+            message["content"] = output
+            return message
+
+
+def main():
+    messages = []
+    
+    while True:
+        user_input = input("Enter a prompt: ")
+        print()
+        messages.append({"role": "user", "content": user_input})
+        message = chat(messages)
+        messages.append(message)
+        print("\n\n")
+
+
+if __name__ == "__main__":
+    main()

examples/python-simplechat/readme.md

@@ -0,0 +1,24 @@
+# Simple Chat Example
+
+The **chat** endpoint is one of two ways to generate text from an LLM with Ollama. At a high level you provide the endpoint an array of objects with a role and content specified. Then with each output and prompt, you add more of those role/content objects, which builds up the history.
+
+## Review the Code
+
+You can see in the **chat** function that actually calling the endpoint is done simply with:
+
+```python
+r = requests.post(
+  "http://0.0.0.0:11434/api/chat",
+  json={"model": model, "messages": messages, "stream": True},
+)
+```
+
+With the **generate** endpoint, you need to provide a `prompt`. But with **chat**, you provide `messages`. And the resulting stream of responses includes a `message` object with a `content` field.
+
+The final JSON object doesn't provide the full content, so you will need to build the content yourself.
+
+In the **main** function, we collect `user_input` and add it as a message to our messages and that is passed to the chat function. When the LLM is done responding the output is added as another message.
+
+## Next Steps
+
+In this example, all generations are kept. You might want to experiment with summarizing everything older than 10 conversations to enable longer history with less context being used.

examples/typescript-functioncalling/extractemail.ts

@@ -0,0 +1,118 @@
+import { Ollama } from "ollama-node";
+import { readFile } from "fs/promises";
+
+// function to be called on events
+function reportEvents(name: string, date: string, location: string) {
+  const nameString = name ? `${name}` : `an event`;
+  const dateString = date ? ` on ${date}` : ``;
+  const locationString = location ? ` at ${location}` : ``;
+  console.log(`You have an event: ${nameString}${dateString}${locationString}`)
+}
+
+// function to be called on addresses
+function reportAddresses(address) {
+  for (const field in address) {
+    if (address[field]) {
+      if (field === "city") {
+        const city = address.city;
+        const state = address.state ? `, ${address.state}` : '';
+        const zip = address.zip ? ` ${address.zip}` : '';
+        console.log(`${city}${state}${zip}`);
+        break;
+      } else {
+        console.log(`${address[field]}`);
+      }
+    }
+  }
+  console.log(``);
+}
+
+async function main() {
+
+  const ollama = new Ollama();
+
+  const systemprompt = `You will be given a text along with a prompt and a schema. You will have to extract the information requested in the prompt from the text and generate output in JSON observing the schema provided. If the schema shows a type of integer or number, you must only show a integer for that field. A string should always be a valid string. If a value is unknown, leave it empty. Output the JSON with extra spaces to ensure that it pretty prints.`
+
+  const schema = {
+    "eventsQuantity": {
+      "type": "integer",
+      "description": "The number of events in the source text"
+    },
+    "addressesQuantity": {
+      "type": "integer",
+      "description": "The number of addresses in the source text"
+    },
+    "events": [{
+      name: {
+        "type": "string",
+        description: "Name of the event"
+      },
+      "date": {
+        "type": "string",
+        "description": "Date of the event"
+      },
+      "location": {
+        "type": "string",
+        "description": "Location of the event"
+      },
+      "extraInfo": {
+        "type": "string",
+        "description": "Any extra information that is provided about the event."
+      }
+    }],
+    "people": [{
+      "name": {
+        "type": "string",
+        "description": "Name of the person"
+      },
+      "company": {
+        "type": "string",
+        "description": "Name of the company where they work"
+      },
+      "street": {
+        "type": "string",
+        "description": "Street address of the person or company. This is only the street name and the numerical address. Do not include city, state, or zip of the address in this field."
+      },
+      "city": {
+        "type": "string",
+        "description": "City portion of the address of the person or company"
+      },
+      "state": {
+        "type": "string",
+        "description": "State portion of the address of the person or company"
+      },
+      "zip": {
+        "type": "string",
+        "description": "Zip code of the person or company"
+      },
+      "extraInfo": {
+        "type": "string",
+        "description": "Any extra information that is provided about the location."
+      }
+    }]
+  }
+
+  const textcontent = await readFile("./info.txt", "utf-8").then((text) => text.split(" ").slice(0, 2000).join(" "));
+
+  const prompt = `The source text is a series of emails that have been put into a single file. They are separated by three dashes. Review the source text and determine the full address of the person sending each of the emails as well as any events that we need to track. If they provide a company address use that. If any extra info is provided, such as a description of the place, or a floor, add it to extraInfo. The first field in the address JSON is quantity of events and should be set to the number of events tracked and the second field should be set to the number of addresses tracked in the file. Don't stuff an event into the output that isn't an event. Only add data to the mostly appropriate field. Don't make up fields that aren't in the schema. If there isn't a value for a field, use null. Output should be in JSON.\n\nSchema: \n${JSON.stringify(schema, null, 2)}\n\nSource Text:\n${textcontent}`
+
+  await ollama.setModel("neural-chat");
+  ollama.setSystemPrompt(systemprompt);
+  ollama.setJSONFormat(true);
+  const data = await ollama.generate(prompt);
+  const output = JSON.parse(data.output);
+  const events = output.events;
+  const addresses = output.people;
+
+  console.log(`Here are your ${output.eventsQuantity} events:`);
+  for (const event of events) {
+    reportEvents(event.name, event.date, event.location);
+  }
+
+  console.log(`\n\nHere are your ${output.addressesQuantity} addresses:`);
+  for (const address of addresses) {
+    reportAddresses(address);
+  }
+}
+
+main();

examples/typescript-functioncalling/extractwp.ts

@@ -0,0 +1,38 @@
+import { Ollama } from "ollama-node";
+import { readFile } from "fs/promises";
+
+async function main() {
+
+  const ollama = new Ollama();
+
+  // Set the system prompt to prepare the model to receive a prompt and a schema and set some rules for the output.
+  const systemprompt = `You will be given a text along with a prompt and a schema. You will have to extract the information requested in the prompt from the text and generate output in JSON observing the schema provided. If the schema shows a type of integer or number, you must only show a integer for that field. A string should always be a valid string. If a value is unknown, leave it empty. Output the JSON with extra spaces to ensure that it pretty prints.`
+
+  const schema = {
+    "people": [{
+      "name": {
+        "type": "string",
+        "description": "Name of the person"
+      },
+      "title": {
+        "type": "string",
+        "description": "Title of the person"
+      }
+    }],
+  }
+
+  // Depending on the model chosen, you may be limited by the size of the context window, so limit the context to 2000 words.
+  const textcontent = await readFile("./wp.txt", "utf-8").then((text) => text.split(" ").slice(0, 2000).join(" "));
+
+  // Specific instructions for this task
+  const prompt = `Review the source text and determine the 10 most important people to focus on. Then extract the name and title for those people. Output should be in JSON.\n\nSchema: \n${JSON.stringify(schema, null, 2)}\n\nSource Text:\n${textcontent}`
+
+  await ollama.setModel("neural-chat");
+  ollama.setSystemPrompt(systemprompt);
+
+  // setJSONFormat is the equivalent of setting 'format: json' in the API
+  ollama.setJSONFormat(true);
+  await ollama.streamingGenerate(prompt, (word) => { process.stdout.write(word) })
+}
+
+main();

examples/typescript-functioncalling/info.txt

@@ -0,0 +1,17 @@
+---
+Hi matt, 
+
+thanks for letting me know that you are going to come today, November 16, for my tea party. My address is 123 Falk St on Bainbridge Island. I live in the house with the red door. I will be home all day so just come by whenever you want.
+
+Fred
+
+---
+Great, send the check to our office at 1917 1st St, Seattle, WA 98101. I will let you know when we receive it.
+
+Mark Richardson
+Big Corp
+---
+We are looking forward to seeing you at our Local AI Meetup. It will be held on December 3. It will be at the offices of Enormous Co. Our address is 344 1st Ave, Seattle, WA 98101. We will be meeting in the conference room on the 3rd floor.
+
+Barbara Reilly
+Enormous Co.

examples/typescript-functioncalling/package-lock.json

@@ -0,0 +1,519 @@
+{
+  "name": "typescript-functioncalling",
+  "lockfileVersion": 3,
+  "requires": true,
+  "packages": {
+    "": {
+      "dependencies": {
+        "ollama-node": "^0.1.27"
+      },
+      "devDependencies": {
+        "tsx": "^4.1.2",
+        "typescript": "^5.2.2"
+      }
+    },
+    "node_modules/@esbuild/android-arm": {
+      "version": "0.18.20",
+      "resolved": "https://registry.npmjs.org/@esbuild/android-arm/-/android-arm-0.18.20.tgz",
+      "integrity": "sha512-fyi7TDI/ijKKNZTUJAQqiG5T7YjJXgnzkURqmGj13C6dCqckZBLdl4h7bkhHt/t0WP+zO9/zwroDvANaOqO5Sw==",
+      "cpu": [
+        "arm"
+      ],
+      "dev": true,
+      "optional": true,
+      "os": [
+        "android"
+      ],
+      "engines": {
+        "node": ">=12"
+      }
+    },
+    "node_modules/@esbuild/android-arm64": {
+      "version": "0.18.20",
+      "resolved": "https://registry.npmjs.org/@esbuild/android-arm64/-/android-arm64-0.18.20.tgz",
+      "integrity": "sha512-Nz4rJcchGDtENV0eMKUNa6L12zz2zBDXuhj/Vjh18zGqB44Bi7MBMSXjgunJgjRhCmKOjnPuZp4Mb6OKqtMHLQ==",
+      "cpu": [
+        "arm64"
+      ],
+      "dev": true,
+      "optional": true,
+      "os": [
+        "android"
+      ],
+      "engines": {
+        "node": ">=12"
+      }
+    },
+    "node_modules/@esbuild/android-x64": {
+      "version": "0.18.20",
+      "resolved": "https://registry.npmjs.org/@esbuild/android-x64/-/android-x64-0.18.20.tgz",
+      "integrity": "sha512-8GDdlePJA8D6zlZYJV/jnrRAi6rOiNaCC/JclcXpB+KIuvfBN4owLtgzY2bsxnx666XjJx2kDPUmnTtR8qKQUg==",
+      "cpu": [
+        "x64"
+      ],
+      "dev": true,
+      "optional": true,
+      "os": [
+        "android"
+      ],
+      "engines": {
+        "node": ">=12"
+      }
+    },
+    "node_modules/@esbuild/darwin-arm64": {
+      "version": "0.18.20",
+      "resolved": "https://registry.npmjs.org/@esbuild/darwin-arm64/-/darwin-arm64-0.18.20.tgz",
+      "integrity": "sha512-bxRHW5kHU38zS2lPTPOyuyTm+S+eobPUnTNkdJEfAddYgEcll4xkT8DB9d2008DtTbl7uJag2HuE5NZAZgnNEA==",
+      "cpu": [
+        "arm64"
+      ],
+      "dev": true,
+      "optional": true,
+      "os": [
+        "darwin"
+      ],
+      "engines": {
+        "node": ">=12"
+      }
+    },
+    "node_modules/@esbuild/darwin-x64": {
+      "version": "0.18.20",
+      "resolved": "https://registry.npmjs.org/@esbuild/darwin-x64/-/darwin-x64-0.18.20.tgz",
+      "integrity": "sha512-pc5gxlMDxzm513qPGbCbDukOdsGtKhfxD1zJKXjCCcU7ju50O7MeAZ8c4krSJcOIJGFR+qx21yMMVYwiQvyTyQ==",
+      "cpu": [
+        "x64"
+      ],
+      "dev": true,
+      "optional": true,
+      "os": [
+        "darwin"
+      ],
+      "engines": {
+        "node": ">=12"
+      }
+    },
+    "node_modules/@esbuild/freebsd-arm64": {
+      "version": "0.18.20",
+      "resolved": "https://registry.npmjs.org/@esbuild/freebsd-arm64/-/freebsd-arm64-0.18.20.tgz",
+      "integrity": "sha512-yqDQHy4QHevpMAaxhhIwYPMv1NECwOvIpGCZkECn8w2WFHXjEwrBn3CeNIYsibZ/iZEUemj++M26W3cNR5h+Tw==",
+      "cpu": [
+        "arm64"
+      ],
+      "dev": true,
+      "optional": true,
+      "os": [
+        "freebsd"
+      ],
+      "engines": {
+        "node": ">=12"
+      }
+    },
+    "node_modules/@esbuild/freebsd-x64": {
+      "version": "0.18.20",
+      "resolved": "https://registry.npmjs.org/@esbuild/freebsd-x64/-/freebsd-x64-0.18.20.tgz",
+      "integrity": "sha512-tgWRPPuQsd3RmBZwarGVHZQvtzfEBOreNuxEMKFcd5DaDn2PbBxfwLcj4+aenoh7ctXcbXmOQIn8HI6mCSw5MQ==",
+      "cpu": [
+        "x64"
+      ],
+      "dev": true,
+      "optional": true,
+      "os": [
+        "freebsd"
+      ],
+      "engines": {
+        "node": ">=12"
+      }
+    },
+    "node_modules/@esbuild/linux-arm": {
+      "version": "0.18.20",
+      "resolved": "https://registry.npmjs.org/@esbuild/linux-arm/-/linux-arm-0.18.20.tgz",
+      "integrity": "sha512-/5bHkMWnq1EgKr1V+Ybz3s1hWXok7mDFUMQ4cG10AfW3wL02PSZi5kFpYKrptDsgb2WAJIvRcDm+qIvXf/apvg==",
+      "cpu": [
+        "arm"
+      ],
+      "dev": true,
+      "optional": true,
+      "os": [
+        "linux"
+      ],
+      "engines": {
+        "node": ">=12"
+      }
+    },
+    "node_modules/@esbuild/linux-arm64": {
+      "version": "0.18.20",
+      "resolved": "https://registry.npmjs.org/@esbuild/linux-arm64/-/linux-arm64-0.18.20.tgz",
+      "integrity": "sha512-2YbscF+UL7SQAVIpnWvYwM+3LskyDmPhe31pE7/aoTMFKKzIc9lLbyGUpmmb8a8AixOL61sQ/mFh3jEjHYFvdA==",
+      "cpu": [
+        "arm64"
+      ],
+      "dev": true,
+      "optional": true,
+      "os": [
+        "linux"
+      ],
+      "engines": {
+        "node": ">=12"
+      }
+    },
+    "node_modules/@esbuild/linux-ia32": {
+      "version": "0.18.20",
+      "resolved": "https://registry.npmjs.org/@esbuild/linux-ia32/-/linux-ia32-0.18.20.tgz",
+      "integrity": "sha512-P4etWwq6IsReT0E1KHU40bOnzMHoH73aXp96Fs8TIT6z9Hu8G6+0SHSw9i2isWrD2nbx2qo5yUqACgdfVGx7TA==",
+      "cpu": [
+        "ia32"
+      ],
+      "dev": true,
+      "optional": true,
+      "os": [
+        "linux"
+      ],
+      "engines": {
+        "node": ">=12"
+      }
+    },
+    "node_modules/@esbuild/linux-loong64": {
+      "version": "0.18.20",
+      "resolved": "https://registry.npmjs.org/@esbuild/linux-loong64/-/linux-loong64-0.18.20.tgz",
+      "integrity": "sha512-nXW8nqBTrOpDLPgPY9uV+/1DjxoQ7DoB2N8eocyq8I9XuqJ7BiAMDMf9n1xZM9TgW0J8zrquIb/A7s3BJv7rjg==",
+      "cpu": [
+        "loong64"
+      ],
+      "dev": true,
+      "optional": true,
+      "os": [
+        "linux"
+      ],
+      "engines": {
+        "node": ">=12"
+      }
+    },
+    "node_modules/@esbuild/linux-mips64el": {
+      "version": "0.18.20",
+      "resolved": "https://registry.npmjs.org/@esbuild/linux-mips64el/-/linux-mips64el-0.18.20.tgz",
+      "integrity": "sha512-d5NeaXZcHp8PzYy5VnXV3VSd2D328Zb+9dEq5HE6bw6+N86JVPExrA6O68OPwobntbNJ0pzCpUFZTo3w0GyetQ==",
+      "cpu": [
+        "mips64el"
+      ],
+      "dev": true,
+      "optional": true,
+      "os": [
+        "linux"
+      ],
+      "engines": {
+        "node": ">=12"
+      }
+    },
+    "node_modules/@esbuild/linux-ppc64": {
+      "version": "0.18.20",
+      "resolved": "https://registry.npmjs.org/@esbuild/linux-ppc64/-/linux-ppc64-0.18.20.tgz",
+      "integrity": "sha512-WHPyeScRNcmANnLQkq6AfyXRFr5D6N2sKgkFo2FqguP44Nw2eyDlbTdZwd9GYk98DZG9QItIiTlFLHJHjxP3FA==",
+      "cpu": [
+        "ppc64"
+      ],
+      "dev": true,
+      "optional": true,
+      "os": [
+        "linux"
+      ],
+      "engines": {
+        "node": ">=12"
+      }
+    },
+    "node_modules/@esbuild/linux-riscv64": {
+      "version": "0.18.20",
+      "resolved": "https://registry.npmjs.org/@esbuild/linux-riscv64/-/linux-riscv64-0.18.20.tgz",
+      "integrity": "sha512-WSxo6h5ecI5XH34KC7w5veNnKkju3zBRLEQNY7mv5mtBmrP/MjNBCAlsM2u5hDBlS3NGcTQpoBvRzqBcRtpq1A==",
+      "cpu": [
+        "riscv64"
+      ],
+      "dev": true,
+      "optional": true,
+      "os": [
+        "linux"
+      ],
+      "engines": {
+        "node": ">=12"
+      }
+    },
+    "node_modules/@esbuild/linux-s390x": {
+      "version": "0.18.20",
+      "resolved": "https://registry.npmjs.org/@esbuild/linux-s390x/-/linux-s390x-0.18.20.tgz",
+      "integrity": "sha512-+8231GMs3mAEth6Ja1iK0a1sQ3ohfcpzpRLH8uuc5/KVDFneH6jtAJLFGafpzpMRO6DzJ6AvXKze9LfFMrIHVQ==",
+      "cpu": [
+        "s390x"
+      ],
+      "dev": true,
+      "optional": true,
+      "os": [
+        "linux"
+      ],
+      "engines": {
+        "node": ">=12"
+      }
+    },
+    "node_modules/@esbuild/linux-x64": {
+      "version": "0.18.20",
+      "resolved": "https://registry.npmjs.org/@esbuild/linux-x64/-/linux-x64-0.18.20.tgz",
+      "integrity": "sha512-UYqiqemphJcNsFEskc73jQ7B9jgwjWrSayxawS6UVFZGWrAAtkzjxSqnoclCXxWtfwLdzU+vTpcNYhpn43uP1w==",
+      "cpu": [
+        "x64"
+      ],
+      "dev": true,
+      "optional": true,
+      "os": [
+        "linux"
+      ],
+      "engines": {
+        "node": ">=12"
+      }
+    },
+    "node_modules/@esbuild/netbsd-x64": {
+      "version": "0.18.20",
+      "resolved": "https://registry.npmjs.org/@esbuild/netbsd-x64/-/netbsd-x64-0.18.20.tgz",
+      "integrity": "sha512-iO1c++VP6xUBUmltHZoMtCUdPlnPGdBom6IrO4gyKPFFVBKioIImVooR5I83nTew5UOYrk3gIJhbZh8X44y06A==",
+      "cpu": [
+        "x64"
+      ],
+      "dev": true,
+      "optional": true,
+      "os": [
+        "netbsd"
+      ],
+      "engines": {
+        "node": ">=12"
+      }
+    },
+    "node_modules/@esbuild/openbsd-x64": {
+      "version": "0.18.20",
+      "resolved": "https://registry.npmjs.org/@esbuild/openbsd-x64/-/openbsd-x64-0.18.20.tgz",
+      "integrity": "sha512-e5e4YSsuQfX4cxcygw/UCPIEP6wbIL+se3sxPdCiMbFLBWu0eiZOJ7WoD+ptCLrmjZBK1Wk7I6D/I3NglUGOxg==",
+      "cpu": [
+        "x64"
+      ],
+      "dev": true,
+      "optional": true,
+      "os": [
+        "openbsd"
+      ],
+      "engines": {
+        "node": ">=12"
+      }
+    },
+    "node_modules/@esbuild/sunos-x64": {
+      "version": "0.18.20",
+      "resolved": "https://registry.npmjs.org/@esbuild/sunos-x64/-/sunos-x64-0.18.20.tgz",
+      "integrity": "sha512-kDbFRFp0YpTQVVrqUd5FTYmWo45zGaXe0X8E1G/LKFC0v8x0vWrhOWSLITcCn63lmZIxfOMXtCfti/RxN/0wnQ==",
+      "cpu": [
+        "x64"
+      ],
+      "dev": true,
+      "optional": true,
+      "os": [
+        "sunos"
+      ],
+      "engines": {
+        "node": ">=12"
+      }
+    },
+    "node_modules/@esbuild/win32-arm64": {
+      "version": "0.18.20",
+      "resolved": "https://registry.npmjs.org/@esbuild/win32-arm64/-/win32-arm64-0.18.20.tgz",
+      "integrity": "sha512-ddYFR6ItYgoaq4v4JmQQaAI5s7npztfV4Ag6NrhiaW0RrnOXqBkgwZLofVTlq1daVTQNhtI5oieTvkRPfZrePg==",
+      "cpu": [
+        "arm64"
+      ],
+      "dev": true,
+      "optional": true,
+      "os": [
+        "win32"
+      ],
+      "engines": {
+        "node": ">=12"
+      }
+    },
+    "node_modules/@esbuild/win32-ia32": {
+      "version": "0.18.20",
+      "resolved": "https://registry.npmjs.org/@esbuild/win32-ia32/-/win32-ia32-0.18.20.tgz",
+      "integrity": "sha512-Wv7QBi3ID/rROT08SABTS7eV4hX26sVduqDOTe1MvGMjNd3EjOz4b7zeexIR62GTIEKrfJXKL9LFxTYgkyeu7g==",
+      "cpu": [
+        "ia32"
+      ],
+      "dev": true,
+      "optional": true,
+      "os": [
+        "win32"
+      ],
+      "engines": {
+        "node": ">=12"
+      }
+    },
+    "node_modules/@esbuild/win32-x64": {
+      "version": "0.18.20",
+      "resolved": "https://registry.npmjs.org/@esbuild/win32-x64/-/win32-x64-0.18.20.tgz",
+      "integrity": "sha512-kTdfRcSiDfQca/y9QIkng02avJ+NCaQvrMejlsB3RRv5sE9rRoeBPISaZpKxHELzRxZyLvNts1P27W3wV+8geQ==",
+      "cpu": [
+        "x64"
+      ],
+      "dev": true,
+      "optional": true,
+      "os": [
+        "win32"
+      ],
+      "engines": {
+        "node": ">=12"
+      }
+    },
+    "node_modules/@types/node": {
+      "version": "20.9.0",
+      "resolved": "https://registry.npmjs.org/@types/node/-/node-20.9.0.tgz",
+      "integrity": "sha512-nekiGu2NDb1BcVofVcEKMIwzlx4NjHlcjhoxxKBNLtz15Y1z7MYf549DFvkHSId02Ax6kGwWntIBPC3l/JZcmw==",
+      "dependencies": {
+        "undici-types": "~5.26.4"
+      }
+    },
+    "node_modules/buffer-from": {
+      "version": "1.1.2",
+      "resolved": "https://registry.npmjs.org/buffer-from/-/buffer-from-1.1.2.tgz",
+      "integrity": "sha512-E+XQCRwSbaaiChtv6k6Dwgc+bx+Bs6vuKJHHl5kox/BaKbhiXzqQOwK4cO22yElGp2OCmjwVhT3HmxgyPGnJfQ==",
+      "dev": true
+    },
+    "node_modules/esbuild": {
+      "version": "0.18.20",
+      "resolved": "https://registry.npmjs.org/esbuild/-/esbuild-0.18.20.tgz",
+      "integrity": "sha512-ceqxoedUrcayh7Y7ZX6NdbbDzGROiyVBgC4PriJThBKSVPWnnFHZAkfI1lJT8QFkOwH4qOS2SJkS4wvpGl8BpA==",
+      "dev": true,
+      "hasInstallScript": true,
+      "bin": {
+        "esbuild": "bin/esbuild"
+      },
+      "engines": {
+        "node": ">=12"
+      },
+      "optionalDependencies": {
+        "@esbuild/android-arm": "0.18.20",
+        "@esbuild/android-arm64": "0.18.20",
+        "@esbuild/android-x64": "0.18.20",
+        "@esbuild/darwin-arm64": "0.18.20",
+        "@esbuild/darwin-x64": "0.18.20",
+        "@esbuild/freebsd-arm64": "0.18.20",
+        "@esbuild/freebsd-x64": "0.18.20",
+        "@esbuild/linux-arm": "0.18.20",
+        "@esbuild/linux-arm64": "0.18.20",
+        "@esbuild/linux-ia32": "0.18.20",
+        "@esbuild/linux-loong64": "0.18.20",
+        "@esbuild/linux-mips64el": "0.18.20",
+        "@esbuild/linux-ppc64": "0.18.20",
+        "@esbuild/linux-riscv64": "0.18.20",
+        "@esbuild/linux-s390x": "0.18.20",
+        "@esbuild/linux-x64": "0.18.20",
+        "@esbuild/netbsd-x64": "0.18.20",
+        "@esbuild/openbsd-x64": "0.18.20",
+        "@esbuild/sunos-x64": "0.18.20",
+        "@esbuild/win32-arm64": "0.18.20",
+        "@esbuild/win32-ia32": "0.18.20",
+        "@esbuild/win32-x64": "0.18.20"
+      }
+    },
+    "node_modules/fsevents": {
+      "version": "2.3.3",
+      "resolved": "https://registry.npmjs.org/fsevents/-/fsevents-2.3.3.tgz",
+      "integrity": "sha512-5xoDfX+fL7faATnagmWPpbFtwh/R77WmMMqqHGS65C3vvB0YHrgF+B1YmZ3441tMj5n63k0212XNoJwzlhffQw==",
+      "dev": true,
+      "hasInstallScript": true,
+      "optional": true,
+      "os": [
+        "darwin"
+      ],
+      "engines": {
+        "node": "^8.16.0 || ^10.6.0 || >=11.0.0"
+      }
+    },
+    "node_modules/get-tsconfig": {
+      "version": "4.7.2",
+      "resolved": "https://registry.npmjs.org/get-tsconfig/-/get-tsconfig-4.7.2.tgz",
+      "integrity": "sha512-wuMsz4leaj5hbGgg4IvDU0bqJagpftG5l5cXIAvo8uZrqn0NJqwtfupTN00VnkQJPcIRrxYrm1Ue24btpCha2A==",
+      "dev": true,
+      "dependencies": {
+        "resolve-pkg-maps": "^1.0.0"
+      },
+      "funding": {
+        "url": "https://github.com/privatenumber/get-tsconfig?sponsor=1"
+      }
+    },
+    "node_modules/ollama-node": {
+      "version": "0.1.27",
+      "resolved": "https://registry.npmjs.org/ollama-node/-/ollama-node-0.1.27.tgz",
+      "integrity": "sha512-tFABPf5P0sXCR5USA31E3tqbge5h/4uf/t5j8/rPvHDo0SDwXeN0kah2J7hIqqkYlO1vLRs0uLC1/Mprgv9t2g==",
+      "dependencies": {
+        "@types/node": "^20.8.4"
+      }
+    },
+    "node_modules/resolve-pkg-maps": {
+      "version": "1.0.0",
+      "resolved": "https://registry.npmjs.org/resolve-pkg-maps/-/resolve-pkg-maps-1.0.0.tgz",
+      "integrity": "sha512-seS2Tj26TBVOC2NIc2rOe2y2ZO7efxITtLZcGSOnHHNOQ7CkiUBfw0Iw2ck6xkIhPwLhKNLS8BO+hEpngQlqzw==",
+      "dev": true,
+      "funding": {
+        "url": "https://github.com/privatenumber/resolve-pkg-maps?sponsor=1"
+      }
+    },
+    "node_modules/source-map": {
+      "version": "0.6.1",
+      "resolved": "https://registry.npmjs.org/source-map/-/source-map-0.6.1.tgz",
+      "integrity": "sha512-UjgapumWlbMhkBgzT7Ykc5YXUT46F0iKu8SGXq0bcwP5dz/h0Plj6enJqjz1Zbq2l5WaqYnrVbwWOWMyF3F47g==",
+      "dev": true,
+      "engines": {
+        "node": ">=0.10.0"
+      }
+    },
+    "node_modules/source-map-support": {
+      "version": "0.5.21",
+      "resolved": "https://registry.npmjs.org/source-map-support/-/source-map-support-0.5.21.tgz",
+      "integrity": "sha512-uBHU3L3czsIyYXKX88fdrGovxdSCoTGDRZ6SYXtSRxLZUzHg5P/66Ht6uoUlHu9EZod+inXhKo3qQgwXUT/y1w==",
+      "dev": true,
+      "dependencies": {
+        "buffer-from": "^1.0.0",
+        "source-map": "^0.6.0"
+      }
+    },
+    "node_modules/tsx": {
+      "version": "4.1.2",
+      "resolved": "https://registry.npmjs.org/tsx/-/tsx-4.1.2.tgz",
+      "integrity": "sha512-1spM1bFV6MP2s4tO4tDC7g52fsaFdtEWdO4GfGdqi20qUgPbnAJqixOyIAvCSx1DDj3YIUB4CD06owTWUsOAuQ==",
+      "dev": true,
+      "dependencies": {
+        "esbuild": "~0.18.20",
+        "get-tsconfig": "^4.7.2",
+        "source-map-support": "^0.5.21"
+      },
+      "bin": {
+        "tsx": "dist/cli.mjs"
+      },
+      "engines": {
+        "node": ">=18.0.0"
+      },
+      "optionalDependencies": {
+        "fsevents": "~2.3.3"
+      }
+    },
+    "node_modules/typescript": {
+      "version": "5.2.2",
+      "resolved": "https://registry.npmjs.org/typescript/-/typescript-5.2.2.tgz",
+      "integrity": "sha512-mI4WrpHsbCIcwT9cF4FZvr80QUeKvsUsUvKDoR+X/7XHQH98xYD8YHZg7ANtz2GtZt/CBq2QJ0thkGJMHfqc1w==",
+      "dev": true,
+      "bin": {
+        "tsc": "bin/tsc",
+        "tsserver": "bin/tsserver"
+      },
+      "engines": {
+        "node": ">=14.17"
+      }
+    },
+    "node_modules/undici-types": {
+      "version": "5.26.5",
+      "resolved": "https://registry.npmjs.org/undici-types/-/undici-types-5.26.5.tgz",
+      "integrity": "sha512-JlCMO+ehdEIKqlFxk6IfVoAUVmgz7cU7zD/h9XZ0qzeosSHmUJVOzSQvvYSYWXkFXC+IfLKSIffhv0sVZup6pA=="
+    }
+  }
+}

examples/typescript-functioncalling/package.json

@@ -0,0 +1,9 @@
+{
+  "dependencies": {
+    "ollama-node": "^0.1.27"
+  },
+  "devDependencies": {
+    "tsx": "^4.1.2",
+    "typescript": "^5.2.2"
+  }
+}

examples/typescript-functioncalling/readme.md

@@ -0,0 +1,28 @@
+# Function calling
+
+![function calling 2023-11-16 16_12_58](https://github.com/jmorganca/ollama/assets/633681/a0acc247-9746-45ab-b325-b65dfbbee4fb)
+
+One of the features added to some models is 'function calling'. It's a bit of a confusing name. It's understandable if you think that means the model can call functions, but that's not what it means. Function calling simply means that the output of the model is formatted in JSON, using a preconfigured schema, and uses the expected types. Then your code can use the output of the model and call functions with it. Using the JSON format in Ollama, you can use any model for function calling. 
+
+The two examples provided can extract information out of the provided texts. The first example uses the first couple of chapters from War and Peace by Lev Nikolayevich Tolstoy, and extracts the names and titles of the characters introduced in the story. The second example uses a more complicated schema to pull out addresses and event information from a series of emails.
+
+## Running the examples
+
+1. Clone this repo and navigate to the `examples/typescript-functioncalling` directory.
+2. Install the dependencies with `npm install`.
+3. Review the `wp.txt` file.
+4. Run `tsx extractwp.ts`.
+5. Review the `info.txt` file.
+6. Run `tsx extractemail.ts`.
+
+## Review the Code
+
+Both examples do roughly the same thing with different source material. They both use the same system prompt, which tells the model to expect some instructions and a schema. Then we inject the schema into the prompt and generate an answer.
+
+The first example, `extractwp.ts`, outputs the resulting JSON to the console, listing the characters introduced at the start of War and Peace. The second example, `extractemail.ts`, is a bit more complicated, extracting two different types of information: addresses and events. It outputs the results to a JSON blob, then the addresses are handed off to one function called `reportAddresses` and the events are handed off to another function called `reportEvents`.
+
+Notice that both examples are using the model from Intel called `neural-chat`. This is not a model tuned for function calling, yet it performs very well at this task.
+
+## Next Steps
+
+Try exporting some of your real emails to the input file and seeing how well the model does. Try pointing the first example at other books. You could even have it cycle through all the sections and maybe add up the number of times any character is seen throughout the book, determining the most important characters. You can also try out different models.

examples/typescript-functioncalling/wp.txt

@@ -0,0 +1,183 @@
+"Well, Prince, so Genoa and Lucca are now just family estates of the Buonapartes. But I warn you, if you don't tell me that this means war, if you still try to defend the infamies and horrors perpetrated by that Antichrist - I really believe he is Antichrist - I will have nothing more to do with you and you are no longer my friend, no longer my 'faithful slave,' as you call yourself! But how do you do? I see I have frightened you - sit down and tell me all the news."
+
+It was in July, 1805, and the speaker was the well-known Anna Pavlovna Scherer, maid of honor and favorite of the Empress Marya Fedorovna. With these words she greeted Prince Vasili Kuragin, a man of high rank and importance, who was the first to arrive at her reception. Anna Pavlovna had had a cough for some days. She was, as she said, suffering from la grippe; grippe being then a new word in St. Petersburg, used only by the elite.
+
+All her invitations without exception, written in French, and delivered by a scarlet-liveried footman that morning, ran as follows:
+
+"If you have nothing better to do, Count (or Prince), and if the prospect of spending an evening with a poor invalid is not too terrible, I shall be very charmed to see you tonight between 7 and 10 - Annette Scherer."
+
+"Heavens! what a virulent attack!" replied the prince, not in the least disconcerted by this reception. He had just entered, wearing an embroidered court uniform, knee breeches, and shoes, and had stars on his breast and a serene expression on his flat face. He spoke in that refined French in which our grandfathers not only spoke but thought, and with the gentle, patronizing intonation natural to a man of importance who had grown old in society and at court. He went up to Anna Pavlovna, kissed her hand, presenting to her his bald, scented, and shining head, and complacently seated himself on the sofa.
+
+"First of all, dear friend, tell me how you are. Set your friend's mind at rest," said he without altering his tone, beneath the politeness and affected sympathy of which indifference and even irony could be discerned.
+
+"Can one be well while suffering morally? Can one be calm in times like these if one has any feeling?" said Anna Pavlovna. "You are staying the whole evening, I hope?"
+
+"And the fete at the English ambassador's? Today is Wednesday. I must put in an appearance there," said the prince. "My daughter is coming for me to take me there."
+
+"I thought today's fete had been canceled. I confess all these festivities and fireworks are becoming wearisome."
+
+"If they had known that you wished it, the entertainment would have been put off," said the prince, who, like a wound-up clock, by force of habit said things he did not even wish to be believed.
+
+"Don't tease! Well, and what has been decided about Novosiltsev's dispatch? You know everything."
+
+"What can one say about it?" replied the prince in a cold, listless tone. "What has been decided? They have decided that Buonaparte has burnt his boats, and I believe that we are ready to burn ours."
+
+Prince Vasili always spoke languidly, like an actor repeating a stale part. Anna Pavlovna Scherer on the contrary, despite her forty years, overflowed with animation and impulsiveness. To be an enthusiast had become her social vocation and, sometimes even when she did not feel like it, she became enthusiastic in order not to disappoint the expectations of those who knew her. The subdued smile which, though it did not suit her faded features, always played round her lips expressed, as in a spoiled child, a continual consciousness of her charming defect, which she neither wished, nor could, nor considered it necessary, to correct.
+
+In the midst of a conversation on political matters Anna Pavlovna burst out:
+
+"Oh, don't speak to me of Austria. Perhaps I don't understand things, but Austria never has wished, and does not wish, for war. She is betraying us! Russia alone must save Europe. Our gracious sovereign recognizes his high vocation and will be true to it. That is the one thing I have faith in! Our good and wonderful sovereign has to perform the noblest role on earth, and he is so virtuous and noble that God will not forsake him. He will fulfill his vocation and crush the hydra of revolution, which has become more terrible than ever in the person of this murderer and villain! We alone must avenge the blood of the just one.... Whom, I ask you, can we rely on?... England with her commercial spirit will not and cannot understand the Emperor Alexander's loftiness of soul. She has refused to evacuate Malta. She wanted to find, and still seeks, some secret motive in our actions. What answer did Novosiltsev get? None. The English have not understood and cannot understand the self-abnegation of our Emperor who wants nothing for himself, but only desires the good of mankind. And what have they promised? Nothing! And what little they have promised they will not perform! Prussia has always declared that Buonaparte is invincible, and that all Europe is powerless before him.... And I don't believe a word that Hardenburg says, or Haugwitz either. This famous Prussian neutrality is just a trap. I have faith only in God and the lofty destiny of our adored monarch. He will save Europe!"
+
+She suddenly paused, smiling at her own impetuosity.
+
+"I think," said the prince with a smile, "that if you had been sent instead of our dear Wintzingerode you would have captured the King of Prussia's consent by assault. You are so eloquent. Will you give me a cup of tea?"
+
+"In a moment. A propos," she added, becoming calm again, "I am expecting two very interesting men tonight, le Vicomte de Mortemart, who is connected with the Montmorencys through the Rohans, one of the best French families. He is one of the genuine emigres, the good ones. And also the Abbe Morio. Do you know that profound thinker? He has been received by the Emperor. Had you heard?"
+
+"I shall be delighted to meet them," said the prince. "But tell me," he added with studied carelessness as if it had only just occurred to him, though the question he was about to ask was the chief motive of his visit, "is it true that the Dowager Empress wants Baron Funke to be appointed first secretary at Vienna? The baron by all accounts is a poor creature."
+
+Prince Vasili wished to obtain this post for his son, but others were trying through the Dowager Empress Marya Fedorovna to secure it for the baron.
+
+Anna Pavlovna almost closed her eyes to indicate that neither she nor anyone else had a right to criticize what the Empress desired or was pleased with.
+
+"Baron Funke has been recommended to the Dowager Empress by her sister," was all she said, in a dry and mournful tone.
+
+As she named the Empress, Anna Pavlovna's face suddenly assumed an expression of profound and sincere devotion and respect mingled with sadness, and this occurred every time she mentioned her illustrious patroness. She added that Her Majesty had deigned to show Baron Funke beaucoup d'estime, and again her face clouded over with sadness.
+
+The prince was silent and looked indifferent. But, with the womanly and courtierlike quickness and tact habitual to her, Anna Pavlovna wished both to rebuke him (for daring to speak as he had done of a man recommended to the Empress) and at the same time to console him, so she said:
+
+"Now about your family. Do you know that since your daughter came out everyone has been enraptured by her? They say she is amazingly beautiful."
+
+The prince bowed to signify his respect and gratitude.
+
+"I often think," she continued after a short pause, drawing nearer to the prince and smiling amiably at him as if to show that political and social topics were ended and the time had come for intimate conversation - "I often think how unfairly sometimes the joys of life are distributed. Why has fate given you two such splendid children? I don't speak of Anatole, your youngest. I don't like him," she added in a tone admitting of no rejoinder and raising her eyebrows. "Two such charming children. And really you appreciate them less than anyone, and so you don't deserve to have them."
+
+And she smiled her ecstatic smile.
+
+"I can't help it," said the prince. "Lavater would have said I lack the bump of paternity."
+
+"Don't joke; I mean to have a serious talk with you. Do you know I am dissatisfied with your younger son? Between ourselves" (and her face assumed its melancholy expression), "he was mentioned at Her Majesty's and you were pitied...."
+
+The prince answered nothing, but she looked at him significantly, awaiting a reply. He frowned.
+
+"What would you have me do?" he said at last. "You know I did all a father could for their education, and they have both turned out fools. Hippolyte is at least a quiet fool, but Anatole is an active one. That is the only difference between them." He said this smiling in a way more natural and animated than usual, so that the wrinkles round his mouth very clearly revealed something unexpectedly coarse and unpleasant.
+
+"And why are children born to such men as you? If you were not a father there would be nothing I could reproach you with," said Anna Pavlovna, looking up pensively.
+
+"I am your faithful slave and to you alone I can confess that my children are the bane of my life. It is the cross I have to bear. That is how I explain it to myself. It can't be helped!"
+
+He said no more, but expressed his resignation to cruel fate by a gesture. Anna Pavlovna meditated.
+
+"Have you never thought of marrying your prodigal son Anatole?" she asked. "They say old maids have a mania for matchmaking, and though I don't feel that weakness in myself as yet, I know a little person who is very unhappy with her father. She is a relation of yours, Princess Mary Bolkonskaya."
+
+Prince Vasili did not reply, though, with the quickness of memory and perception befitting a man of the world, he indicated by a movement of the head that he was considering this information.
+
+"Do you know," he said at last, evidently unable to check the sad current of his thoughts, "that Anatole is costing me forty thousand rubles a year? And," he went on after a pause, "what will it be in five years, if he goes on like this?" Presently he added: "That's what we fathers have to put up with.... Is this princess of yours rich?"
+
+"Her father is very rich and stingy. He lives in the country. He is the well-known Prince Bolkonski who had to retire from the army under the late Emperor, and was nicknamed 'the King of Prussia.' He is very clever but eccentric, and a bore. The poor girl is very unhappy. She has a brother; I think you know him, he married Lise Meinen lately. He is an aide-de-camp of Kutuzov's and will be here tonight."
+
+"Listen, dear Annette," said the prince, suddenly taking Anna Pavlovna's hand and for some reason drawing it downwards. "Arrange that affair for me and I shall always be your most devoted slave-slafe with an f, as a village elder of mine writes in his reports. She is rich and of good family and that's all I want."
+
+And with the familiarity and easy grace peculiar to him, he raised the maid of honor's hand to his lips, kissed it, and swung it to and fro as he lay back in his armchair, looking in another direction.
+
+"Attendez," said Anna Pavlovna, reflecting, "I'll speak to Lise, young Bolkonski's wife, this very evening, and perhaps the thing can be arranged. It shall be on your family's behalf that I'll start my apprenticeship as old maid."
+
+Anna Pavlovna's drawing room was gradually filling. The highest Petersburg society was assembled there: people differing widely in age and character but alike in the social circle to which they belonged. Prince Vasili's daughter, the beautiful Helene, came to take her father to the ambassador's entertainment; she wore a ball dress and her badge as maid of honor. The youthful little Princess Bolkonskaya, known as la femme la plus seduisante de Petersbourg, * was also there. She had been married during the previous winter, and being pregnant did not go to any large gatherings, but only to small receptions. Prince Vasili's son, Hippolyte, had come with Mortemart, whom he introduced. The Abbe Morio and many others had also come.
+
+* The most fascinating woman in Petersburg.
+
+To each new arrival Anna Pavlovna said, "You have not yet seen my aunt," or "You do not know my aunt?" and very gravely conducted him or her to a little old lady, wearing large bows of ribbon in her cap, who had come sailing in from another room as soon as the guests began to arrive; and slowly turning her eyes from the visitor to her aunt, Anna Pavlovna mentioned each one's name and then left them.
+
+Each visitor performed the ceremony of greeting this old aunt whom not one of them knew, not one of them wanted to know, and not one of them cared about; Anna Pavlovna observed these greetings with mournful and solemn interest and silent approval. The aunt spoke to each of them in the same words, about their health and her own, and the health of Her Majesty, "who, thank God, was better today." And each visitor, though politeness prevented his showing impatience, left the old woman with a sense of relief at having performed a vexatious duty and did not return to her the whole evening.
+
+The young Princess Bolkonskaya had brought some work in a gold-embroidered velvet bag. Her pretty little upper lip, on which a delicate dark down was just perceptible, was too short for her teeth, but it lifted all the more sweetly, and was especially charming when she occasionally drew it down to meet the lower lip. As is always the case with a thoroughly attractive woman, her defect - the shortness of her upper lip and her half-open mouth - seemed to be her own special and peculiar form of beauty. Everyone brightened at the sight of this pretty young woman, so soon to become a mother, so full of life and health, and carrying her burden so lightly. Old men and dull dispirited young ones who looked at her, after being in her company and talking to her a little while, felt as if they too were becoming, like her, full of life and health. All who talked to her, and at each word saw her bright smile and the constant gleam of her white teeth, thought that they were in a specially amiable mood that day.
+
+The little princess went round the table with quick, short, swaying steps, her workbag on her arm, and gaily spreading out her dress sat down on a sofa near the silver samovar, as if all she was doing was a pleasure to herself and to all around her. "I have brought my work," said she in French, displaying her bag and addressing all present. "Mind, Annette, I hope you have not played a wicked trick on me," she added, turning to her hostess. "You wrote that it was to be quite a small reception, and just see how badly I am dressed." And she spread out her arms to show her short-waisted, lace-trimmed, dainty gray dress, girdled with a broad ribbon just below the breast.
+
+"Soyez tranquille, Lise, you will always be prettier than anyone else," replied Anna Pavlovna.
+
+"You know," said the princess in the same tone of voice and still in French, turning to a general, "my husband is deserting me? He is going to get himself killed. Tell me what this wretched war is for?" she added, addressing Prince Vasili, and without waiting for an answer she turned to speak to his daughter, the beautiful Helene.
+
+"What a delightful woman this little princess is!" said Prince Vasili to Anna Pavlovna.
+
+One of the next arrivals was a stout, heavily built young man with close-cropped hair, spectacles, the light-colored breeches fashionable at that time, a very high ruffle, and a brown dress coat. This stout young man was an illegitimate son of Count Bezukhov, a well-known grandee of Catherine's time who now lay dying in Moscow. The young man had not yet entered either the military or civil service, as he had only just returned from abroad where he had been educated, and this was his first appearance in society. Anna Pavlovna greeted him with the nod she accorded to the lowest hierarchy in her drawing room. But in spite of this lowest-grade greeting, a look of anxiety and fear, as at the sight of something too large and unsuited to the place, came over her face when she saw Pierre enter. Though he was certainly rather bigger than the other men in the room, her anxiety could only have reference to the clever though shy, but observant and natural, expression which distinguished him from everyone else in that drawing room.
+
+"It is very good of you, Monsieur Pierre, to come and visit a poor invalid," said Anna Pavlovna, exchanging an alarmed glance with her aunt as she conducted him to her.
+
+Pierre murmured something unintelligible, and continued to look round as if in search of something. On his way to the aunt he bowed to the little princess with a pleased smile, as to an intimate acquaintance.
+
+Anna Pavlovna's alarm was justified, for Pierre turned away from the aunt without waiting to hear her speech about Her Majesty's health. Anna Pavlovna in dismay detained him with the words: "Do you know the Abbe Morio? He is a most interesting man."
+
+"Yes, I have heard of his scheme for perpetual peace, and it is very interesting but hardly feasible."
+
+"You think so?" rejoined Anna Pavlovna in order to say something and get away to attend to her duties as hostess. But Pierre now committed a reverse act of impoliteness. First he had left a lady before she had finished speaking to him, and now he continued to speak to another who wished to get away. With his head bent, and his big feet spread apart, he began explaining his reasons for thinking the abbe's plan chimerical.
+
+"We will talk of it later," said Anna Pavlovna with a smile.
+
+And having got rid of this young man who did not know how to behave, she resumed her duties as hostess and continued to listen and watch, ready to help at any point where the conversation might happen to flag. As the foreman of a spinning mill, when he has set the hands to work, goes round and notices here a spindle that has stopped or there one that creaks or makes more noise than it should, and hastens to check the machine or set it in proper motion, so Anna Pavlovna moved about her drawing room, approaching now a silent, now a too-noisy group, and by a word or slight rearrangement kept the conversational machine in steady, proper, and regular motion. But amid these cares her anxiety about Pierre was evident. She kept an anxious watch on him when he approached the group round Mortemart to listen to what was being said there, and again when he passed to another group whose center was the abbe.
+
+Pierre had been educated abroad, and this reception at Anna Pavlovna's was the first he had attended in Russia. He knew that all the intellectual lights of Petersburg were gathered there and, like a child in a toyshop, did not know which way to look, afraid of missing any clever conversation that was to be heard. Seeing the self-confident and refined expression on the faces of those present he was always expecting to hear something very profound. At last he came up to Morio. Here the conversation seemed interesting and he stood waiting for an opportunity to express his own views, as young people are fond of doing.
+
+CHAPTER III
+Anna Pavlovna's reception was in full swing. The spindles hummed steadily and ceaselessly on all sides. With the exception of the aunt, beside whom sat only one elderly lady, who with her thin careworn face was rather out of place in this brilliant society, the whole company had settled into three groups. One, chiefly masculine, had formed round the abbe. Another, of young people, was grouped round the beautiful Princess Helene, Prince Vasili's daughter, and the little Princess Bolkonskaya, very pretty and rosy, though rather too plump for her age. The third group was gathered round Mortemart and Anna Pavlovna.
+
+The vicomte was a nice-looking young man with soft features and polished manners, who evidently considered himself a celebrity but out of politeness modestly placed himself at the disposal of the circle in which he found himself. Anna Pavlovna was obviously serving him up as a treat to her guests. As a clever maitre d'hotel serves up as a specially choice delicacy a piece of meat that no one who had seen it in the kitchen would have cared to eat, so Anna Pavlovna served up to her guests, first the vicomte and then the abbe, as peculiarly choice morsels. The group about Mortemart immediately began discussing the murder of the Duc d'Enghien. The vicomte said that the Duc d'Enghien had perished by his own magnanimity, and that there were particular reasons for Buonaparte's hatred of him.
+
+"Ah, yes! Do tell us all about it, Vicomte," said Anna Pavlovna, with a pleasant feeling that there was something A la Louis XV in the sound of that sentence: "Contez nous cela, Vicomte."
+
+The vicomte bowed and smiled courteously in token of his willingness to comply. Anna Pavlovna arranged a group round him, inviting everyone to listen to his tale.
+
+"The vicomte knew the duc personally," whispered Anna Pavlovna to one of the guests. "The vicomte is a wonderful raconteur," said she to another. "How evidently he belongs to the best society," said she to a third; and the vicomte was served up to the company in the choicest and most advantageous style, like a well-garnished joint of roast beef on a hot dish.
+
+The vicomte wished to begin his story and gave a subtle smile.
+
+"Come over here, Helene, dear," said Anna Pavlovna to the beautiful young princess who was sitting some way off, the center of another group.
+
+The princess smiled. She rose with the same unchanging smile with which she had first entered the room - the smile of a perfectly beautiful woman. With a slight rustle of her white dress trimmed with moss and ivy, with a gleam of white shoulders, glossy hair, and sparkling diamonds, she passed between the men who made way for her, not looking at any of them but smiling on all, as if graciously allowing each the privilege of admiring her beautiful figure and shapely shoulders, back, and bosom - which in the fashion of those days were very much exposed - and she seemed to bring the glamour of a ballroom with her as she moved toward Anna Pavlovna. Helene was so lovely that not only did she not show any trace of coquetry, but on the contrary she even appeared shy of her unquestionable and all too victorious beauty. She seemed to wish, but to be unable, to diminish its effect.
+
+"How lovely!" said everyone who saw her; and the vicomte lifted his shoulders and dropped his eyes as if startled by something extraordinary when she took her seat opposite and beamed upon him also with her unchanging smile.
+
+"Madame, I doubt my ability before such an audience," said he, smilingly inclining his head.
+
+The princess rested her bare round arm on a little table and considered a reply unnecessary. She smilingly waited. All the time the story was being told she sat upright, glancing now at her beautiful round arm, altered in shape by its pressure on the table, now at her still more beautiful bosom, on which she readjusted a diamond necklace. From time to time she smoothed the folds of her dress, and whenever the story produced an effect she glanced at Anna Pavlovna, at once adopted just the expression she saw on the maid of honor's face, and again relapsed into her radiant smile.
+
+The little princess had also left the tea table and followed Helene.
+
+"Wait a moment, I'll get my work.... Now then, what are you thinking of?" she went on, turning to Prince Hippolyte. "Fetch me my workbag."
+
+There was a general movement as the princess, smiling and talking merrily to everyone at once, sat down and gaily arranged herself in her seat.
+
+"Now I am all right," she said, and asking the vicomte to begin, she took up her work.
+
+Prince Hippolyte, having brought the workbag, joined the circle and moving a chair close to hers seated himself beside her.
+
+Le charmant Hippolyte was surprising by his extraordinary resemblance to his beautiful sister, but yet more by the fact that in spite of this resemblance he was exceedingly ugly. His features were like his sister's, but while in her case everything was lit up by a joyous, self-satisfied, youthful, and constant smile of animation, and by the wonderful classic beauty of her figure, his face on the contrary was dulled by imbecility and a constant expression of sullen self-confidence, while his body was thin and weak. His eyes, nose, and mouth all seemed puckered into a vacant, wearied grimace, and his arms and legs always fell into unnatural positions.
+
+"It's not going to be a ghost story?" said he, sitting down beside the princess and hastily adjusting his lorgnette, as if without this instrument he could not begin to speak.
+
+"Why no, my dear fellow," said the astonished narrator, shrugging his shoulders.
+
+"Because I hate ghost stories," said Prince Hippolyte in a tone which showed that he only understood the meaning of his words after he had uttered them.
+
+He spoke with such self-confidence that his hearers could not be sure whether what he said was very witty or very stupid. He was dressed in a dark-green dress coat, knee breeches of the color of cuisse de nymphe effrayee, as he called it, shoes, and silk stockings.
+
+The vicomte told his tale very neatly. It was an anecdote, then current, to the effect that the Duc d'Enghien had gone secretly to Paris to visit Mademoiselle George; that at her house he came upon Bonaparte, who also enjoyed the famous actress' favors, and that in his presence Napoleon happened to fall into one of the fainting fits to which he was subject, and was thus at the duc's mercy. The latter spared him, and this magnanimity Bonaparte subsequently repaid by death.
+
+The story was very pretty and interesting, especially at the point where the rivals suddenly recognized one another; and the ladies looked agitated.
+
+"Charming!" said Anna Pavlovna with an inquiring glance at the little princess.
+
+"Charming!" whispered the little princess, sticking the needle into her work as if to testify that the interest and fascination of the story prevented her from going on with it.
+
+The vicomte appreciated this silent praise and smiling gratefully prepared to continue, but just then Anna Pavlovna, who had kept a watchful eye on the young man who so alarmed her, noticed that he was talking too loudly and vehemently with the abbe, so she hurried to the rescue. Pierre had managed to start a conversation with the abbe about the balance of power, and the latter, evidently interested by the young man's simple-minded eagerness, was explaining his pet theory. Both were talking and listening too eagerly and too naturally, which was why Anna Pavlovna disapproved.
+
+"The means are ... the balance of power in Europe and the rights of the people," the abbe was saying. "It is only necessary for one powerful nation like Russia - barbaric as she is said to be - to place herself disinterestedly at the head of an alliance having for its object the maintenance of the balance of power of Europe, and it would save the world!"
+
+"But how are you to get that balance?" Pierre was beginning.
+
+At that moment Anna Pavlovna came up and, looking severely at Pierre, asked the Italian how he stood Russian climate. The Italian's face instantly changed and assumed an offensively affected, sugary expression, evidently habitual to him when conversing with women.
+
+"I am so enchanted by the brilliancy of the wit and culture of the society, more especially of the feminine society, in which I have had the honor of being received, that I have not yet had time to think of the climate," said he.
+
+Not letting the abbe and Pierre escape, Anna Pavlovna, the more conveniently to keep them under observation, brought them into the larger circle.
+

examples/typescript-simplechat/client.ts

@@ -0,0 +1,77 @@
+import * as readline from "readline";
+
+const model = "llama2";
+type Message = {
+  role: "assistant" | "user" | "system";
+  content: string;
+}
+const messages: Message[] = [{
+  role: "system",
+  content: "You are a helpful AI agent."
+}]
+
+const rl = readline.createInterface({
+  input: process.stdin,
+  output: process.stdout
+})
+
+async function chat(messages: Message[]): Promise<Message> {
+  const body = {
+    model: model,
+    messages: messages
+  }
+
+  const response = await fetch("http://localhost:11434/api/chat", {
+    method: "POST",
+    body: JSON.stringify(body)
+  })
+
+  const reader = response.body?.getReader()
+  if (!reader) {
+    throw new Error("Failed to read response body")
+  }
+  let content = ""
+  while (true) {
+    const { done, value } = await reader.read()
+    if (done) {
+      break;
+    }
+    const rawjson = new TextDecoder().decode(value);
+    const json = JSON.parse(rawjson)
+
+    if (json.done === false) {
+      process.stdout.write(json.message.content);
+      content += json.message.content
+    }
+
+  }
+  return { role: "assistant", content: content };
+}
+
+async function askQuestion(): Promise<void> {
+  return new Promise<void>((resolve) => {
+    rl.question("\n\nAsk a question: (press enter alone to quit)\n\n", async (user_input) => {
+      if (user_input.trim() === "") {
+        rl.close();
+        console.log("Thankyou. Goodbye.\n")
+        console.log("=======\nHere is the message history that was used in this conversation.\n=======\n")
+        messages.forEach(message => {
+          console.log(message)
+        })
+        resolve();
+      } else {
+        console.log();
+        messages.push({ role: "user", content: user_input });
+        messages.push(await chat(messages));
+        await askQuestion(); // Ask the next question
+      }
+    });
+  });
+}
+
+async function main() {
+  await askQuestion();
+
+}
+
+main();

examples/typescript-simplechat/package.json

@@ -0,0 +1 @@
+{ "dependencies": { "@types/node": "^20.10.4", "prompt-sync": "^4.2.0", "readline": "^1.3.0" } }

examples/typescript-simplechat/readme.md

@@ -0,0 +1,39 @@
+# Simple Chat Example
+
+The **chat** endpoint is one of two ways to generate text from an LLM with Ollama. At a high level you provide the endpoint an array of message objects with a role and content specified. Then with each output and prompt, you add more messages, which builds up the history.
+
+## Run the Example
+
+There are a few ways to run this, just like any Typescript code:
+
+1. Compile with `tsc` and then run it with `node client.js`.
+2. Install `tsx` and run it with `tsx client.ts`.
+3. Install `bun` and run it with `bun client.ts`.
+
+## Review the Code
+
+You can see in the **chat** function that is actually calling the endpoint is simply done with:
+
+```typescript
+const body = {
+  model: model,
+  messages: messages
+}
+
+const response = await fetch("http://localhost:11434/api/chat", {
+  method: "POST",
+  body: JSON.stringify(body)
+})
+```
+
+With the **generate** endpoint, you need to provide a `prompt`. But with **chat**, you provide `messages`. And the resulting stream of responses includes a `message` object with a `content` field.
+
+The final JSON object doesn't provide the full content, so you will need to build the content yourself. In this example, **chat** takes the full array of messages and outputs the resulting message from this call of the chat endpoint.
+
+In the **askQuestion** function, we collect `user_input` and add it as a message to our messages and that is passed to the chat function. When the LLM is done responding the output is added as another message to the messages array.
+
+At the end, you will see a printout of all the messages.
+
+## Next Steps
+
+In this example, all generations are kept. You might want to experiment with summarizing everything older than 10 conversations to enable longer history with less context being used.

format/bytes.go

@@ -1,23 +1,47 @@
 package format
 
-import "fmt"
+import (
+	"fmt"
+	"math"
+)
 
 const (
 	Byte     = 1
 	KiloByte = Byte * 1000
 	MegaByte = KiloByte * 1000
 	GigaByte = MegaByte * 1000
+	TeraByte = GigaByte * 1000
 )
 
 func HumanBytes(b int64) string {
+	var value float64
+	var unit string
+
 	switch {
-	case b > GigaByte:
-		return fmt.Sprintf("%.1f GB", float64(b)/GigaByte)
-	case b > MegaByte:
-		return fmt.Sprintf("%.1f MB", float64(b)/MegaByte)
-	case b > KiloByte:
-		return fmt.Sprintf("%.1f KB", float64(b)/KiloByte)
+	case b >= TeraByte:
+		value = float64(b) / TeraByte
+		unit = "TB"
+	case b >= GigaByte:
+		value = float64(b) / GigaByte
+		unit = "GB"
+	case b >= MegaByte:
+		value = float64(b) / MegaByte
+		unit = "MB"
+	case b >= KiloByte:
+		value = float64(b) / KiloByte
+		unit = "KB"
 	default:
 		return fmt.Sprintf("%d B", b)
 	}
+
+	switch {
+	case value >= 100:
+		return fmt.Sprintf("%d %s", int(value), unit)
+	case value >= 10:
+		return fmt.Sprintf("%d %s", int(value), unit)
+	case value != math.Trunc(value):
+		return fmt.Sprintf("%.1f %s", value, unit)
+	default:
+		return fmt.Sprintf("%d %s", int(value), unit)
+	}
 }

go.mod

@@ -5,14 +5,15 @@ go 1.20
 require (
 	github.com/emirpasic/gods v1.18.1
 	github.com/gin-gonic/gin v1.9.1
-	github.com/mattn/go-runewidth v0.0.14
-	github.com/mitchellh/colorstring v0.0.0-20190213212951-d06e56a500db
 	github.com/olekukonko/tablewriter v0.0.5
 	github.com/spf13/cobra v1.7.0
 	golang.org/x/sync v0.3.0
 )
 
-require github.com/rivo/uniseg v0.2.0 // indirect
+require (
+	github.com/mattn/go-runewidth v0.0.14 // indirect
+	github.com/rivo/uniseg v0.2.0 // indirect
+)
 
 require (
 	github.com/bytedance/sonic v1.9.1 // indirect

go.sum

@@ -63,8 +63,6 @@ github.com/mattn/go-isatty v0.0.19/go.mod h1:W+V8PltTTMOvKvAeJH7IuucS94S2C6jfK/D
 github.com/mattn/go-runewidth v0.0.9/go.mod h1:H031xJmbD/WCDINGzjvQ9THkh0rPKHF+m2gUSrubnMI=
 github.com/mattn/go-runewidth v0.0.14 h1:+xnbZSEeDbOIg5/mE6JF0w6n9duR1l3/WmbinWVwUuU=
 github.com/mattn/go-runewidth v0.0.14/go.mod h1:Jdepj2loyihRzMpdS35Xk/zdY8IAYHsh153qUoGf23w=
-github.com/mitchellh/colorstring v0.0.0-20190213212951-d06e56a500db h1:62I3jR2EmQ4l5rM/4FEfDWcRD+abF5XlKShorW5LRoQ=
-github.com/mitchellh/colorstring v0.0.0-20190213212951-d06e56a500db/go.mod h1:l0dey0ia/Uv7NcFFVbCLtqEBQbrT4OCwCSKTEv6enCw=
 github.com/modern-go/concurrent v0.0.0-20180228061459-e0a39a4cb421/go.mod h1:6dJC0mAP4ikYIbvyc7fijjWJddQyLn8Ig3JB5CqoB9Q=
 github.com/modern-go/concurrent v0.0.0-20180306012644-bacd9c7ef1dd h1:TRLaZ9cD/w8PVh93nsPXa1VrQ6jlwL5oN8l14QlcNfg=
 github.com/modern-go/concurrent v0.0.0-20180306012644-bacd9c7ef1dd/go.mod h1:6dJC0mAP4ikYIbvyc7fijjWJddQyLn8Ig3JB5CqoB9Q=

llm/falcon.go

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

llm/ggml.go

@@ -7,9 +7,10 @@ import (
 )
 
 type GGML struct {
-	magic uint32
 	container
 	model
+
+	Size int64
 }
 
 const (
@@ -82,7 +83,7 @@ type model interface {
 
 type container interface {
 	Name() string
-	Decode(io.Reader) (model, error)
+	Decode(*readSeekOffset) (model, error)
 }
 
 type containerGGML struct{}
@@ -91,7 +92,9 @@ func (c *containerGGML) Name() string {
 	return "ggml"
 }
 
-func (c *containerGGML) Decode(r io.Reader) (model, error) {
+func (c *containerGGML) Decode(ro *readSeekOffset) (model, error) {
+	// file contents aren't decoded
+	ro.Seek(0, io.SeekEnd)
 	return nil, nil
 }
 
@@ -103,9 +106,9 @@ func (c *containerGGMF) Name() string {
 	return "ggmf"
 }
 
-func (c *containerGGMF) Decode(r io.Reader) (model, error) {
+func (c *containerGGMF) Decode(ro *readSeekOffset) (model, error) {
 	var version uint32
-	binary.Read(r, binary.LittleEndian, &version)
+	binary.Read(ro, binary.LittleEndian, &version)
 
 	switch version {
 	case 1:
@@ -114,6 +117,10 @@ func (c *containerGGMF) Decode(r io.Reader) (model, error) {
 	}
 
 	c.version = version
+
+	// remaining file contents aren't decoded
+	ro.Seek(0, io.SeekEnd)
+
 	return nil, nil
 }
 
@@ -125,9 +132,9 @@ func (c *containerGGJT) Name() string {
 	return "ggjt"
 }
 
-func (c *containerGGJT) Decode(r io.Reader) (model, error) {
+func (c *containerGGJT) Decode(ro *readSeekOffset) (model, error) {
 	var version uint32
-	binary.Read(r, binary.LittleEndian, &version)
+	binary.Read(ro, binary.LittleEndian, &version)
 
 	switch version {
 	case 1, 2, 3:
@@ -139,7 +146,11 @@ func (c *containerGGJT) Decode(r io.Reader) (model, error) {
 
 	// different model types may have different layouts for hyperparameters
 	var llama llamaModel
-	binary.Read(r, binary.LittleEndian, &llama.hyperparameters)
+	binary.Read(ro, binary.LittleEndian, &llama.hyperparameters)
+
+	// remaining file contents aren't decoded
+	ro.Seek(0, io.SeekEnd)
+
 	return &llama, nil
 }
 
@@ -151,9 +162,9 @@ func (c *containerLORA) Name() string {
 	return "ggla"
 }
 
-func (c *containerLORA) Decode(r io.Reader) (model, error) {
+func (c *containerLORA) Decode(ro *readSeekOffset) (model, error) {
 	var version uint32
-	binary.Read(r, binary.LittleEndian, &version)
+	binary.Read(ro, binary.LittleEndian, &version)
 
 	switch version {
 	case 1:
@@ -162,6 +173,10 @@ func (c *containerLORA) Decode(r io.Reader) (model, error) {
 	}
 
 	c.version = version
+
+	// remaining file contents aren't decoded
+	ro.Seek(0, io.SeekEnd)
+
 	return nil, nil
 }
 
@@ -180,33 +195,61 @@ const (
 )
 
 func DecodeGGML(r io.ReadSeeker) (*GGML, error) {
-	var ggml GGML
-	binary.Read(r, binary.LittleEndian, &ggml.magic)
+	ro := readSeekOffset{ReadSeeker: r}
 
-	switch ggml.magic {
+	var magic uint32
+	if err := binary.Read(&ro, binary.LittleEndian, &magic); err != nil {
+		return nil, err
+	}
+
+	var c container
+	switch magic {
 	case FILE_MAGIC_GGML:
-		ggml.container = &containerGGML{}
+		c = &containerGGML{}
 	case FILE_MAGIC_GGMF:
-		ggml.container = &containerGGMF{}
+		c = &containerGGMF{}
 	case FILE_MAGIC_GGJT:
-		ggml.container = &containerGGJT{}
+		c = &containerGGJT{}
 	case FILE_MAGIC_GGLA:
-		ggml.container = &containerLORA{}
+		c = &containerLORA{}
 	case FILE_MAGIC_GGUF_LE:
-		ggml.container = &containerGGUF{bo: binary.LittleEndian}
+		c = &containerGGUF{bo: binary.LittleEndian}
 	case FILE_MAGIC_GGUF_BE:
-		ggml.container = &containerGGUF{bo: binary.BigEndian}
+		c = &containerGGUF{bo: binary.BigEndian}
 	default:
 		return nil, errors.New("invalid file magic")
 	}
 
-	model, err := ggml.Decode(r)
+	model, err := c.Decode(&ro)
 	if err != nil {
 		return nil, err
 	}
 
-	ggml.model = model
-
 	// final model type
-	return &ggml, nil
+	return &GGML{
+		container: c,
+		model:     model,
+		Size:      ro.offset,
+	}, nil
+}
+
+type readSeekOffset struct {
+	io.ReadSeeker
+	offset int64
+}
+
+func (rso *readSeekOffset) Seek(offset int64, whence int) (int64, error) {
+	offset, err := rso.ReadSeeker.Seek(offset, whence)
+	if err != nil {
+		return 0, err
+	}
+
+	rso.offset = offset
+	return offset, nil
+}
+
+func (rso *readSeekOffset) Read(p []byte) (int, error) {
+	n, err := rso.ReadSeeker.Read(p)
+	rso.offset += int64(n)
+	return n, err
 }

llm/gguf.go

@@ -23,26 +23,24 @@ type containerGGUF struct {
 		NumTensor uint64
 		NumKV     uint64
 	}
-
-	parameters uint64
 }
 
 func (c *containerGGUF) Name() string {
 	return "gguf"
 }
 
-func (c *containerGGUF) Decode(r io.Reader) (model, error) {
-	binary.Read(r, c.bo, &c.Version)
+func (c *containerGGUF) Decode(rso *readSeekOffset) (model, error) {
+	binary.Read(rso, c.bo, &c.Version)
 
 	switch c.Version {
 	case 1:
-		binary.Read(r, c.bo, &c.V1)
+		binary.Read(rso, c.bo, &c.V1)
 	default:
-		binary.Read(r, c.bo, &c.V2)
+		binary.Read(rso, c.bo, &c.V2)
 	}
 
 	model := newGGUFModel(c)
-	if err := model.Decode(r); err != nil {
+	if err := model.Decode(rso); err != nil {
 		return nil, err
 	}
 
@@ -67,9 +65,23 @@ const (
 
 type kv map[string]any
 
+type tensor struct {
+	name   string
+	kind   uint32
+	offset uint64
+	size   uint64
+
+	// shape is the number of elements in each dimension
+	shape [4]uint64
+}
+
 type ggufModel struct {
 	*containerGGUF
+
 	kv
+	tensors []tensor
+
+	parameters uint64
 }
 
 func newGGUFModel(container *containerGGUF) *ggufModel {
@@ -96,8 +108,7 @@ func (llm *ggufModel) NumKV() uint64 {
 }
 
 func (llm *ggufModel) ModelFamily() string {
-	t, ok := llm.kv["general.architecture"].(string)
-	if ok {
+	if t, ok := llm.kv["general.architecture"].(string); ok {
 		return t
 	}
 
@@ -109,82 +120,60 @@ func (llm *ggufModel) ModelType() string {
 		return format.HumanNumber(llm.parameters)
 	}
 
-	switch llm.ModelFamily() {
-	case "llama":
-		if blocks, ok := llm.kv["llama.block_count"].(uint32); ok {
-			heads, headsOK := llm.kv["llama.head_count"].(uint32)
-			headKVs, headsKVsOK := llm.kv["llama.head_count_kv"].(uint32)
-			if headsOK && headsKVsOK && heads/headKVs == 8 {
-				return "70B"
-			}
-
-			return llamaModelType(blocks)
-		}
-	case "falcon":
-		if blocks, ok := llm.kv["falcon.block_count"].(uint32); ok {
-			return falconModelType(blocks)
-		}
-	case "starcoder":
-		if blocks, ok := llm.kv["starcoder.block_count"].(uint32); ok {
-			return starCoderModelType(blocks)
-		}
-	}
-
 	return "unknown"
 }
 
 func (llm *ggufModel) FileType() string {
-	t, ok := llm.kv["general.file_type"].(uint32)
-	if ok {
+	if t, ok := llm.kv["general.file_type"].(uint32); ok {
 		return fileType(t)
 	}
 
 	return "unknown"
 }
 
-func (llm *ggufModel) Decode(r io.Reader) error {
+func (llm *ggufModel) Decode(rso *readSeekOffset) error {
 	// decode key-values
 	for i := 0; uint64(i) < llm.NumKV(); i++ {
-		k, err := llm.readString(r)
+		k, err := llm.readString(rso)
 		if err != nil {
 			return err
 		}
 
-		vtype := llm.readU32(r)
+		vtype := llm.readU32(rso)
 
 		var v any
 		switch vtype {
 		case ggufTypeUint8:
-			v = llm.readU8(r)
+			v = llm.readU8(rso)
 		case ggufTypeInt8:
-			v = llm.readI8(r)
+			v = llm.readI8(rso)
 		case ggufTypeUint16:
-			v = llm.readU16(r)
+			v = llm.readU16(rso)
 		case ggufTypeInt16:
-			v = llm.readI16(r)
+			v = llm.readI16(rso)
 		case ggufTypeUint32:
-			v = llm.readU32(r)
+			v = llm.readU32(rso)
 		case ggufTypeInt32:
-			v = llm.readI32(r)
+			v = llm.readI32(rso)
 		case ggufTypeUint64:
-			v = llm.readU64(r)
+			v = llm.readU64(rso)
 		case ggufTypeInt64:
-			v = llm.readI64(r)
+			v = llm.readI64(rso)
 		case ggufTypeFloat32:
-			v = llm.readF32(r)
+			v = llm.readF32(rso)
 		case ggufTypeFloat64:
-			v = llm.readF64(r)
+			v = llm.readF64(rso)
 		case ggufTypeBool:
-			v = llm.readBool(r)
+			v = llm.readBool(rso)
 		case ggufTypeString:
-			s, err := llm.readString(r)
+			s, err := llm.readString(rso)
 			if err != nil {
 				return err
 			}
 
 			v = s
 		case ggufTypeArray:
-			a, err := llm.readArray(r)
+			a, err := llm.readArray(rso)
 			if err != nil {
 				return err
 			}
@@ -199,21 +188,85 @@ func (llm *ggufModel) Decode(r io.Reader) error {
 
 	// decode tensors
 	for i := 0; uint64(i) < llm.NumTensor(); i++ {
-		if _, err := llm.readString(r); err != nil {
+		name, err := llm.readString(rso)
+		if err != nil {
 			return err
 		}
 
-		dimensions := llm.readU32(r)
+		// dims is the number of dimensions in the tensor
+		dims := llm.readU32(rso)
 
-		var elements uint64 = 1
-		for i := 0; uint32(i) < dimensions; i++ {
-			elements *= llm.readU64(r)
+		shape := [4]uint64{1, 1, 1, 1}
+		for i := 0; uint32(i) < dims; i++ {
+			shape[i] = llm.readU64(rso)
 		}
 
-		llm.readU32(r) // type
-		llm.readU64(r) // offset
+		kind := llm.readU32(rso)
+		offset := llm.readU64(rso)
 
-		llm.parameters += elements
+		var blockSize uint64
+		switch {
+		case kind < 2:
+			blockSize = 1
+		case kind < 10:
+			blockSize = 32
+		default:
+			blockSize = 256
+		}
+
+		var typeSize uint64
+		switch kind {
+		case 0: // FP32
+			typeSize = 4
+		case 1: // FP16
+			typeSize = 2
+		case 2: // Q4_0
+			typeSize = 2 + blockSize/2
+		case 3: // Q4_1
+			typeSize = 2 + 2 + blockSize/2
+		case 6: // Q5_0
+			typeSize = 2 + 4 + blockSize/2
+		case 7: // Q5_1
+			typeSize = 2 + 2 + 4 + blockSize/2
+		case 8: // Q8_0
+			typeSize = 2 + blockSize
+		case 9: // Q8_1
+			typeSize = 4 + 4 + blockSize
+		case 10: // Q2_K
+			typeSize = blockSize/16 + blockSize/4 + 2 + 2
+		case 11: // Q3_K
+			typeSize = blockSize/8 + blockSize/4 + 12 + 2
+		case 12: // Q4_K
+			typeSize = 2 + 2 + 12 + blockSize/2
+		case 13: // Q5_K
+			typeSize = 2 + 2 + 12 + blockSize/8 + blockSize/2
+		case 14: // Q6_K
+			typeSize = blockSize/2 + blockSize/4 + blockSize/16 + 2
+		}
+
+		parameters := shape[0] * shape[1] * shape[2] * shape[3]
+		size := parameters * typeSize / blockSize
+
+		llm.tensors = append(llm.tensors, tensor{
+			name:   name,
+			kind:   kind,
+			offset: offset,
+			size:   size,
+			shape:  shape,
+		})
+
+		llm.parameters += parameters
+	}
+
+	alignment, ok := llm.kv["general.alignment"].(uint32)
+	if !ok {
+		alignment = 32
+	}
+
+	rso.Seek(int64(alignment)-rso.offset%int64(alignment), io.SeekCurrent)
+	for _, tensor := range llm.tensors {
+		padded := (int64(tensor.size) + int64(alignment) - 1) & ^(int64(alignment) - 1)
+		rso.Seek(padded, io.SeekCurrent)
 	}
 
 	return nil
@@ -335,7 +388,7 @@ func (llm *ggufModel) readArrayV1(r io.Reader) (arr []any, err error) {
 		case ggufTypeUint8:
 			arr = append(arr, llm.readU8(r))
 		case ggufTypeInt8:
-			arr = append(arr, llm.readU8(r))
+			arr = append(arr, llm.readI8(r))
 		case ggufTypeUint16:
 			arr = append(arr, llm.readU16(r))
 		case ggufTypeInt16:
@@ -376,7 +429,7 @@ func (llm *ggufModel) readArray(r io.Reader) (arr []any, err error) {
 		case ggufTypeUint8:
 			arr = append(arr, llm.readU8(r))
 		case ggufTypeInt8:
-			arr = append(arr, llm.readU8(r))
+			arr = append(arr, llm.readI8(r))
 		case ggufTypeUint16:
 			arr = append(arr, llm.readU16(r))
 		case ggufTypeInt16:

llm/llama.cpp/generate_darwin_amd64.go

@@ -7,13 +7,12 @@ package llm
 //go:generate git -C ggml apply ../patches/0002-34B-model-support.patch
 //go:generate git -C ggml apply ../patches/0003-metal-fix-synchronization-in-new-matrix-multiplicati.patch
 //go:generate git -C ggml apply ../patches/0004-metal-add-missing-barriers-for-mul-mat-2699.patch
-//go:generate cmake -S ggml -B ggml/build/cpu -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=x86_64 -DCMAKE_OSX_ARCHITECTURES=x86_64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
+//go:generate cmake -S ggml -B ggml/build/cpu -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_NAME=Darwin -DCMAKE_SYSTEM_PROCESSOR=x86_64 -DCMAKE_OSX_ARCHITECTURES=x86_64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
 //go:generate cmake --build ggml/build/cpu --target server --config Release
 //go:generate mv ggml/build/cpu/bin/server ggml/build/cpu/bin/ollama-runner
 
 //go:generate git submodule update --force gguf
 //go:generate git -C gguf apply ../patches/0001-update-default-log-target.patch
-//go:generate git -C gguf apply ../patches/0001-metal-handle-ggml_scale-for-n-4-0-close-3754.patch
-//go:generate cmake -S gguf -B gguf/build/cpu -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=x86_64 -DCMAKE_OSX_ARCHITECTURES=x86_64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
+//go:generate cmake -S gguf -B gguf/build/cpu -DLLAMA_METAL=off -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_NAME=Darwin -DCMAKE_SYSTEM_PROCESSOR=x86_64 -DCMAKE_OSX_ARCHITECTURES=x86_64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0 -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=on
 //go:generate cmake --build gguf/build/cpu --target server --config Release
 //go:generate mv gguf/build/cpu/bin/server gguf/build/cpu/bin/ollama-runner

llm/llama.cpp/generate_darwin_arm64.go

@@ -13,7 +13,6 @@ package llm
 
 //go:generate git submodule update --force gguf
 //go:generate git -C gguf apply ../patches/0001-update-default-log-target.patch
-//go:generate git -C gguf apply ../patches/0001-metal-handle-ggml_scale-for-n-4-0-close-3754.patch
 //go:generate cmake -S gguf -B gguf/build/metal -DLLAMA_METAL=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=arm64 -DCMAKE_OSX_ARCHITECTURES=arm64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
 //go:generate cmake --build gguf/build/metal --target server --config Release
 //go:generate mv gguf/build/metal/bin/server gguf/build/metal/bin/ollama-runner

llm/llama.cpp/generate_linux.go

@@ -21,6 +21,6 @@ package llm
 //go:generate cmake -S ggml -B ggml/build/cuda -DLLAMA_CUBLAS=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on
 //go:generate cmake --build ggml/build/cuda --target server --config Release
 //go:generate mv ggml/build/cuda/bin/server ggml/build/cuda/bin/ollama-runner
-//go:generate cmake -S gguf -B gguf/build/cuda -DLLAMA_CUBLAS=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off
+//go:generate cmake -S gguf -B gguf/build/cuda -DLLAMA_CUBLAS=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off -DLLAMA_CUDA_PEER_MAX_BATCH_SIZE=0
 //go:generate cmake --build gguf/build/cuda --target server --config Release
 //go:generate mv gguf/build/cuda/bin/server gguf/build/cuda/bin/ollama-runner

llm/llama.cpp/generate_windows.go

@@ -14,3 +14,11 @@ package llm
 //go:generate cmake -S gguf -B gguf/build/cpu -DLLAMA_K_QUANTS=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off
 //go:generate cmake --build gguf/build/cpu --target server --config Release
 //go:generate cmd /c move gguf\build\cpu\bin\Release\server.exe gguf\build\cpu\bin\Release\ollama-runner.exe
+
+//go:generate cmake -S ggml -B ggml/build/cuda -DLLAMA_CUBLAS=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on
+//go:generate cmake --build ggml/build/cuda --target server --config Release
+//go:generate cmd /c move ggml\build\cuda\bin\Release\server.exe ggml\build\cuda\bin\Release\ollama-runner.exe
+
+//go:generate cmake -S gguf -B gguf/build/cuda -DLLAMA_CUBLAS=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off
+//go:generate cmake --build gguf/build/cuda --target server --config Release
+//go:generate cmd /c move gguf\build\cuda\bin\Release\server.exe gguf\build\cuda\bin\Release\ollama-runner.exe

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

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

llm/llama.go

@@ -59,6 +59,7 @@ ws ::= ([ \t\n] ws)?
 var llamaCppEmbed embed.FS
 
 type ModelRunner struct {
+	Type        string // "gguf" or "ggml"
 	Path        string // path to the model runner executable
 	Accelerated bool
 }
@@ -71,24 +72,26 @@ func chooseRunners(workDir, runnerType string) []ModelRunner {
 	// IMPORTANT: the order of the runners in the array is the priority order
 	switch runtime.GOOS {
 	case "darwin":
-		runners = []ModelRunner{
-			{Path: path.Join(buildPath, "metal", "bin", "ollama-runner")},
-			{Path: path.Join(buildPath, "cpu", "bin", "ollama-runner")},
+		if runtime.GOARCH == "arm64" {
+			runners = []ModelRunner{{Type: runnerType, Path: path.Join(buildPath, "metal", "bin", "ollama-runner")}}
+		} else {
+			runners = []ModelRunner{{Type: runnerType, Path: path.Join(buildPath, "cpu", "bin", "ollama-runner")}}
 		}
 	case "linux":
 		runners = []ModelRunner{
-			{Path: path.Join(buildPath, "cuda", "bin", "ollama-runner"), Accelerated: true},
-			{Path: path.Join(buildPath, "cpu", "bin", "ollama-runner")},
+			{Type: runnerType, Path: path.Join(buildPath, "cuda", "bin", "ollama-runner"), Accelerated: true},
+			{Type: runnerType, Path: path.Join(buildPath, "cpu", "bin", "ollama-runner")},
 		}
 	case "windows":
 		// TODO: select windows GPU runner here when available
 		runners = []ModelRunner{
-			{Path: path.Join(buildPath, "cpu", "bin", "Release", "ollama-runner.exe")},
+			{Type: runnerType, Path: path.Join(buildPath, "cuda", "bin", "Release", "ollama-runner.exe"), Accelerated: true},
+			{Type: runnerType, Path: path.Join(buildPath, "cpu", "bin", "Release", "ollama-runner.exe")},
 		}
 	default:
 		log.Printf("unknown OS, running on CPU: %s", runtime.GOOS)
 		runners = []ModelRunner{
-			{Path: path.Join(buildPath, "cpu", "bin", "ollama-runner")},
+			{Type: runnerType, Path: path.Join(buildPath, "cpu", "bin", "ollama-runner")},
 		}
 	}
 
@@ -146,6 +149,7 @@ func chooseRunners(workDir, runnerType string) []ModelRunner {
 	for _, r := range runners {
 		// clean the ModelRunner paths so that they match the OS we are running on
 		localRunnersByPriority = append(localRunnersByPriority, ModelRunner{
+			Type:        r.Type,
 			Path:        filepath.Clean(path.Join(workDir, r.Path)),
 			Accelerated: r.Accelerated,
 		})
@@ -219,13 +223,19 @@ type Running struct {
 	*StatusWriter            // captures error messages from the llama runner process
 }
 
+type ImageData struct {
+	Data []byte `json:"data"`
+	ID   int    `json:"id"`
+}
+
 type llama struct {
 	api.Options
+	ImageData []ImageData
 	Running
 }
 
 var (
-	errNvidiaSMI     = errors.New("nvidia-smi command failed")
+	errNvidiaSMI     = errors.New("warning: gpu support may not be enabled, check that you have installed GPU drivers: nvidia-smi command failed")
 	errAvailableVRAM = errors.New("not enough VRAM available, falling back to CPU only")
 )
 
@@ -268,7 +278,7 @@ func NumGPU(numLayer, fileSizeBytes int64, opts api.Options) int {
 	if opts.NumGPU != -1 {
 		return opts.NumGPU
 	}
-	if runtime.GOOS == "linux" {
+	if runtime.GOOS == "linux" || runtime.GOOS == "windows" {
 		freeBytes, err := CheckVRAM()
 		if err != nil {
 			if !errors.Is(err, errNvidiaSMI) {
@@ -323,7 +333,7 @@ func (w *StatusWriter) Write(b []byte) (int, error) {
 	return os.Stderr.Write(b)
 }
 
-func newLlama(model string, adapters []string, runners []ModelRunner, numLayers int64, opts api.Options) (*llama, error) {
+func newLlama(model string, adapters, projectors []string, runners []ModelRunner, numLayers int64, opts api.Options) (*llama, error) {
 	fileInfo, err := os.Stat(model)
 	if err != nil {
 		return nil, err
@@ -342,6 +352,10 @@ func newLlama(model string, adapters []string, runners []ModelRunner, numLayers
 		"--embedding",
 	}
 
+	if opts.MainGPU > 0 {
+		params = append(params, "--main-gpu", fmt.Sprintf("%d", opts.MainGPU))
+	}
+
 	if opts.RopeFrequencyBase > 0 {
 		params = append(params, "--rope-freq-base", fmt.Sprintf("%f", opts.RopeFrequencyBase))
 	}
@@ -359,6 +373,11 @@ func newLlama(model string, adapters []string, runners []ModelRunner, numLayers
 		params = append(params, "--lora", adapters[0])
 	}
 
+	if len(projectors) > 0 {
+		// TODO: applying multiple projectors is not supported by the llama.cpp server yet
+		params = append(params, "--mmproj", projectors[0])
+	}
+
 	if opts.NumThread > 0 {
 		params = append(params, "--threads", fmt.Sprintf("%d", opts.NumThread))
 	}
@@ -391,11 +410,13 @@ func newLlama(model string, adapters []string, runners []ModelRunner, numLayers
 		}
 
 		port := rand.Intn(65535-49152) + 49152 // get a random port in the ephemeral range
+		params := append(params, "--port", strconv.Itoa(port))
+
 		ctx, cancel := context.WithCancel(context.Background())
 		cmd := exec.CommandContext(
 			ctx,
 			runner.Path,
-			append(params, "--port", strconv.Itoa(port))...,
+			params...,
 		)
 
 		var libraryPaths []string
@@ -524,25 +545,43 @@ type prediction struct {
 }
 
 const maxBufferSize = 512 * format.KiloByte
+const maxRetries = 6
 
-func (llm *llama) Predict(ctx context.Context, prevContext []int, prompt string, format string, fn func(api.GenerateResponse)) error {
-	prevConvo, err := llm.Decode(ctx, prevContext)
-	if err != nil {
-		return err
+type PredictOpts struct {
+	Prompt string
+	Format string
+	Images []api.ImageData
+}
+
+type PredictResult struct {
+	Content            string
+	Done               bool
+	PromptEvalCount    int
+	PromptEvalDuration time.Duration
+	EvalCount          int
+	EvalDuration       time.Duration
+}
+
+// IsRetryable checks if the line matches a condition that can be retried
+func isRetryable(line []byte) bool {
+	return bytes.Contains(line, []byte("slot unavailable"))
+}
+
+func (llm *llama) Predict(ctx context.Context, predict PredictOpts, fn func(PredictResult)) error {
+	imageData := llm.ImageData
+	if len(predict.Images) > 0 {
+		for cnt, i := range predict.Images {
+			imageData = append(imageData, ImageData{Data: i, ID: cnt})
+		}
 	}
-
-	// Remove leading spaces from prevConvo if present
-	prevConvo = strings.TrimPrefix(prevConvo, " ")
-
-	var nextContext strings.Builder
-	nextContext.WriteString(prevConvo)
-	nextContext.WriteString(prompt)
+	log.Printf("loaded %d images", len(imageData))
 
 	request := map[string]any{
-		"prompt":            nextContext.String(),
+		"prompt":            predict.Prompt,
 		"stream":            true,
 		"n_predict":         llm.NumPredict,
 		"n_keep":            llm.NumKeep,
+		"main_gpu":          llm.MainGPU,
 		"temperature":       llm.Temperature,
 		"top_k":             llm.TopK,
 		"top_p":             llm.TopP,
@@ -558,103 +597,121 @@ func (llm *llama) Predict(ctx context.Context, prevContext []int, prompt string,
 		"penalize_nl":       llm.PenalizeNewline,
 		"seed":              llm.Seed,
 		"stop":              llm.Stop,
+		"image_data":        imageData,
 	}
 
-	if format == "json" {
+	if predict.Format == "json" {
 		request["grammar"] = jsonGrammar
 	}
 
-	// Handling JSON marshaling with special characters unescaped.
-	buffer := &bytes.Buffer{}
-	enc := json.NewEncoder(buffer)
-	enc.SetEscapeHTML(false)
-
-	if err := enc.Encode(request); err != nil {
-		return fmt.Errorf("failed to marshal data: %v", err)
-	}
-
-	endpoint := fmt.Sprintf("http://127.0.0.1:%d/completion", llm.Port)
-	req, err := http.NewRequestWithContext(ctx, http.MethodPost, endpoint, buffer)
-	if err != nil {
-		return fmt.Errorf("error creating POST request: %v", err)
-	}
-	req.Header.Set("Content-Type", "application/json")
-
-	resp, err := http.DefaultClient.Do(req)
-	if err != nil {
-		return fmt.Errorf("POST predict: %v", err)
-	}
-	defer resp.Body.Close()
-
-	if resp.StatusCode >= 400 {
-		bodyBytes, err := io.ReadAll(resp.Body)
-		if err != nil {
-			return fmt.Errorf("failed reading llm error response: %w", err)
+	retryDelay := 100 * time.Microsecond
+	for retries := 0; retries < maxRetries; retries++ {
+		if retries > 0 {
+			time.Sleep(retryDelay) // wait before retrying
+			retryDelay *= 2        // exponential backoff
 		}
-		log.Printf("llm predict error: %s", bodyBytes)
-		return fmt.Errorf("%s", bodyBytes)
-	}
 
-	scanner := bufio.NewScanner(resp.Body)
-	// increase the buffer size to avoid running out of space
-	buf := make([]byte, 0, maxBufferSize)
-	scanner.Buffer(buf, maxBufferSize)
-	for scanner.Scan() {
-		select {
-		case <-ctx.Done():
-			// This handles the request cancellation
-			return ctx.Err()
-		default:
-			line := scanner.Bytes()
-			if len(line) == 0 {
-				continue
+		// Handling JSON marshaling with special characters unescaped.
+		buffer := &bytes.Buffer{}
+		enc := json.NewEncoder(buffer)
+		enc.SetEscapeHTML(false)
+
+		if err := enc.Encode(request); err != nil {
+			return fmt.Errorf("failed to marshal data: %v", err)
+		}
+
+		endpoint := fmt.Sprintf("http://127.0.0.1:%d/completion", llm.Port)
+		req, err := http.NewRequestWithContext(ctx, http.MethodPost, endpoint, buffer)
+		if err != nil {
+			return fmt.Errorf("error creating POST request: %v", err)
+		}
+		req.Header.Set("Content-Type", "application/json")
+
+		resp, err := http.DefaultClient.Do(req)
+		if err != nil {
+			return fmt.Errorf("POST predict: %v", err)
+		}
+		defer resp.Body.Close()
+
+		if resp.StatusCode >= 400 {
+			bodyBytes, err := io.ReadAll(resp.Body)
+			if err != nil {
+				return fmt.Errorf("failed reading llm error response: %w", err)
 			}
+			log.Printf("llm predict error: %s", bodyBytes)
+			return fmt.Errorf("%s", bodyBytes)
+		}
+
+		scanner := bufio.NewScanner(resp.Body)
+		// increase the buffer size to avoid running out of space
+		buf := make([]byte, 0, maxBufferSize)
+		scanner.Buffer(buf, maxBufferSize)
+
+		retryNeeded := false
+		for scanner.Scan() {
+			select {
+			case <-ctx.Done():
+				// This handles the request cancellation
+				return ctx.Err()
+			default:
+				line := scanner.Bytes()
+				if len(line) == 0 {
+					continue
+				}
+
+				if isRetryable(line) {
+					retryNeeded = true
+					break
+				}
+
+				evt, ok := bytes.CutPrefix(line, []byte("data: "))
+				if !ok {
+					return fmt.Errorf("error parsing llm response stream: %s", line)
+				}
 
-			if evt, ok := bytes.CutPrefix(line, []byte("data: ")); ok {
 				var p prediction
 				if err := json.Unmarshal(evt, &p); err != nil {
 					return fmt.Errorf("error unmarshaling llm prediction response: %v", err)
 				}
 
 				if p.Content != "" {
-					fn(api.GenerateResponse{Response: p.Content})
-					nextContext.WriteString(p.Content)
+					fn(PredictResult{
+						Content: 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{
+					fn(PredictResult{
 						Done:               true,
-						Context:            embd,
 						PromptEvalCount:    p.Timings.PromptN,
 						PromptEvalDuration: parseDurationMs(p.Timings.PromptMS),
 						EvalCount:          p.Timings.PredictedN,
 						EvalDuration:       parseDurationMs(p.Timings.PredictedMS),
 					})
-
 					return nil
 				}
 			}
 		}
-	}
 
-	if err := scanner.Err(); err != nil {
-		if strings.Contains(err.Error(), "unexpected EOF") {
-			// this means the llama runner subprocess crashed
-			llm.Close()
-			if llm.StatusWriter != nil && llm.StatusWriter.LastErrMsg != "" {
-				return fmt.Errorf("llama runner exited: %v", llm.StatusWriter.LastErrMsg)
+		if err := scanner.Err(); err != nil {
+			if strings.Contains(err.Error(), "unexpected EOF") {
+				// this means the llama runner subprocess crashed
+				llm.Close()
+				if llm.StatusWriter != nil && llm.StatusWriter.LastErrMsg != "" {
+					return fmt.Errorf("llama runner exited: %v", llm.StatusWriter.LastErrMsg)
+				}
+				return fmt.Errorf("llama runner exited, you may not have enough available memory to run this model")
 			}
-			return fmt.Errorf("llama runner exited, you may not have enough available memory to run this model")
+			return fmt.Errorf("error reading llm response: %v", err)
+		}
+
+		if !retryNeeded {
+			return nil // success
 		}
-		return fmt.Errorf("error reading llm response: %v", err)
 	}
 
-	return nil
+	// should never reach here ideally
+	return fmt.Errorf("max retries exceeded")
 }
 
 type TokenizeRequest struct {

llm/llm.go

@@ -14,7 +14,7 @@ import (
 )
 
 type LLM interface {
-	Predict(context.Context, []int, string, string, func(api.GenerateResponse)) error
+	Predict(context.Context, PredictOpts, func(PredictResult)) error
 	Embedding(context.Context, string) ([]float64, error)
 	Encode(context.Context, string) ([]int, error)
 	Decode(context.Context, []int) (string, error)
@@ -23,7 +23,7 @@ type LLM interface {
 	Ping(context.Context) error
 }
 
-func New(workDir, model string, adapters []string, opts api.Options) (LLM, error) {
+func New(workDir, model string, adapters, projectors []string, opts api.Options) (LLM, error) {
 	if _, err := os.Stat(model); err != nil {
 		return nil, err
 	}
@@ -41,20 +41,13 @@ func New(workDir, model string, adapters []string, opts api.Options) (LLM, error
 
 	if runtime.GOOS == "darwin" {
 		switch ggml.FileType() {
-		case "Q8_0":
+		case "F32", "Q5_0", "Q5_1", "Q8_0":
 			if ggml.Name() != "gguf" && opts.NumGPU != 0 {
 				// GGML Q8_0 do not support Metal API and will
 				// cause the runner to segmentation fault so disable GPU
 				log.Printf("WARNING: GPU disabled for F32, Q5_0, Q5_1, and Q8_0")
 				opts.NumGPU = 0
 			}
-		case "F32", "Q5_0", "Q5_1":
-			if opts.NumGPU != 0 {
-				// F32, Q5_0, Q5_1, and Q8_0 do not support Metal API and will
-				// cause the runner to segmentation fault so disable GPU
-				log.Printf("WARNING: GPU disabled for F32, Q5_0, Q5_1, and Q8_0")
-				opts.NumGPU = 0
-			}
 		}
 
 		var requiredMemory int64
@@ -89,9 +82,9 @@ func New(workDir, model string, adapters []string, opts api.Options) (LLM, error
 		opts.NumGQA = 0
 		opts.RopeFrequencyBase = 0.0
 		opts.RopeFrequencyScale = 0.0
-		return newLlama(model, adapters, chooseRunners(workDir, "gguf"), ggml.NumLayers(), opts)
+		return newLlama(model, adapters, projectors, chooseRunners(workDir, "gguf"), ggml.NumLayers(), opts)
 	case "ggml", "ggmf", "ggjt", "ggla":
-		return newLlama(model, adapters, chooseRunners(workDir, "ggml"), ggml.NumLayers(), opts)
+		return newLlama(model, adapters, projectors, chooseRunners(workDir, "ggml"), ggml.NumLayers(), opts)
 	default:
 		return nil, fmt.Errorf("unknown ggml type: %s", ggml.ModelFamily())
 	}

llm/starcoder.go

@@ -1,23 +0,0 @@
-package llm
-
-const (
-	starCoderModelType1B  = 24
-	starCoderModelType3B  = 36
-	starCoderModelType7B  = 42
-	starCoderModelType15B = 40
-)
-
-func starCoderModelType(numLayer uint32) string {
-	switch numLayer {
-	case 24:
-		return "1B"
-	case 36:
-		return "3B"
-	case 42:
-		return "7B"
-	case 40:
-		return "15B"
-	default:
-		return "unknown"
-	}
-}

parser/parser.go

@@ -37,10 +37,13 @@ func Parse(reader io.Reader) ([]Command, error) {
 		switch string(bytes.ToUpper(fields[0])) {
 		case "FROM":
 			command.Name = "model"
-			command.Args = string(fields[1])
+			command.Args = string(bytes.TrimSpace(fields[1]))
 			// copy command for validation
 			modelCommand = command
-		case "LICENSE", "TEMPLATE", "SYSTEM", "PROMPT", "ADAPTER":
+		case "ADAPTER":
+			command.Name = string(bytes.ToLower(fields[0]))
+			command.Args = string(bytes.TrimSpace(fields[1]))
+		case "LICENSE", "TEMPLATE", "SYSTEM", "PROMPT":
 			command.Name = string(bytes.ToLower(fields[0]))
 			command.Args = string(fields[1])
 		case "PARAMETER":
@@ -50,7 +53,7 @@ func Parse(reader io.Reader) ([]Command, error) {
 			}
 
 			command.Name = string(fields[0])
-			command.Args = string(fields[1])
+			command.Args = string(bytes.TrimSpace(fields[1]))
 		case "EMBED":
 			return nil, fmt.Errorf("deprecated command: EMBED is no longer supported, use the /embed API endpoint instead")
 		default:

progress/bar.go

@@ -0,0 +1,215 @@
+package progress
+
+import (
+	"fmt"
+	"os"
+	"strings"
+	"time"
+
+	"github.com/jmorganca/ollama/format"
+	"golang.org/x/term"
+)
+
+type Bar struct {
+	message      string
+	messageWidth int
+
+	maxValue     int64
+	initialValue int64
+	currentValue int64
+
+	started time.Time
+	stopped time.Time
+
+	maxBuckets int
+	buckets    []bucket
+}
+
+type bucket struct {
+	updated time.Time
+	value   int64
+}
+
+func NewBar(message string, maxValue, initialValue int64) *Bar {
+	b := Bar{
+		message:      message,
+		messageWidth: -1,
+		maxValue:     maxValue,
+		initialValue: initialValue,
+		currentValue: initialValue,
+		started:      time.Now(),
+		maxBuckets:   10,
+	}
+
+	if initialValue >= maxValue {
+		b.stopped = time.Now()
+	}
+
+	return &b
+}
+
+// formatDuration limits the rendering of a time.Duration to 2 units
+func formatDuration(d time.Duration) string {
+	switch {
+	case d >= 100*time.Hour:
+		return "99h+"
+	case d >= time.Hour:
+		return fmt.Sprintf("%dh%dm", int(d.Hours()), int(d.Minutes())%60)
+	default:
+		return d.Round(time.Second).String()
+	}
+}
+
+func (b *Bar) String() string {
+	termWidth, _, err := term.GetSize(int(os.Stderr.Fd()))
+	if err != nil {
+		termWidth = 80
+	}
+
+	var pre strings.Builder
+	if len(b.message) > 0 {
+		message := strings.TrimSpace(b.message)
+		if b.messageWidth > 0 && len(message) > b.messageWidth {
+			message = message[:b.messageWidth]
+		}
+
+		fmt.Fprintf(&pre, "%s", message)
+		if padding := b.messageWidth - pre.Len(); padding > 0 {
+			pre.WriteString(repeat(" ", padding))
+		}
+
+		pre.WriteString(" ")
+	}
+
+	fmt.Fprintf(&pre, "%3.0f%%", b.percent())
+
+	var suf strings.Builder
+	// max 13 characters: "999 MB/999 MB"
+	if b.stopped.IsZero() {
+		curValue := format.HumanBytes(b.currentValue)
+		suf.WriteString(repeat(" ", 6-len(curValue)))
+		suf.WriteString(curValue)
+		suf.WriteString("/")
+
+		maxValue := format.HumanBytes(b.maxValue)
+		suf.WriteString(repeat(" ", 6-len(maxValue)))
+		suf.WriteString(maxValue)
+	} else {
+		maxValue := format.HumanBytes(b.maxValue)
+		suf.WriteString(repeat(" ", 6-len(maxValue)))
+		suf.WriteString(maxValue)
+		suf.WriteString(repeat(" ", 7))
+	}
+
+	rate := b.rate()
+	// max 10 characters: "  999 MB/s"
+	if b.stopped.IsZero() && rate > 0 {
+		suf.WriteString("  ")
+		humanRate := format.HumanBytes(int64(rate))
+		suf.WriteString(repeat(" ", 6-len(humanRate)))
+		suf.WriteString(humanRate)
+		suf.WriteString("/s")
+	} else {
+		suf.WriteString(repeat(" ", 10))
+	}
+
+	// max 8 characters: "  59m59s"
+	if b.stopped.IsZero() && rate > 0 {
+		suf.WriteString("  ")
+		var remaining time.Duration
+		if rate > 0 {
+			remaining = time.Duration(int64(float64(b.maxValue-b.currentValue)/rate)) * time.Second
+		}
+
+		humanRemaining := formatDuration(remaining)
+		suf.WriteString(repeat(" ", 6-len(humanRemaining)))
+		suf.WriteString(humanRemaining)
+	} else {
+		suf.WriteString(repeat(" ", 8))
+	}
+
+	var mid strings.Builder
+	// add 5 extra spaces: 2 boundary characters and 1 space at each end
+	f := termWidth - pre.Len() - suf.Len() - 5
+	n := int(float64(f) * b.percent() / 100)
+
+	mid.WriteString(" ▕")
+
+	if n > 0 {
+		mid.WriteString(repeat("█", n))
+	}
+
+	if f-n > 0 {
+		mid.WriteString(repeat(" ", f-n))
+	}
+
+	mid.WriteString("▏ ")
+
+	return pre.String() + mid.String() + suf.String()
+}
+
+func (b *Bar) Set(value int64) {
+	if value >= b.maxValue {
+		value = b.maxValue
+	}
+
+	b.currentValue = value
+	if b.currentValue >= b.maxValue {
+		b.stopped = time.Now()
+	}
+
+	// throttle bucket updates to 1 per second
+	if len(b.buckets) == 0 || time.Since(b.buckets[len(b.buckets)-1].updated) > time.Second {
+		b.buckets = append(b.buckets, bucket{
+			updated: time.Now(),
+			value:   value,
+		})
+
+		if len(b.buckets) > b.maxBuckets {
+			b.buckets = b.buckets[1:]
+		}
+	}
+}
+
+func (b *Bar) percent() float64 {
+	if b.maxValue > 0 {
+		return float64(b.currentValue) / float64(b.maxValue) * 100
+	}
+
+	return 0
+}
+
+func (b *Bar) rate() float64 {
+	var numerator, denominator float64
+
+	if !b.stopped.IsZero() {
+		numerator = float64(b.currentValue - b.initialValue)
+		denominator = b.stopped.Sub(b.started).Round(time.Second).Seconds()
+	} else {
+		switch len(b.buckets) {
+		case 0:
+			// noop
+		case 1:
+			numerator = float64(b.buckets[0].value - b.initialValue)
+			denominator = b.buckets[0].updated.Sub(b.started).Round(time.Second).Seconds()
+		default:
+			first, last := b.buckets[0], b.buckets[len(b.buckets)-1]
+			numerator = float64(last.value - first.value)
+			denominator = last.updated.Sub(first.updated).Round(time.Second).Seconds()
+		}
+	}
+
+	if denominator != 0 {
+		return numerator / denominator
+	}
+
+	return 0
+}
+
+func repeat(s string, n int) string {
+	if n > 0 {
+		return strings.Repeat(s, n)
+	}
+
+	return ""
+}

progress/progress.go

@@ -0,0 +1,113 @@
+package progress
+
+import (
+	"fmt"
+	"io"
+	"sync"
+	"time"
+)
+
+type State interface {
+	String() string
+}
+
+type Progress struct {
+	mu sync.Mutex
+	w  io.Writer
+
+	pos int
+
+	ticker *time.Ticker
+	states []State
+}
+
+func NewProgress(w io.Writer) *Progress {
+	p := &Progress{w: w}
+	go p.start()
+	return p
+}
+
+func (p *Progress) stop() bool {
+	for _, state := range p.states {
+		if spinner, ok := state.(*Spinner); ok {
+			spinner.Stop()
+		}
+	}
+
+	if p.ticker != nil {
+		p.ticker.Stop()
+		p.ticker = nil
+		p.render()
+		return true
+	}
+
+	return false
+}
+
+func (p *Progress) Stop() bool {
+	stopped := p.stop()
+	if stopped {
+		fmt.Fprint(p.w, "\n")
+	}
+	return stopped
+}
+
+func (p *Progress) StopAndClear() bool {
+	fmt.Fprint(p.w, "\033[?25l")
+	defer fmt.Fprint(p.w, "\033[?25h")
+
+	stopped := p.stop()
+	if stopped {
+		// clear all progress lines
+		for i := 0; i < p.pos; i++ {
+			if i > 0 {
+				fmt.Fprint(p.w, "\033[A")
+			}
+			fmt.Fprint(p.w, "\033[2K\033[1G")
+		}
+	}
+
+	return stopped
+}
+
+func (p *Progress) Add(key string, state State) {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+
+	p.states = append(p.states, state)
+}
+
+func (p *Progress) render() error {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+
+	fmt.Fprint(p.w, "\033[?25l")
+	defer fmt.Fprint(p.w, "\033[?25h")
+
+	// clear already rendered progress lines
+	for i := 0; i < p.pos; i++ {
+		if i > 0 {
+			fmt.Fprint(p.w, "\033[A")
+		}
+		fmt.Fprint(p.w, "\033[2K\033[1G")
+	}
+
+	// render progress lines
+	for i, state := range p.states {
+		fmt.Fprint(p.w, state.String())
+		if i < len(p.states)-1 {
+			fmt.Fprint(p.w, "\n")
+		}
+	}
+
+	p.pos = len(p.states)
+
+	return nil
+}
+
+func (p *Progress) start() {
+	p.ticker = time.NewTicker(100 * time.Millisecond)
+	for range p.ticker.C {
+		p.render()
+	}
+}

progress/spinner.go

@@ -0,0 +1,73 @@
+package progress
+
+import (
+	"fmt"
+	"strings"
+	"time"
+)
+
+type Spinner struct {
+	message      string
+	messageWidth int
+
+	parts []string
+
+	value int
+
+	ticker  *time.Ticker
+	started time.Time
+	stopped time.Time
+}
+
+func NewSpinner(message string) *Spinner {
+	s := &Spinner{
+		message: message,
+		parts: []string{
+			"⠋", "⠙", "⠹", "⠸", "⠼", "⠴", "⠦", "⠧", "⠇", "⠏",
+		},
+		started: time.Now(),
+	}
+	go s.start()
+	return s
+}
+
+func (s *Spinner) String() string {
+	var sb strings.Builder
+	if len(s.message) > 0 {
+		message := strings.TrimSpace(s.message)
+		if s.messageWidth > 0 && len(message) > s.messageWidth {
+			message = message[:s.messageWidth]
+		}
+
+		fmt.Fprintf(&sb, "%s", message)
+		if padding := s.messageWidth - sb.Len(); padding > 0 {
+			sb.WriteString(strings.Repeat(" ", padding))
+		}
+
+		sb.WriteString(" ")
+	}
+
+	if s.stopped.IsZero() {
+		spinner := s.parts[s.value]
+		sb.WriteString(spinner)
+		sb.WriteString(" ")
+	}
+
+	return sb.String()
+}
+
+func (s *Spinner) start() {
+	s.ticker = time.NewTicker(100 * time.Millisecond)
+	for range s.ticker.C {
+		s.value = (s.value + 1) % len(s.parts)
+		if !s.stopped.IsZero() {
+			return
+		}
+	}
+}
+
+func (s *Spinner) Stop() {
+	if s.stopped.IsZero() {
+		s.stopped = time.Now()
+	}
+}

progressbar/LICENSE

@@ -1,21 +0,0 @@
-MIT License
-
-Copyright (c) 2017 Zack
-
-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.

progressbar/README.md

@@ -1,121 +0,0 @@
-# progressbar
-
-[![CI](https://github.com/schollz/progressbar/actions/workflows/ci.yml/badge.svg?branch=main&event=push)](https://github.com/schollz/progressbar/actions/workflows/ci.yml)
-[![go report card](https://goreportcard.com/badge/github.com/schollz/progressbar)](https://goreportcard.com/report/github.com/schollz/progressbar) 
-[![coverage](https://img.shields.io/badge/coverage-84%25-brightgreen.svg)](https://gocover.io/github.com/schollz/progressbar)
-[![godocs](https://godoc.org/github.com/schollz/progressbar?status.svg)](https://godoc.org/github.com/schollz/progressbar/v3) 
-
-A very simple thread-safe progress bar which should work on every OS without problems. I needed a progressbar for [croc](https://github.com/schollz/croc) and everything I tried had problems, so I made another one. In order to be OS agnostic I do not plan to support [multi-line outputs](https://github.com/schollz/progressbar/issues/6).
-
-
-## Install
-
-```
-go get -u github.com/schollz/progressbar/v3
-```
-
-## Usage 
-
-### Basic usage
-
-```golang
-bar := progressbar.Default(100)
-for i := 0; i < 100; i++ {
-    bar.Add(1)
-    time.Sleep(40 * time.Millisecond)
-}
-```
-
-which looks like:
-
-![Example of basic bar](examples/basic/basic.gif)
-
-
-### I/O operations
-
-The `progressbar` implements an `io.Writer` so it can automatically detect the number of bytes written to a stream, so you can use it as a progressbar for an `io.Reader`.
-
-```golang
-req, _ := http.NewRequest("GET", "https://dl.google.com/go/go1.14.2.src.tar.gz", nil)
-resp, _ := http.DefaultClient.Do(req)
-defer resp.Body.Close()
-
-f, _ := os.OpenFile("go1.14.2.src.tar.gz", os.O_CREATE|os.O_WRONLY, 0644)
-defer f.Close()
-
-bar := progressbar.DefaultBytes(
-    resp.ContentLength,
-    "downloading",
-)
-io.Copy(io.MultiWriter(f, bar), resp.Body)
-```
-
-which looks like:
-
-![Example of download bar](examples/download/download.gif)
-
-
-### Progress bar with unknown length
-
-A progressbar with unknown length is a spinner. Any bar with -1 length will automatically convert it to a spinner with a customizable spinner type. For example, the above code can be run and set the `resp.ContentLength` to `-1`.
-
-which looks like:
-
-![Example of download bar with unknown length](examples/download-unknown/download-unknown.gif)
-
-
-### Customization
-
-There is a lot of customization that you can do - change the writer, the color, the width, description, theme, etc. See [all the options](https://pkg.go.dev/github.com/schollz/progressbar/v3?tab=doc#Option).
-
-```golang
-bar := progressbar.NewOptions(1000,
-    progressbar.OptionSetWriter(ansi.NewAnsiStdout()),
-    progressbar.OptionEnableColorCodes(true),
-    progressbar.OptionShowBytes(true),
-    progressbar.OptionSetWidth(15),
-    progressbar.OptionSetDescription("[cyan][1/3][reset] Writing moshable file..."),
-    progressbar.OptionSetTheme(progressbar.Theme{
-        Saucer:        "[green]=[reset]",
-        SaucerHead:    "[green]>[reset]",
-        SaucerPadding: " ",
-        BarStart:      "[",
-        BarEnd:        "]",
-    }))
-for i := 0; i < 1000; i++ {
-    bar.Add(1)
-    time.Sleep(5 * time.Millisecond)
-}
-```
-
-which looks like:
-
-![Example of customized bar](examples/customization/customization.gif)
-
-
-## Contributing
-
-Pull requests are welcome. Feel free to...
-
-- Revise documentation
-- Add new features
-- Fix bugs
-- Suggest improvements
-
-## Thanks
-
-Thanks [@Dynom](https://github.com/dynom) for massive improvements in version 2.0!
-
-Thanks [@CrushedPixel](https://github.com/CrushedPixel) for adding descriptions and color code support!
-
-Thanks [@MrMe42](https://github.com/MrMe42) for adding some minor features!
-
-Thanks [@tehstun](https://github.com/tehstun) for some great PRs!
-
-Thanks [@Benzammour](https://github.com/Benzammour) and [@haseth](https://github.com/haseth) for helping create v3!
-
-Thanks [@briandowns](https://github.com/briandowns) for compiling the list of spinners.
-
-## License
-
-MIT

progressbar/spinners.go

@@ -1,80 +0,0 @@
-package progressbar
-
-var spinners = map[int][]string{
-	0:  {"←", "↖", "↑", "↗", "→", "↘", "↓", "↙"},
-	1:  {"▁", "▃", "▄", "▅", "▆", "▇", "█", "▇", "▆", "▅", "▄", "▃", "▁"},
-	2:  {"▖", "▘", "▝", "▗"},
-	3:  {"┤", "┘", "┴", "└", "├", "┌", "┬", "┐"},
-	4:  {"◢", "◣", "◤", "◥"},
-	5:  {"◰", "◳", "◲", "◱"},
-	6:  {"◴", "◷", "◶", "◵"},
-	7:  {"◐", "◓", "◑", "◒"},
-	8:  {".", "o", "O", "@", "*"},
-	9:  {"|", "/", "-", "\\"},
-	10: {"◡◡", "⊙⊙", "◠◠"},
-	11: {"⣾", "⣽", "⣻", "⢿", "⡿", "⣟", "⣯", "⣷"},
-	12: {">))'>", " >))'>", "  >))'>", "   >))'>", "    >))'>", "   <'((<", "  <'((<", " <'((<"},
-	13: {"⠁", "⠂", "⠄", "⡀", "⢀", "⠠", "⠐", "⠈"},
-	14: {"⠋", "⠙", "⠹", "⠸", "⠼", "⠴", "⠦", "⠧", "⠇", "⠏"},
-	15: {"a", "b", "c", "d", "e", "f", "g", "h", "i", "j", "k", "l", "m", "n", "o", "p", "q", "r", "s", "t", "u", "v", "w", "x", "y", "z"},
-	16: {"▉", "▊", "▋", "▌", "▍", "▎", "▏", "▎", "▍", "▌", "▋", "▊", "▉"},
-	17: {"■", "□", "▪", "▫"},
-	18: {"←", "↑", "→", "↓"},
-	19: {"╫", "╪"},
-	20: {"⇐", "⇖", "⇑", "⇗", "⇒", "⇘", "⇓", "⇙"},
-	21: {"⠁", "⠁", "⠉", "⠙", "⠚", "⠒", "⠂", "⠂", "⠒", "⠲", "⠴", "⠤", "⠄", "⠄", "⠤", "⠠", "⠠", "⠤", "⠦", "⠖", "⠒", "⠐", "⠐", "⠒", "⠓", "⠋", "⠉", "⠈", "⠈"},
-	22: {"⠈", "⠉", "⠋", "⠓", "⠒", "⠐", "⠐", "⠒", "⠖", "⠦", "⠤", "⠠", "⠠", "⠤", "⠦", "⠖", "⠒", "⠐", "⠐", "⠒", "⠓", "⠋", "⠉", "⠈"},
-	23: {"⠁", "⠉", "⠙", "⠚", "⠒", "⠂", "⠂", "⠒", "⠲", "⠴", "⠤", "⠄", "⠄", "⠤", "⠴", "⠲", "⠒", "⠂", "⠂", "⠒", "⠚", "⠙", "⠉", "⠁"},
-	24: {"⠋", "⠙", "⠚", "⠒", "⠂", "⠂", "⠒", "⠲", "⠴", "⠦", "⠖", "⠒", "⠐", "⠐", "⠒", "⠓", "⠋"},
-	25: {"ｦ", "ｧ", "ｨ", "ｩ", "ｪ", "ｫ", "ｬ", "ｭ", "ｮ", "ｯ", "ｱ", "ｲ", "ｳ", "ｴ", "ｵ", "ｶ", "ｷ", "ｸ", "ｹ", "ｺ", "ｻ", "ｼ", "ｽ", "ｾ", "ｿ", "ﾀ", "ﾁ", "ﾂ", "ﾃ", "ﾄ", "ﾅ", "ﾆ", "ﾇ", "ﾈ", "ﾉ", "ﾊ", "ﾋ", "ﾌ", "ﾍ", "ﾎ", "ﾏ", "ﾐ", "ﾑ", "ﾒ", "ﾓ", "ﾔ", "ﾕ", "ﾖ", "ﾗ", "ﾘ", "ﾙ", "ﾚ", "ﾛ", "ﾜ", "ﾝ"},
-	26: {".", "..", "..."},
-	27: {"▁", "▂", "▃", "▄", "▅", "▆", "▇", "█", "▉", "▊", "▋", "▌", "▍", "▎", "▏", "▏", "▎", "▍", "▌", "▋", "▊", "▉", "█", "▇", "▆", "▅", "▄", "▃", "▂", "▁"},
-	28: {".", "o", "O", "°", "O", "o", "."},
-	29: {"+", "x"},
-	30: {"v", "<", "^", ">"},
-	31: {">>--->", " >>--->", "  >>--->", "   >>--->", "    >>--->", "    <---<<", "   <---<<", "  <---<<", " <---<<", "<---<<"},
-	32: {"|", "||", "|||", "||||", "|||||", "|||||||", "||||||||", "|||||||", "||||||", "|||||", "||||", "|||", "||", "|"},
-	33: {"[          ]", "[=         ]", "[==        ]", "[===       ]", "[====      ]", "[=====     ]", "[======    ]", "[=======   ]", "[========  ]", "[========= ]", "[==========]"},
-	34: {"(*---------)", "(-*--------)", "(--*-------)", "(---*------)", "(----*-----)", "(-----*----)", "(------*---)", "(-------*--)", "(--------*-)", "(---------*)"},
-	35: {"█▒▒▒▒▒▒▒▒▒", "███▒▒▒▒▒▒▒", "█████▒▒▒▒▒", "███████▒▒▒", "██████████"},
-	36: {"[                    ]", "[=>                  ]", "[===>                ]", "[=====>              ]", "[======>             ]", "[========>           ]", "[==========>         ]", "[============>       ]", "[==============>     ]", "[================>   ]", "[==================> ]", "[===================>]"},
-	37: {"ဝ", "၀"},
-	38: {"▌", "▀", "▐▄"},
-	39: {"🌍", "🌎", "🌏"},
-	40: {"◜", "◝", "◞", "◟"},
-	41: {"⬒", "⬔", "⬓", "⬕"},
-	42: {"⬖", "⬘", "⬗", "⬙"},
-	43: {"[>>>          >]", "[]>>>>        []", "[]  >>>>      []", "[]    >>>>    []", "[]      >>>>  []", "[]        >>>>[]", "[>>          >>]"},
-	44: {"♠", "♣", "♥", "♦"},
-	45: {"➞", "➟", "➠", "➡", "➠", "➟"},
-	46: {"  |  ", ` \   `, "_    ", ` \   `, "  |  ", "   / ", "    _", "   / "},
-	47: {"  . . . .", ".   . . .", ". .   . .", ". . .   .", ". . . .  ", ". . . . ."},
-	48: {" |     ", "  /    ", "   _   ", `    \  `, "     | ", `    \  `, "   _   ", "  /    "},
-	49: {"⎺", "⎻", "⎼", "⎽", "⎼", "⎻"},
-	50: {"▹▹▹▹▹", "▸▹▹▹▹", "▹▸▹▹▹", "▹▹▸▹▹", "▹▹▹▸▹", "▹▹▹▹▸"},
-	51: {"[    ]", "[   =]", "[  ==]", "[ ===]", "[====]", "[=== ]", "[==  ]", "[=   ]"},
-	52: {"( ●    )", "(  ●   )", "(   ●  )", "(    ● )", "(     ●)", "(    ● )", "(   ●  )", "(  ●   )", "( ●    )"},
-	53: {"✶", "✸", "✹", "✺", "✹", "✷"},
-	54: {"▐|\\____________▌", "▐_|\\___________▌", "▐__|\\__________▌", "▐___|\\_________▌", "▐____|\\________▌", "▐_____|\\_______▌", "▐______|\\______▌", "▐_______|\\_____▌", "▐________|\\____▌", "▐_________|\\___▌", "▐__________|\\__▌", "▐___________|\\_▌", "▐____________|\\▌", "▐____________/|▌", "▐___________/|_▌", "▐__________/|__▌", "▐_________/|___▌", "▐________/|____▌", "▐_______/|_____▌", "▐______/|______▌", "▐_____/|_______▌", "▐____/|________▌", "▐___/|_________▌", "▐__/|__________▌", "▐_/|___________▌", "▐/|____________▌"},
-	55: {"▐⠂       ▌", "▐⠈       ▌", "▐ ⠂      ▌", "▐ ⠠      ▌", "▐  ⡀     ▌", "▐  ⠠     ▌", "▐   ⠂    ▌", "▐   ⠈    ▌", "▐    ⠂   ▌", "▐    ⠠   ▌", "▐     ⡀  ▌", "▐     ⠠  ▌", "▐      ⠂ ▌", "▐      ⠈ ▌", "▐       ⠂▌", "▐       ⠠▌", "▐       ⡀▌", "▐      ⠠ ▌", "▐      ⠂ ▌", "▐     ⠈  ▌", "▐     ⠂  ▌", "▐    ⠠   ▌", "▐    ⡀   ▌", "▐   ⠠    ▌", "▐   ⠂    ▌", "▐  ⠈     ▌", "▐  ⠂     ▌", "▐ ⠠      ▌", "▐ ⡀      ▌", "▐⠠       ▌"},
-	56: {"¿", "?"},
-	57: {"⢹", "⢺", "⢼", "⣸", "⣇", "⡧", "⡗", "⡏"},
-	58: {"⢄", "⢂", "⢁", "⡁", "⡈", "⡐", "⡠"},
-	59: {".  ", ".. ", "...", " ..", "  .", "   "},
-	60: {".", "o", "O", "°", "O", "o", "."},
-	61: {"▓", "▒", "░"},
-	62: {"▌", "▀", "▐", "▄"},
-	63: {"⊶", "⊷"},
-	64: {"▪", "▫"},
-	65: {"□", "■"},
-	66: {"▮", "▯"},
-	67: {"-", "=", "≡"},
-	68: {"d", "q", "p", "b"},
-	69: {"∙∙∙", "●∙∙", "∙●∙", "∙∙●", "∙∙∙"},
-	70: {"🌑 ", "🌒 ", "🌓 ", "🌔 ", "🌕 ", "🌖 ", "🌗 ", "🌘 "},
-	71: {"☗", "☖"},
-	72: {"⧇", "⧆"},
-	73: {"◉", "◎"},
-	74: {"㊂", "㊀", "㊁"},
-	75: {"⦾", "⦿"},
-}

readline/history.go

@@ -43,9 +43,12 @@ func (h *History) Init() error {
 	}
 
 	path := filepath.Join(home, ".ollama", "history")
+	if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
+		return err
+	}
+
 	h.Filename = path
 
-	//todo check if the file exists
 	f, err := os.OpenFile(path, os.O_CREATE|os.O_RDONLY, 0600)
 	if err != nil {
 		if errors.Is(err, os.ErrNotExist) {

readline/readline.go

@@ -24,6 +24,7 @@ type Instance struct {
 	Prompt   *Prompt
 	Terminal *Terminal
 	History  *History
+	Pasting  bool
 }
 
 func New(prompt Prompt) (*Instance, error) {
@@ -46,7 +47,7 @@ func New(prompt Prompt) (*Instance, error) {
 
 func (i *Instance) Readline() (string, error) {
 	prompt := i.Prompt.Prompt
-	if i.Prompt.UseAlt {
+	if i.Prompt.UseAlt || i.Pasting {
 		prompt = i.Prompt.AltPrompt
 	}
 	fmt.Print(prompt)
@@ -63,12 +64,13 @@ func (i *Instance) Readline() (string, error) {
 	var esc bool
 	var escex bool
 	var metaDel bool
-	var pasteMode PasteMode
 
 	var currentLineBuf []rune
 
 	for {
-		if buf.IsEmpty() {
+		// don't show placeholder when pasting unless we're in multiline mode
+		showPlaceholder := !i.Pasting || i.Prompt.UseAlt
+		if buf.IsEmpty() && showPlaceholder {
 			ph := i.Prompt.Placeholder
 			if i.Prompt.UseAlt {
 				ph = i.Prompt.AltPlaceholder
@@ -119,9 +121,9 @@ func (i *Instance) Readline() (string, error) {
 					code += string(r)
 				}
 				if code == CharBracketedPasteStart {
-					pasteMode = PasteModeStart
+					i.Pasting = true
 				} else if code == CharBracketedPasteEnd {
-					pasteMode = PasteModeEnd
+					i.Pasting = false
 				}
 			case KeyDel:
 				if buf.Size() > 0 {
@@ -145,6 +147,8 @@ func (i *Instance) Readline() (string, error) {
 				buf.MoveLeftWord()
 			case 'f':
 				buf.MoveRightWord()
+			case CharBackspace:
+				buf.DeleteWord()
 			case CharEscapeEx:
 				escex = true
 			}
@@ -187,6 +191,8 @@ func (i *Instance) Readline() (string, error) {
 			buf.ClearScreen()
 		case CharCtrlW:
 			buf.DeleteWord()
+		case CharCtrlZ:
+			return handleCharCtrlZ(fd, termios)
 		case CharEnter:
 			output := buf.String()
 			if output != "" {
@@ -194,12 +200,7 @@ func (i *Instance) Readline() (string, error) {
 			}
 			buf.MoveToEnd()
 			fmt.Println()
-			switch pasteMode {
-			case PasteModeStart:
-				output = `"""` + output
-			case PasteModeEnd:
-				output = output + `"""`
-			}
+
 			return output, nil
 		default:
 			if metaDel {

readline/readline_unix.go

@@ -0,0 +1,18 @@
+//go:build !windows
+
+package readline
+
+import (
+	"syscall"
+)
+
+func handleCharCtrlZ(fd int, termios *Termios) (string, error) {
+	if err := UnsetRawMode(fd, termios); err != nil {
+		return "", err
+	}
+
+	syscall.Kill(0, syscall.SIGSTOP)
+
+	// on resume...
+	return "", nil
+}

readline/readline_windows.go

@@ -0,0 +1,6 @@
+package readline
+
+func handleCharCtrlZ(fd int, state *State) (string, error) {
+	// not supported
+	return "", nil
+}

readline/types.go

@@ -76,11 +76,3 @@ const (
 	CharBracketedPasteStart = "00~"
 	CharBracketedPasteEnd   = "01~"
 )
-
-type PasteMode int
-
-const (
-	PastModeOff = iota
-	PasteModeStart
-	PasteModeEnd
-)

scripts/build_darwin.sh

@@ -10,6 +10,7 @@ mkdir -p dist
 for TARGETARCH in arm64 amd64; do
     GOOS=darwin GOARCH=$TARGETARCH go generate ./...
     GOOS=darwin GOARCH=$TARGETARCH go build -o dist/ollama-darwin-$TARGETARCH
+    rm -rf llm/llama.cpp/*/build
 done
 
 lipo -create -output dist/ollama dist/ollama-darwin-*

scripts/build_docker.sh

@@ -10,6 +10,8 @@ docker buildx build \
     --platform=linux/arm64,linux/amd64 \
     --build-arg=VERSION \
     --build-arg=GOFLAGS \
+    --cache-from type=local,src=.cache \
+    --cache-to type=local,dest=.cache \
     -f Dockerfile \
-    -t ollama \
+    -t ollama/ollama:$VERSION \
     .

scripts/install.sh

@@ -181,6 +181,9 @@ install_cuda_driver_apt() {
         debian)
             status 'Enabling contrib sources...'
             $SUDO sed 's/main/contrib/' < /etc/apt/sources.list | $SUDO tee /etc/apt/sources.list.d/contrib.list > /dev/null
+            if [ -f "/etc/apt/sources.list.d/debian.sources" ]; then
+                $SUDO sed 's/main/contrib/' < /etc/apt/sources.list.d/debian.sources | $SUDO tee /etc/apt/sources.list.d/contrib.sources > /dev/null
+            fi
             ;;
     esac
 
@@ -214,7 +217,7 @@ fi
 
 if ! check_gpu nvidia-smi || [ -z "$(nvidia-smi | grep -o "CUDA Version: [0-9]*\.[0-9]*")" ]; then
     case $OS_NAME in
-        centos|rhel) install_cuda_driver_yum 'rhel' $OS_VERSION ;;
+        centos|rhel) install_cuda_driver_yum 'rhel' $(echo $OS_VERSION | cut -d '.' -f 1) ;;
         rocky) install_cuda_driver_yum 'rhel' $(echo $OS_VERSION | cut -c1) ;;
         fedora) install_cuda_driver_yum $OS_NAME $OS_VERSION ;;
         amzn) install_cuda_driver_yum 'fedora' '35' ;;
@@ -227,7 +230,8 @@ fi
 if ! lsmod | grep -q nvidia; then
     KERNEL_RELEASE="$(uname -r)"
     case $OS_NAME in
-        centos|rhel|rocky|amzn) $SUDO $PACKAGE_MANAGER -y install kernel-devel-$KERNEL_RELEASE kernel-headers-$KERNEL_RELEASE ;;
+        rocky) $SUDO $PACKAGE_MANAGER -y install kernel-devel kernel-headers ;;
+        centos|rhel|amzn) $SUDO $PACKAGE_MANAGER -y install kernel-devel-$KERNEL_RELEASE kernel-headers-$KERNEL_RELEASE ;;
         fedora) $SUDO $PACKAGE_MANAGER -y install kernel-devel-$KERNEL_RELEASE ;;
         debian|ubuntu) $SUDO apt-get -y install linux-headers-$KERNEL_RELEASE ;;
         *) exit ;;

scripts/push_docker.sh

@@ -10,6 +10,7 @@ docker buildx build \
     --platform=linux/arm64,linux/amd64 \
     --build-arg=VERSION \
     --build-arg=GOFLAGS \
+    --cache-from type=local,src=.cache \
     -f Dockerfile \
     -t ollama/ollama -t ollama/ollama:$VERSION \
     .

server/download.go

@@ -7,6 +7,7 @@ import (
 	"fmt"
 	"io"
 	"log"
+	"math"
 	"net/http"
 	"net/url"
 	"os"
@@ -53,8 +54,8 @@ type blobDownloadPart struct {
 
 const (
 	numDownloadParts          = 64
-	minDownloadPartSize int64 = 32 * 1000 * 1000
-	maxDownloadPartSize int64 = 256 * 1000 * 1000
+	minDownloadPartSize int64 = 100 * format.MegaByte
+	maxDownloadPartSize int64 = 1000 * format.MegaByte
 )
 
 func (p *blobDownloadPart) Name() string {
@@ -147,7 +148,6 @@ func (b *blobDownload) run(ctx context.Context, requestURL *url.URL, opts *Regis
 			continue
 		}
 
-		i := i
 		g.Go(func() error {
 			var err error
 			for try := 0; try < maxRetries; try++ {
@@ -158,12 +158,11 @@ func (b *blobDownload) run(ctx context.Context, requestURL *url.URL, opts *Regis
 					// return immediately if the context is canceled or the device is out of space
 					return err
 				case err != nil:
-					log.Printf("%s part %d attempt %d failed: %v, retrying", b.Digest[7:19], i, try, err)
+					sleep := time.Second * time.Duration(math.Pow(2, float64(try)))
+					log.Printf("%s part %d attempt %d failed: %v, retrying in %s", b.Digest[7:19], part.N, try, err, sleep)
+					time.Sleep(sleep)
 					continue
 				default:
-					if try > 0 {
-						log.Printf("%s part %d completed after %d retries", b.Digest[7:19], i, try)
-					}
 					return nil
 				}
 			}
@@ -285,7 +284,7 @@ func (b *blobDownload) Wait(ctx context.Context, fn func(api.ProgressResponse))
 		}
 
 		fn(api.ProgressResponse{
-			Status:    fmt.Sprintf("downloading %s", b.Digest),
+			Status:    fmt.Sprintf("pulling %s", b.Digest[7:19]),
 			Digest:    b.Digest,
 			Total:     b.Total,
 			Completed: b.Completed.Load(),
@@ -304,7 +303,7 @@ type downloadOpts struct {
 	fn      func(api.ProgressResponse)
 }
 
-const maxRetries = 3
+const maxRetries = 6
 
 var errMaxRetriesExceeded = errors.New("max retries exceeded")
 
@@ -322,7 +321,7 @@ func downloadBlob(ctx context.Context, opts downloadOpts) error {
 		return err
 	default:
 		opts.fn(api.ProgressResponse{
-			Status:    fmt.Sprintf("downloading %s", opts.digest),
+			Status:    fmt.Sprintf("pulling %s", opts.digest[7:19]),
 			Digest:    opts.digest,
 			Total:     fi.Size(),
 			Completed: fi.Size(),

server/images_test.go

@@ -1,23 +1,98 @@
 package server
 
 import (
+	"strings"
 	"testing"
 
 	"github.com/jmorganca/ollama/api"
 )
 
-func TestModelPrompt(t *testing.T) {
-	var m Model
-	req := api.GenerateRequest{
-		Template: "a{{ .Prompt }}b",
-		Prompt:   "<h1>",
+func TestChat(t *testing.T) {
+	tests := []struct {
+		name     string
+		template string
+		msgs     []api.Message
+		want     string
+		wantErr  string
+	}{
+		{
+			name:     "Single Message",
+			template: "[INST] {{ .System }} {{ .Prompt }} [/INST]",
+			msgs: []api.Message{
+				{
+					Role:    "system",
+					Content: "You are a Wizard.",
+				},
+				{
+					Role:    "user",
+					Content: "What are the potion ingredients?",
+				},
+			},
+			want: "[INST] You are a Wizard. What are the potion ingredients? [/INST]",
+		},
+		{
+			name:     "Message History",
+			template: "[INST] {{ .System }} {{ .Prompt }} [/INST]",
+			msgs: []api.Message{
+				{
+					Role:    "system",
+					Content: "You are a Wizard.",
+				},
+				{
+					Role:    "user",
+					Content: "What are the potion ingredients?",
+				},
+				{
+					Role:    "assistant",
+					Content: "sugar",
+				},
+				{
+					Role:    "user",
+					Content: "Anything else?",
+				},
+			},
+			want: "[INST] You are a Wizard. What are the potion ingredients? [/INST]sugar[INST]  Anything else? [/INST]",
+		},
+		{
+			name:     "Assistant Only",
+			template: "[INST] {{ .System }} {{ .Prompt }} [/INST]",
+			msgs: []api.Message{
+				{
+					Role:    "assistant",
+					Content: "everything nice",
+				},
+			},
+			want: "[INST]   [/INST]everything nice",
+		},
+		{
+			name: "Invalid Role",
+			msgs: []api.Message{
+				{
+					Role:    "not-a-role",
+					Content: "howdy",
+				},
+			},
+			wantErr: "invalid role: not-a-role",
+		},
 	}
-	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)
+
+	for _, tt := range tests {
+		m := Model{
+			Template: tt.template,
+		}
+		t.Run(tt.name, func(t *testing.T) {
+			got, _, err := m.ChatPrompt(tt.msgs)
+			if tt.wantErr != "" {
+				if err == nil {
+					t.Errorf("ChatPrompt() expected error, got nil")
+				}
+				if !strings.Contains(err.Error(), tt.wantErr) {
+					t.Errorf("ChatPrompt() error = %v, wantErr %v", err, tt.wantErr)
+				}
+			}
+			if got != tt.want {
+				t.Errorf("ChatPrompt() got = %v, want %v", got, tt.want)
+			}
+		})
 	}
 }

server/layers.go

@@ -0,0 +1,109 @@
+package server
+
+import (
+	"crypto/sha256"
+	"fmt"
+	"io"
+	"os"
+	"runtime"
+	"strings"
+
+	"golang.org/x/exp/slices"
+)
+
+type Layers struct {
+	items []*Layer
+}
+
+func (ls *Layers) Add(layer *Layer) {
+	if layer.Size > 0 {
+		ls.items = append(ls.items, layer)
+	}
+}
+
+func (ls *Layers) Replace(layer *Layer) {
+	if layer.Size > 0 {
+		mediatype := layer.MediaType
+		layers := slices.DeleteFunc(ls.items, func(l *Layer) bool {
+			return l.MediaType == mediatype
+		})
+
+		ls.items = append(layers, layer)
+	}
+}
+
+type Layer struct {
+	MediaType string `json:"mediaType"`
+	Digest    string `json:"digest"`
+	Size      int64  `json:"size"`
+	From      string `json:"from,omitempty"`
+
+	tempFileName string
+}
+
+func NewLayer(r io.Reader, mediatype string) (*Layer, error) {
+	blobs, err := GetBlobsPath("")
+	if err != nil {
+		return nil, err
+	}
+
+	delimiter := ":"
+	if runtime.GOOS == "windows" {
+		delimiter = "-"
+	}
+
+	pattern := strings.Join([]string{"sha256", "*-partial"}, delimiter)
+	temp, err := os.CreateTemp(blobs, pattern)
+	if err != nil {
+		return nil, err
+	}
+	defer temp.Close()
+
+	sha256sum := sha256.New()
+	n, err := io.Copy(io.MultiWriter(temp, sha256sum), r)
+	if err != nil {
+		return nil, err
+	}
+
+	return &Layer{
+		MediaType:    mediatype,
+		Digest:       fmt.Sprintf("sha256:%x", sha256sum.Sum(nil)),
+		Size:         n,
+		tempFileName: temp.Name(),
+	}, nil
+}
+
+func NewLayerFromLayer(digest, mediatype, from string) (*Layer, error) {
+	blob, err := GetBlobsPath(digest)
+	if err != nil {
+		return nil, err
+	}
+
+	fi, err := os.Stat(blob)
+	if err != nil {
+		return nil, err
+	}
+
+	return &Layer{
+		MediaType: mediatype,
+		Digest:    digest,
+		Size:      fi.Size(),
+		From:      from,
+	}, nil
+}
+
+func (l *Layer) Commit() (bool, error) {
+	// always remove temp
+	defer os.Remove(l.tempFileName)
+
+	blob, err := GetBlobsPath(l.Digest)
+	if err != nil {
+		return false, err
+	}
+
+	if _, err := os.Stat(blob); err != nil {
+		return true, os.Rename(l.tempFileName, blob)
+	}
+
+	return false, nil
+}

server/manifests.go

@@ -0,0 +1,34 @@
+package server
+
+import (
+	"bytes"
+	"encoding/json"
+	"os"
+	"path/filepath"
+)
+
+func WriteManifest(name string, config *Layer, layers []*Layer) error {
+	manifest := ManifestV2{
+		SchemaVersion: 2,
+		MediaType:     "application/vnd.docker.distribution.manifest.v2+json",
+		Config:        config,
+		Layers:        layers,
+	}
+
+	var b bytes.Buffer
+	if err := json.NewEncoder(&b).Encode(manifest); err != nil {
+		return err
+	}
+
+	modelpath := ParseModelPath(name)
+	manifestPath, err := modelpath.GetManifestPath()
+	if err != nil {
+		return err
+	}
+
+	if err := os.MkdirAll(filepath.Dir(manifestPath), 0755); err != nil {
+		return err
+	}
+
+	return os.WriteFile(manifestPath, b.Bytes(), 0644)
+}

server/modelpath.go

@@ -67,6 +67,20 @@ func ParseModelPath(name string) ModelPath {
 	return mp
 }
 
+var errModelPathInvalid = errors.New("invalid model path")
+
+func (mp ModelPath) Validate() error {
+	if mp.Repository == "" {
+		return fmt.Errorf("%w: model repository name is required", errModelPathInvalid)
+	}
+
+	if strings.Contains(mp.Tag, ":") {
+		return fmt.Errorf("%w: ':' (colon) is not allowed in tag names", errModelPathInvalid)
+	}
+
+	return nil
+}
+
 func (mp ModelPath) GetNamespaceRepository() string {
 	return fmt.Sprintf("%s/%s", mp.Namespace, mp.Repository)
 }

server/routes.go

@@ -26,6 +26,7 @@ import (
 
 	"github.com/jmorganca/ollama/api"
 	"github.com/jmorganca/ollama/llm"
+	"github.com/jmorganca/ollama/parser"
 	"github.com/jmorganca/ollama/version"
 )
 
@@ -58,17 +59,26 @@ var loaded struct {
 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(ctx context.Context, workDir string, model *Model, reqOpts map[string]interface{}, sessionDuration time.Duration) error {
+func load(c *gin.Context, modelName string, reqOpts map[string]interface{}, sessionDuration time.Duration) (*Model, error) {
+	model, err := GetModel(modelName)
+	if err != nil {
+		return nil, err
+	}
+
+	workDir := c.GetString("workDir")
+
 	opts := api.DefaultOptions()
 	if err := opts.FromMap(model.Options); err != nil {
 		log.Printf("could not load model options: %v", err)
-		return err
+		return nil, err
 	}
 
 	if err := opts.FromMap(reqOpts); err != nil {
-		return err
+		return nil, err
 	}
 
+	ctx := c.Request.Context()
+
 	// check if the loaded model is still running in a subprocess, in case something unexpected happened
 	if loaded.runner != nil {
 		if err := loaded.runner.Ping(ctx); err != nil {
@@ -95,7 +105,7 @@ func load(ctx context.Context, workDir string, model *Model, reqOpts map[string]
 			loaded.Options = nil
 		}
 
-		llmRunner, err := llm.New(workDir, model.ModelPath, model.AdapterPaths, opts)
+		llmRunner, err := llm.New(workDir, model.ModelPath, model.AdapterPaths, model.ProjectorPaths, opts)
 		if err != nil {
 			// some older models are not compatible with newer versions of llama.cpp
 			// show a generalized compatibility error until there is a better way to
@@ -104,7 +114,7 @@ func load(ctx context.Context, workDir string, model *Model, reqOpts map[string]
 				err = fmt.Errorf("%v: this model may be incompatible with your version of Ollama. If you previously pulled this model, try updating it by running `ollama pull %s`", err, model.ShortName)
 			}
 
-			return err
+			return nil, err
 		}
 
 		loaded.Model = model
@@ -138,7 +148,7 @@ func load(ctx context.Context, workDir string, model *Model, reqOpts map[string]
 	}
 
 	loaded.expireTimer.Reset(sessionDuration)
-	return nil
+	return model, nil
 }
 
 func GenerateHandler(c *gin.Context) {
@@ -146,9 +156,9 @@ func GenerateHandler(c *gin.Context) {
 	defer loaded.mu.Unlock()
 
 	checkpointStart := time.Now()
-
 	var req api.GenerateRequest
 	err := c.ShouldBindJSON(&req)
+
 	switch {
 	case errors.Is(err, io.EOF):
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
@@ -171,88 +181,150 @@ func GenerateHandler(c *gin.Context) {
 		return
 	}
 
-	model, err := GetModel(req.Model)
+	sessionDuration := defaultSessionDuration
+	model, err := load(c, req.Model, req.Options, sessionDuration)
 	if err != nil {
 		var pErr *fs.PathError
-		if errors.As(err, &pErr) {
+		switch {
+		case errors.As(err, &pErr):
 			c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found, try pulling it first", req.Model)})
-			return
+		case errors.Is(err, api.ErrInvalidOpts):
+			c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		default:
+			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 		}
-		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
 		return
 	}
 
-	workDir := c.GetString("workDir")
-
-	// TODO: set this duration from the request if specified
-	sessionDuration := defaultSessionDuration
-	if err := load(c.Request.Context(), workDir, model, req.Options, sessionDuration); err != nil {
-		if errors.Is(err, api.ErrInvalidOpts) {
-			c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
-			return
-		}
-		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+	// an empty request loads the model
+	if req.Prompt == "" && req.Template == "" && req.System == "" {
+		c.JSON(http.StatusOK, api.GenerateResponse{
+			CreatedAt: time.Now().UTC(),
+			Model:     req.Model,
+			Done:      true})
 		return
 	}
 
 	checkpointLoaded := time.Now()
 
-	prompt := req.Prompt
-	if !req.Raw {
-		prompt, err = model.Prompt(req)
+	var prompt string
+	switch {
+	case req.Raw:
+		prompt = req.Prompt
+	case req.Prompt != "":
+		if req.Template != "" {
+			// override the default model template
+			model.Template = req.Template
+		}
+
+		var rebuild strings.Builder
+		if req.Context != nil {
+			// TODO: context is deprecated, at some point the context logic within this conditional should be removed
+			prevCtx, err := loaded.runner.Decode(c.Request.Context(), req.Context)
+			if err != nil {
+				c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+				return
+			}
+
+			// Remove leading spaces from prevCtx if present
+			prevCtx = strings.TrimPrefix(prevCtx, " ")
+			rebuild.WriteString(prevCtx)
+		}
+		p, err := model.Prompt(PromptVars{
+			System: req.System,
+			Prompt: req.Prompt,
+			First:  len(req.Context) == 0,
+		})
 		if err != nil {
 			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 			return
 		}
+		rebuild.WriteString(p)
+		prompt = rebuild.String()
 	}
 
 	ch := make(chan any)
+	var generated strings.Builder
 	go func() {
 		defer close(ch)
-		// an empty request loads the model
-		if req.Prompt == "" && req.Template == "" && req.System == "" {
-			ch <- api.GenerateResponse{CreatedAt: time.Now().UTC(), Model: req.Model, Done: true}
-			return
-		}
 
-		fn := func(r api.GenerateResponse) {
+		fn := func(r llm.PredictResult) {
+			// Update model expiration
 			loaded.expireAt = time.Now().Add(sessionDuration)
 			loaded.expireTimer.Reset(sessionDuration)
 
-			r.Model = req.Model
-			r.CreatedAt = time.Now().UTC()
+			// Build up the full response
+			if _, err := generated.WriteString(r.Content); err != nil {
+				ch <- gin.H{"error": err.Error()}
+				return
+			}
+
+			resp := api.GenerateResponse{
+				Model:     req.Model,
+				CreatedAt: time.Now().UTC(),
+				Done:      r.Done,
+				Response:  r.Content,
+				Metrics: api.Metrics{
+					PromptEvalCount:    r.PromptEvalCount,
+					PromptEvalDuration: r.PromptEvalDuration,
+					EvalCount:          r.EvalCount,
+					EvalDuration:       r.EvalDuration,
+				},
+			}
+
 			if r.Done {
-				r.TotalDuration = time.Since(checkpointStart)
-				r.LoadDuration = checkpointLoaded.Sub(checkpointStart)
+				resp.TotalDuration = time.Since(checkpointStart)
+				resp.LoadDuration = checkpointLoaded.Sub(checkpointStart)
+
+				if !req.Raw {
+					embd, err := loaded.runner.Encode(c.Request.Context(), prompt+generated.String())
+					if err != nil {
+						ch <- gin.H{"error": err.Error()}
+						return
+					}
+					resp.Context = embd
+				}
 			}
 
-			if req.Raw {
-				// in raw mode the client must manage history on their own
-				r.Context = nil
-			}
-
-			ch <- r
+			ch <- resp
 		}
 
-		if err := loaded.runner.Predict(c.Request.Context(), req.Context, prompt, req.Format, fn); err != nil {
+		// Start prediction
+		predictReq := llm.PredictOpts{
+			Prompt: prompt,
+			Format: req.Format,
+			Images: req.Images,
+		}
+		if err := loaded.runner.Predict(c.Request.Context(), predictReq, fn); err != nil {
 			ch <- gin.H{"error": err.Error()}
 		}
 	}()
 
 	if req.Stream != nil && !*req.Stream {
-		var response api.GenerateResponse
-		generated := ""
+		// Accumulate responses into the final response
+		var final api.GenerateResponse
+		var sb strings.Builder
 		for resp := range ch {
-			if r, ok := resp.(api.GenerateResponse); ok {
-				generated += r.Response
-				response = r
-			} else {
-				c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+			switch r := resp.(type) {
+			case api.GenerateResponse:
+				sb.WriteString(r.Response)
+				final = r
+			case gin.H:
+				if errorMsg, ok := r["error"].(string); ok {
+					c.JSON(http.StatusInternalServerError, gin.H{"error": errorMsg})
+					return
+				} else {
+					c.JSON(http.StatusInternalServerError, gin.H{"error": "unexpected error format in response"})
+					return
+				}
+			default:
+				c.JSON(http.StatusInternalServerError, gin.H{"error": "unexpected error"})
 				return
 			}
 		}
-		response.Response = generated
-		c.JSON(http.StatusOK, response)
+
+		final.Response = sb.String()
+		c.JSON(http.StatusOK, final)
 		return
 	}
 
@@ -279,15 +351,18 @@ func EmbeddingHandler(c *gin.Context) {
 		return
 	}
 
-	model, err := GetModel(req.Model)
+	sessionDuration := defaultSessionDuration
+	_, err = load(c, req.Model, req.Options, sessionDuration)
 	if err != nil {
-		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
-		return
-	}
-
-	workDir := c.GetString("workDir")
-	if err := load(c.Request.Context(), workDir, model, req.Options, 5*time.Minute); err != nil {
-		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		var pErr *fs.PathError
+		switch {
+		case errors.As(err, &pErr):
+			c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found, try pulling it first", req.Model)})
+		case errors.Is(err, api.ErrInvalidOpts):
+			c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		default:
+			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		}
 		return
 	}
 
@@ -409,8 +484,36 @@ func CreateModelHandler(c *gin.Context) {
 		return
 	}
 
-	if req.Name == "" || req.Path == "" {
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "name and path are required"})
+	if req.Name == "" {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "name is required"})
+		return
+	}
+
+	if err := ParseModelPath(req.Name).Validate(); err != nil {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		return
+	}
+
+	if req.Path == "" && req.Modelfile == "" {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "path or modelfile are required"})
+		return
+	}
+
+	var modelfile io.Reader = strings.NewReader(req.Modelfile)
+	if req.Path != "" && req.Modelfile == "" {
+		mf, err := os.Open(req.Path)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("error reading modelfile: %s", err)})
+			return
+		}
+		defer mf.Close()
+
+		modelfile = mf
+	}
+
+	commands, err := parser.Parse(modelfile)
+	if err != nil {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
 		return
 	}
 
@@ -424,7 +527,7 @@ func CreateModelHandler(c *gin.Context) {
 		ctx, cancel := context.WithCancel(c.Request.Context())
 		defer cancel()
 
-		if err := CreateModel(ctx, req.Name, req.Path, fn); err != nil {
+		if err := CreateModel(ctx, req.Name, filepath.Dir(req.Path), commands, fn); err != nil {
 			ch <- gin.H{"error": err.Error()}
 		}
 	}()
@@ -513,10 +616,19 @@ func GetModelInfo(name string) (*api.ShowResponse, error) {
 		return nil, err
 	}
 
+	modelDetails := api.ModelDetails{
+		Format:            model.Config.ModelFormat,
+		Family:            model.Config.ModelFamily,
+		Families:          model.Config.ModelFamilies,
+		ParameterSize:     model.Config.ModelType,
+		QuantizationLevel: model.Config.FileType,
+	}
+
 	resp := &api.ShowResponse{
 		License:  strings.Join(model.License, "\n"),
 		System:   model.System,
 		Template: model.Template,
+		Details:  modelDetails,
 	}
 
 	mf, err := ShowModelfile(model)
@@ -566,25 +678,42 @@ func ListModelsHandler(c *gin.Context) {
 		return
 	}
 
+	modelResponse := func(modelName string) (api.ModelResponse, error) {
+		model, err := GetModel(modelName)
+		if err != nil {
+			return api.ModelResponse{}, err
+		}
+
+		modelDetails := api.ModelDetails{
+			Format:            model.Config.ModelFormat,
+			Family:            model.Config.ModelFamily,
+			Families:          model.Config.ModelFamilies,
+			ParameterSize:     model.Config.ModelType,
+			QuantizationLevel: model.Config.FileType,
+		}
+
+		return api.ModelResponse{
+			Name:    model.ShortName,
+			Size:    model.Size,
+			Digest:  model.Digest,
+			Details: modelDetails,
+		}, nil
+	}
+
 	walkFunc := func(path string, info os.FileInfo, _ error) error {
 		if !info.IsDir() {
 			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, digest, err := GetManifest(mp)
+			resp, err := modelResponse(tag)
 			if err != nil {
 				log.Printf("skipping file: %s", fp)
 				return nil
 			}
 
-			models = append(models, api.ModelResponse{
-				Name:       mp.GetShortTagname(),
-				Size:       manifest.GetTotalSize(),
-				Digest:     digest,
-				ModifiedAt: info.ModTime(),
-			})
+			resp.ModifiedAt = info.ModTime()
+			models = append(models, resp)
 		}
 
 		return nil
@@ -615,6 +744,11 @@ func CopyModelHandler(c *gin.Context) {
 		return
 	}
 
+	if err := ParseModelPath(req.Destination).Validate(); err != nil {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		return
+	}
+
 	if err := CopyModel(req.Source, req.Destination); err != nil {
 		if os.IsNotExist(err) {
 			c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found", req.Source)})
@@ -625,6 +759,41 @@ func CopyModelHandler(c *gin.Context) {
 	}
 }
 
+func HeadBlobHandler(c *gin.Context) {
+	path, err := GetBlobsPath(c.Param("digest"))
+	if err != nil {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		return
+	}
+
+	if _, err := os.Stat(path); err != nil {
+		c.AbortWithStatusJSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("blob %q not found", c.Param("digest"))})
+		return
+	}
+
+	c.Status(http.StatusOK)
+}
+
+func CreateBlobHandler(c *gin.Context) {
+	layer, err := NewLayer(c.Request.Body, "")
+	if err != nil {
+		c.AbortWithStatusJSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+
+	if layer.Digest != c.Param("digest") {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("digest mismatch, expected %q, got %q", c.Param("digest"), layer.Digest)})
+		return
+	}
+
+	if _, err := layer.Commit(); err != nil {
+		c.AbortWithStatusJSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+
+	c.Status(http.StatusCreated)
+}
+
 var defaultAllowOrigins = []string{
 	"localhost",
 	"127.0.0.1",
@@ -678,12 +847,15 @@ func Serve(ln net.Listener, allowOrigins []string) error {
 
 	r.POST("/api/pull", PullModelHandler)
 	r.POST("/api/generate", GenerateHandler)
+	r.POST("/api/chat", ChatHandler)
 	r.POST("/api/embeddings", EmbeddingHandler)
 	r.POST("/api/create", CreateModelHandler)
 	r.POST("/api/push", PushModelHandler)
 	r.POST("/api/copy", CopyModelHandler)
 	r.DELETE("/api/delete", DeleteModelHandler)
 	r.POST("/api/show", ShowModelHandler)
+	r.POST("/api/blobs/:digest", CreateBlobHandler)
+	r.HEAD("/api/blobs/:digest", HeadBlobHandler)
 
 	for _, method := range []string{http.MethodGet, http.MethodHead} {
 		r.Handle(method, "/", func(c *gin.Context) {
@@ -691,6 +863,9 @@ func Serve(ln net.Listener, allowOrigins []string) error {
 		})
 
 		r.Handle(method, "/api/tags", ListModelsHandler)
+		r.Handle(method, "/api/version", func(c *gin.Context) {
+			c.JSON(http.StatusOK, gin.H{"version": version.Version})
+		})
 	}
 
 	log.Printf("Listening on %s (version %s)", ln.Addr(), version.Version)
@@ -713,7 +888,7 @@ func Serve(ln net.Listener, allowOrigins []string) error {
 	if runtime.GOOS == "linux" {
 		// check compatibility to log warnings
 		if _, err := llm.CheckVRAM(); err != nil {
-			log.Printf("Warning: GPU support may not be enabled, check you have installed GPU drivers: %v", err)
+			log.Print(err.Error())
 		}
 	}
 
@@ -769,3 +944,136 @@ func streamResponse(c *gin.Context, ch chan any) {
 		return true
 	})
 }
+
+func ChatHandler(c *gin.Context) {
+	loaded.mu.Lock()
+	defer loaded.mu.Unlock()
+
+	checkpointStart := time.Now()
+
+	var req api.ChatRequest
+	err := c.ShouldBindJSON(&req)
+	switch {
+	case errors.Is(err, io.EOF):
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
+		return
+	case err != nil:
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		return
+	}
+
+	// validate the request
+	switch {
+	case req.Model == "":
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "model is required"})
+		return
+	case len(req.Format) > 0 && req.Format != "json":
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "format must be json"})
+		return
+	}
+
+	sessionDuration := defaultSessionDuration
+	model, err := load(c, req.Model, req.Options, sessionDuration)
+	if err != nil {
+		var pErr *fs.PathError
+		switch {
+		case errors.As(err, &pErr):
+			c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found, try pulling it first", req.Model)})
+		case errors.Is(err, api.ErrInvalidOpts):
+			c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		default:
+			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		}
+		return
+	}
+
+	// an empty request loads the model
+	if len(req.Messages) == 0 {
+		c.JSON(http.StatusOK, api.ChatResponse{CreatedAt: time.Now().UTC(), Model: req.Model, Done: true})
+		return
+	}
+
+	checkpointLoaded := time.Now()
+
+	prompt, images, err := model.ChatPrompt(req.Messages)
+	if err != nil {
+		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		return
+	}
+
+	ch := make(chan any)
+
+	go func() {
+		defer close(ch)
+
+		fn := func(r llm.PredictResult) {
+			// Update model expiration
+			loaded.expireAt = time.Now().Add(sessionDuration)
+			loaded.expireTimer.Reset(sessionDuration)
+
+			resp := api.ChatResponse{
+				Model:     req.Model,
+				CreatedAt: time.Now().UTC(),
+				Done:      r.Done,
+				Metrics: api.Metrics{
+					PromptEvalCount:    r.PromptEvalCount,
+					PromptEvalDuration: r.PromptEvalDuration,
+					EvalCount:          r.EvalCount,
+					EvalDuration:       r.EvalDuration,
+				},
+			}
+
+			if r.Done {
+				resp.TotalDuration = time.Since(checkpointStart)
+				resp.LoadDuration = checkpointLoaded.Sub(checkpointStart)
+			} else {
+				resp.Message = &api.Message{Role: "assistant", Content: r.Content}
+			}
+
+			ch <- resp
+		}
+
+		// Start prediction
+		predictReq := llm.PredictOpts{
+			Prompt: prompt,
+			Format: req.Format,
+			Images: images,
+		}
+		if err := loaded.runner.Predict(c.Request.Context(), predictReq, fn); err != nil {
+			ch <- gin.H{"error": err.Error()}
+		}
+	}()
+
+	if req.Stream != nil && !*req.Stream {
+		// Accumulate responses into the final response
+		var final api.ChatResponse
+		var sb strings.Builder
+		for resp := range ch {
+			switch r := resp.(type) {
+			case api.ChatResponse:
+				if r.Message != nil {
+					sb.WriteString(r.Message.Content)
+				}
+
+				final = r
+			case gin.H:
+				if errorMsg, ok := r["error"].(string); ok {
+					c.JSON(http.StatusInternalServerError, gin.H{"error": errorMsg})
+					return
+				} else {
+					c.JSON(http.StatusInternalServerError, gin.H{"error": "unexpected error format in response"})
+					return
+				}
+			default:
+				c.JSON(http.StatusInternalServerError, gin.H{"error": "unexpected error"})
+				return
+			}
+		}
+
+		final.Message = &api.Message{Role: "assistant", Content: sb.String()}
+		c.JSON(http.StatusOK, final)
+		return
+	}
+
+	streamResponse(c, ch)
+}

server/upload.go

@@ -8,6 +8,7 @@ import (
 	"hash"
 	"io"
 	"log"
+	"math"
 	"net/http"
 	"net/url"
 	"os"
@@ -35,6 +36,8 @@ type blobUpload struct {
 
 	context.CancelFunc
 
+	file *os.File
+
 	done       bool
 	err        error
 	references atomic.Int32
@@ -42,8 +45,8 @@ type blobUpload struct {
 
 const (
 	numUploadParts          = 64
-	minUploadPartSize int64 = 95 * 1000 * 1000
-	maxUploadPartSize int64 = 1000 * 1000 * 1000
+	minUploadPartSize int64 = 100 * format.MegaByte
+	maxUploadPartSize int64 = 1000 * format.MegaByte
 )
 
 func (b *blobUpload) Prepare(ctx context.Context, requestURL *url.URL, opts *RegistryOptions) error {
@@ -55,7 +58,7 @@ func (b *blobUpload) Prepare(ctx context.Context, requestURL *url.URL, opts *Reg
 	if b.From != "" {
 		values := requestURL.Query()
 		values.Add("mount", b.Digest)
-		values.Add("from", b.From)
+		values.Add("from", ParseModelPath(b.From).GetNamespaceRepository())
 		requestURL.RawQuery = values.Encode()
 	}
 
@@ -77,6 +80,14 @@ func (b *blobUpload) Prepare(ctx context.Context, requestURL *url.URL, opts *Reg
 
 	b.Total = fi.Size()
 
+	// http.StatusCreated indicates a blob has been mounted
+	// ref: https://distribution.github.io/distribution/spec/api/#cross-repository-blob-mount
+	if resp.StatusCode == http.StatusCreated {
+		b.Completed.Store(b.Total)
+		b.done = true
+		return nil
+	}
+
 	var size = b.Total / numUploadParts
 	switch {
 	case size < minUploadPartSize:
@@ -92,7 +103,7 @@ func (b *blobUpload) Prepare(ctx context.Context, requestURL *url.URL, opts *Reg
 		}
 
 		// set part.N to the current number of parts
-		b.Parts = append(b.Parts, blobUploadPart{blobUpload: b, N: len(b.Parts), Offset: offset, Size: size})
+		b.Parts = append(b.Parts, blobUploadPart{N: len(b.Parts), Offset: offset, Size: size})
 		offset += size
 	}
 
@@ -120,12 +131,12 @@ func (b *blobUpload) Run(ctx context.Context, opts *RegistryOptions) {
 		return
 	}
 
-	f, err := os.Open(p)
+	b.file, err = os.Open(p)
 	if err != nil {
 		b.err = err
 		return
 	}
-	defer f.Close()
+	defer b.file.Close()
 
 	g, inner := errgroup.WithContext(ctx)
 	g.SetLimit(numUploadParts)
@@ -137,15 +148,16 @@ func (b *blobUpload) Run(ctx context.Context, opts *RegistryOptions) {
 			g.Go(func() error {
 				var err error
 				for try := 0; try < maxRetries; try++ {
-					part.ReadSeeker = io.NewSectionReader(f, part.Offset, part.Size)
-					err = b.uploadChunk(inner, http.MethodPatch, requestURL, part, opts)
+					err = b.uploadPart(inner, http.MethodPatch, requestURL, part, opts)
 					switch {
 					case errors.Is(err, context.Canceled):
 						return err
 					case errors.Is(err, errMaxRetriesExceeded):
 						return err
 					case err != nil:
-						log.Printf("%s part %d attempt %d failed: %v, retrying", b.Digest[7:19], part.N, try, err)
+						sleep := time.Second * time.Duration(math.Pow(2, float64(try)))
+						log.Printf("%s part %d attempt %d failed: %v, retrying in %s", b.Digest[7:19], part.N, try, err, sleep)
+						time.Sleep(sleep)
 						continue
 					}
 
@@ -164,6 +176,7 @@ func (b *blobUpload) Run(ctx context.Context, opts *RegistryOptions) {
 
 	requestURL := <-b.nextURL
 
+	// calculate md5 checksum and add it to the commit request
 	var sb strings.Builder
 	for _, part := range b.Parts {
 		sb.Write(part.Sum(nil))
@@ -180,30 +193,43 @@ func (b *blobUpload) Run(ctx context.Context, opts *RegistryOptions) {
 	headers.Set("Content-Type", "application/octet-stream")
 	headers.Set("Content-Length", "0")
 
-	resp, err := makeRequestWithRetry(ctx, http.MethodPut, requestURL, headers, nil, opts)
-	if err != nil {
-		b.err = err
-		return
+	for try := 0; try < maxRetries; try++ {
+		var resp *http.Response
+		resp, err = makeRequestWithRetry(ctx, http.MethodPut, requestURL, headers, nil, opts)
+		if errors.Is(err, context.Canceled) {
+			break
+		} else if err != nil {
+			sleep := time.Second * time.Duration(math.Pow(2, float64(try)))
+			log.Printf("%s complete upload attempt %d failed: %v, retrying in %s", b.Digest[7:19], try, err, sleep)
+			time.Sleep(sleep)
+			continue
+		}
+		defer resp.Body.Close()
+		break
 	}
-	defer resp.Body.Close()
 
+	b.err = err
 	b.done = true
 }
 
-func (b *blobUpload) uploadChunk(ctx context.Context, method string, requestURL *url.URL, part *blobUploadPart, opts *RegistryOptions) error {
-	part.Reset()
-
+func (b *blobUpload) uploadPart(ctx context.Context, method string, requestURL *url.URL, part *blobUploadPart, opts *RegistryOptions) error {
 	headers := make(http.Header)
 	headers.Set("Content-Type", "application/octet-stream")
 	headers.Set("Content-Length", fmt.Sprintf("%d", part.Size))
-	headers.Set("X-Redirect-Uploads", "1")
 
 	if method == http.MethodPatch {
+		headers.Set("X-Redirect-Uploads", "1")
 		headers.Set("Content-Range", fmt.Sprintf("%d-%d", part.Offset, part.Offset+part.Size-1))
 	}
 
-	resp, err := makeRequest(ctx, method, requestURL, headers, io.TeeReader(part.ReadSeeker, io.MultiWriter(part, part.Hash)), opts)
+	sr := io.NewSectionReader(b.file, part.Offset, part.Size)
+
+	md5sum := md5.New()
+	w := &progressWriter{blobUpload: b}
+
+	resp, err := makeRequest(ctx, method, requestURL, headers, io.TeeReader(sr, io.MultiWriter(w, md5sum)), opts)
 	if err != nil {
+		w.Rollback()
 		return err
 	}
 	defer resp.Body.Close()
@@ -215,11 +241,13 @@ func (b *blobUpload) uploadChunk(ctx context.Context, method string, requestURL
 
 	nextURL, err := url.Parse(location)
 	if err != nil {
+		w.Rollback()
 		return err
 	}
 
 	switch {
 	case resp.StatusCode == http.StatusTemporaryRedirect:
+		w.Rollback()
 		b.nextURL <- nextURL
 
 		redirectURL, err := resp.Location()
@@ -227,15 +255,18 @@ func (b *blobUpload) uploadChunk(ctx context.Context, method string, requestURL
 			return err
 		}
 
+		// retry uploading to the redirect URL
 		for try := 0; try < maxRetries; try++ {
-			err = b.uploadChunk(ctx, http.MethodPut, redirectURL, part, nil)
+			err = b.uploadPart(ctx, http.MethodPut, redirectURL, part, nil)
 			switch {
 			case errors.Is(err, context.Canceled):
 				return err
 			case errors.Is(err, errMaxRetriesExceeded):
 				return err
 			case err != nil:
-				log.Printf("%s part %d attempt %d failed: %v, retrying", b.Digest[7:19], part.N, try, err)
+				sleep := time.Second * time.Duration(math.Pow(2, float64(try)))
+				log.Printf("%s part %d attempt %d failed: %v, retrying in %s", b.Digest[7:19], part.N, try, err, sleep)
+				time.Sleep(sleep)
 				continue
 			}
 
@@ -245,6 +276,7 @@ func (b *blobUpload) uploadChunk(ctx context.Context, method string, requestURL
 		return fmt.Errorf("%w: %w", errMaxRetriesExceeded, err)
 
 	case resp.StatusCode == http.StatusUnauthorized:
+		w.Rollback()
 		auth := resp.Header.Get("www-authenticate")
 		authRedir := ParseAuthRedirectString(auth)
 		token, err := getAuthToken(ctx, authRedir)
@@ -255,18 +287,20 @@ func (b *blobUpload) uploadChunk(ctx context.Context, method string, requestURL
 		opts.Token = token
 		fallthrough
 	case resp.StatusCode >= http.StatusBadRequest:
+		w.Rollback()
 		body, err := io.ReadAll(resp.Body)
 		if err != nil {
 			return err
 		}
 
-		return fmt.Errorf("http status %d %s: %s", resp.StatusCode, resp.Status, body)
+		return fmt.Errorf("http status %s: %s", resp.Status, body)
 	}
 
 	if method == http.MethodPatch {
 		b.nextURL <- nextURL
 	}
 
+	part.Hash = md5sum
 	return nil
 }
 
@@ -293,7 +327,7 @@ func (b *blobUpload) Wait(ctx context.Context, fn func(api.ProgressResponse)) er
 		}
 
 		fn(api.ProgressResponse{
-			Status:    fmt.Sprintf("uploading %s", b.Digest),
+			Status:    fmt.Sprintf("pushing %s", b.Digest[7:19]),
 			Digest:    b.Digest,
 			Total:     b.Total,
 			Completed: b.Completed.Load(),
@@ -311,25 +345,23 @@ type blobUploadPart struct {
 	Offset int64
 	Size   int64
 	hash.Hash
+}
 
+type progressWriter struct {
 	written int64
-
-	io.ReadSeeker
 	*blobUpload
 }
 
-func (p *blobUploadPart) Write(b []byte) (n int, err error) {
+func (p *progressWriter) Write(b []byte) (n int, err error) {
 	n = len(b)
 	p.written += int64(n)
 	p.Completed.Add(int64(n))
 	return n, nil
 }
 
-func (p *blobUploadPart) Reset() {
-	p.Seek(0, io.SeekStart)
-	p.Completed.Add(-int64(p.written))
+func (p *progressWriter) Rollback() {
+	p.Completed.Add(-p.written)
 	p.written = 0
-	p.Hash = md5.New()
 }
 
 func uploadBlob(ctx context.Context, mp ModelPath, layer *Layer, opts *RegistryOptions, fn func(api.ProgressResponse)) error {
@@ -344,7 +376,7 @@ func uploadBlob(ctx context.Context, mp ModelPath, layer *Layer, opts *RegistryO
 	default:
 		defer resp.Body.Close()
 		fn(api.ProgressResponse{
-			Status:    fmt.Sprintf("uploading %s", layer.Digest),
+			Status:    fmt.Sprintf("pushing %s", layer.Digest[7:19]),
 			Digest:    layer.Digest,
 			Total:     layer.Size,
 			Completed: layer.Size,