build with cmake

ignore nil map values

.gitignore

@@ -2,5 +2,8 @@
 .vscode
 .env
 .venv
+.swp
 dist
 ollama
+/ggml-metal.metal
+build

CMakeLists.txt

@@ -0,0 +1,40 @@
+cmake_minimum_required(VERSION 3.14)  # 3.11 or later for FetchContent, but some features might require newer versions
+
+project(llama_cpp)
+
+include(FetchContent)
+
+FetchContent_Declare(
+    llama_cpp_gguf
+    GIT_REPOSITORY    https://github.com/ggerganov/llama.cpp.git
+    GIT_TAG           6381d4e
+)
+
+FetchContent_Declare(
+    llama_cpp_ggml
+    GIT_REPOSITORY    https://github.com/ggerganov/llama.cpp.git
+    GIT_TAG           dadbed9
+)
+
+FetchContent_MakeAvailable(llama_cpp_ggml)
+
+add_subdirectory(${llama_cpp_ggml_SOURCE_DIR}/examples EXCLUDE_FROM_ALL)
+add_executable(llama_cpp ${llama_cpp_ggml_SOURCE_DIR}/examples/server/server.cpp)
+include_directories(${llama_cpp_ggml_SOURCE_DIR})
+include_directories(${llama_cpp_ggml_SOURCE_DIR}/examples)
+target_compile_features(llama_cpp PRIVATE cxx_std_11)
+target_link_libraries(llama_cpp PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+
+if (APPLE)
+    add_executable(llama_cpp_metal ${llama_cpp_ggml_SOURCE_DIR}/examples/server/server.cpp)
+    target_compile_options(llama_cpp_metal PRIVATE -DLLAMA_STATIC=ON -DLLAMA_METAL=ON -DGGML_USE_METAL=1)
+    target_compile_features(llama_cpp_metal PRIVATE cxx_std_11)
+    target_link_libraries(llama_cpp_metal PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+    configure_file(${llama_cpp_SOURCE_DIR}/ggml-metal.metal ${CMAKE_BINARY_DIR}/ggml-metal.metal COPYONLY)
+else()
+    add_executable(llama_cpp_cublas ${llama_cpp_ggml_SOURCE_DIR}/examples/server/server.cpp)
+    target_compile_definitions(llama_cpp_cublas PRIVATE -DLLAMA_STATIC=ON -DLLAMA_CUBLAS=ON)
+    target_compile_options(llama_cpp_cublas PRIVATE -DLLAMA_CUBLAS=ON -DLLAMA_STATIC=ON)
+    target_compile_features(llama_cpp_cublas PRIVATE cxx_std_11)
+    target_link_libraries(llama_cpp_cublas PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+endif()

README.md

@@ -1,76 +1,129 @@
-![ollama](https://github.com/jmorganca/ollama/assets/251292/961f99bb-251a-4eec-897d-1ba99997ad0f)
+<div align="center">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" height="200px" srcset="https://github.com/jmorganca/ollama/assets/3325447/56ea1849-1284-4645-8970-956de6e51c3c">
+    <img alt="logo" height="200px" src="https://github.com/jmorganca/ollama/assets/3325447/0d0b44e2-8f4a-4e99-9b52-a5c1c741c8f7">
+  </picture>
+</div>
 
 # Ollama
 
-Run large language models with `llama.cpp`.
+[![Discord](https://dcbadge.vercel.app/api/server/ollama?style=flat&compact=true)](https://discord.gg/ollama)
 
-> Note: certain models that can be run with Ollama are intended for research and/or non-commercial use only.
+Run, create, and share large language models (LLMs).
 
-### Features
+> Note: Ollama is in early preview. Please report any issues you find.
 
-- Download and run popular large language models
-- Switch between multiple models on the fly
-- Hardware acceleration where available (Metal, CUDA)
-- Fast inference server written in Go, powered by [llama.cpp](https://github.com/ggerganov/llama.cpp)
-- REST API to use with your application (python, typescript SDKs coming soon)
+## Download
 
-## Install
-
-- [Download](https://ollama.ai/download) for macOS with Apple Silicon (Intel coming soon)
-- Download for Windows (coming soon)
-
-You can also build the [binary from source](#building).
+- [Download](https://ollama.ai/download) for macOS
+- Download for Windows and Linux (coming soon)
+- Build [from source](#building)
 
 ## Quickstart
 
-Run a fast and simple model.
+To run and chat with [Llama 2](https://ai.meta.com/llama), the new model by Meta:
 
 ```
-ollama run orca
+ollama run llama2
 ```
 
-## Example models
+## Model library
 
-### 💬 Chat
+Ollama supports a list of open-source models available on [ollama.ai/library](https://ollama.ai/library "ollama model library") 
 
-Have a conversation.
+Here are some example open-source models that can be downloaded: 
+
+| Model                    | Parameters | Size  | Download                        |
+| ------------------------ | ---------- | ----- | ------------------------------- |
+| Llama2                   | 7B         | 3.8GB | `ollama pull llama2`            |
+| Llama2 13B               | 13B        | 7.3GB | `ollama pull llama2:13b`        |
+| Llama2 70B               | 70B        | 39GB  | `ollama pull llama2:70b`        |
+| Llama2 Uncensored        | 7B         | 3.8GB | `ollama pull llama2-uncensored` |
+| Orca Mini                | 3B         | 1.9GB | `ollama pull orca-mini`         |
+| Vicuna                   | 7B         | 3.8GB | `ollama pull vicuna`            |
+| Nous-Hermes              | 7B         | 3.8GB | `ollama pull nous-hermes`       |
+| Nous-Hermes 13B          | 13B        | 7.3GB | `ollama pull nous-hermes:13b`   |
+| Wizard Vicuna Uncensored | 13B        | 7.3GB | `ollama pull wizard-vicuna`     |
+
+> Note: You should have at least 8 GB of RAM to run the 3B models, 16 GB to run the 7B models, and 32 GB to run the 13B models.
+
+## Examples
+
+### Run a model
 
 ```
-ollama run vicuna "Why is the sky blue?"
+ollama run llama2
+>>> hi
+Hello! How can I help you today?
 ```
 
-### 🗺️ Instructions
-
-Get a helping hand.
+For multiline input, you can wrap text with `"""`:
 
 ```
-ollama run orca "Write an email to my boss."
+>>> """Hello,
+... world!
+... """
+I'm a basic program that prints the famous "Hello, world!" message to the console.
 ```
 
-### 🔎 Ask questions about documents
+### Create a custom model
 
-Send the contents of a document and ask questions about it.
+Pull a base model:
 
 ```
-ollama run nous-hermes "$(cat input.txt)", please summarize this story
+ollama pull llama2
 ```
 
-### 📖 Storytelling
+> To update a model to the latest version, run `ollama pull llama2` again. The model will be updated (if necessary).
 
-Venture into the unknown.
+Create a `Modelfile`:
 
 ```
-ollama run nous-hermes "Once upon a time"
+FROM llama2
+
+# set the temperature to 1 [higher is more creative, lower is more coherent]
+PARAMETER temperature 1
+
+# set the system prompt
+SYSTEM """
+You are Mario from Super Mario Bros. Answer as Mario, the assistant, only.
+"""
 ```
 
-## Advanced usage
-
-### Run a local model
+Next, create and run the model:
 
 ```
-ollama run ~/Downloads/vicuna-7b-v1.3.ggmlv3.q4_1.bin
+ollama create mario -f ./Modelfile
+ollama run mario
+>>> hi
+Hello! It's your friend Mario.
 ```
 
+For more examples, see the [examples](./examples) directory. For more information on creating a Modelfile, see the [Modelfile](./docs/modelfile.md) documentation.
+
+### Pull a model from the registry
+
+```
+ollama pull orca
+```
+
+### Listing local models
+
+```
+ollama list
+```
+
+## Model packages
+
+### Overview
+
+Ollama bundles model weights, configuration, and data into a single package, defined by a [Modelfile](./docs/modelfile.md).
+
+<picture>
+  <source media="(prefers-color-scheme: dark)" height="480" srcset="https://github.com/jmorganca/ollama/assets/251292/2fd96b5f-191b-45c1-9668-941cfad4eb70">
+  <img alt="logo" height="480" src="https://github.com/jmorganca/ollama/assets/251292/2fd96b5f-191b-45c1-9668-941cfad4eb70">
+</picture>
+
 ## Building
 
 ```
@@ -80,29 +133,34 @@ go build .
 To run it start the server:
 
 ```
-./ollama server &
+./ollama serve &
 ```
 
 Finally, run a model!
 
 ```
-./ollama run ~/Downloads/vicuna-7b-v1.3.ggmlv3.q4_1.bin
+./ollama run llama2
 ```
 
-## API Reference
+## REST API
 
-### `POST /api/pull`
+> See the [API documentation](./docs/api.md) for all endpoints.
 
-Download a model
+Ollama has an API for running and managing models. For example to generate text from a model:
 
 ```
-curl -X POST http://localhost:11343/api/pull -d '{"model": "orca"}'
+curl -X POST http://localhost:11434/api/generate -d '{
+  "model": "llama2",
+  "prompt":"Why is the sky blue?"
+}'
 ```
 
-### `POST /api/generate`
+## Tools using Ollama
 
-Complete a prompt
-
-```
-curl -X POST http://localhost:11434/api/generate -d '{"model": "orca", "prompt": "hello!"}'
-```
+- [LangChain](https://python.langchain.com/docs/integrations/llms/ollama) and [LangChain.js](https://js.langchain.com/docs/modules/model_io/models/llms/integrations/ollama) with a question-answering [example](https://js.langchain.com/docs/use_cases/question_answering/local_retrieval_qa).
+- [Continue](https://github.com/continuedev/continue) - embeds Ollama inside Visual Studio Code. The extension lets you highlight code to add to the prompt, ask questions in the sidebar, and generate code inline.
+- [LiteLLM](https://github.com/BerriAI/litellm) a lightweight python package to simplify LLM API calls
+- [Discord AI Bot](https://github.com/mekb-turtle/discord-ai-bot) - interact with Ollama as a chatbot on Discord.
+- [Raycast Ollama](https://github.com/MassimilianoPasquini97/raycast_ollama) - Raycast extension to use Ollama for local llama inference on Raycast.
+- [Simple HTML UI for Ollama](https://github.com/rtcfirefly/ollama-ui)
+- [Emacs client](https://github.com/zweifisch/ollama) for Ollama

api/client.go

@@ -9,10 +9,18 @@ import (
 	"io"
 	"net/http"
 	"net/url"
+	"os"
+	"strings"
+)
+
+const DefaultHost = "localhost:11434"
+
+var (
+	envHost = os.Getenv("OLLAMA_HOST")
 )
 
 type Client struct {
-	base    url.URL
+	Base    url.URL
 	HTTP    http.Client
 	Headers http.Header
 }
@@ -27,22 +35,40 @@ func checkError(resp *http.Response, body []byte) error {
 	err := json.Unmarshal(body, &apiError)
 	if err != nil {
 		// Use the full body as the message if we fail to decode a response.
-		apiError.Message = string(body)
+		apiError.ErrorMessage = string(body)
 	}
 
 	return apiError
 }
 
-func NewClient(hosts ...string) *Client {
-	host := "127.0.0.1:11434"
-	if len(hosts) > 0 {
-		host = hosts[0]
+// Host returns the default host to use for the client. It is determined in the following order:
+// 1. The OLLAMA_HOST environment variable
+// 2. The default host (localhost:11434)
+func Host() string {
+	if envHost != "" {
+		return envHost
+	}
+	return DefaultHost
+}
+
+// FromEnv creates a new client using Host() as the host. An error is returns
+// if the host is invalid.
+func FromEnv() (*Client, error) {
+	h := Host()
+	if !strings.HasPrefix(h, "http://") && !strings.HasPrefix(h, "https://") {
+		h = "http://" + h
 	}
 
-	return &Client{
-		base: url.URL{Scheme: "http", Host: host},
-		HTTP: http.Client{},
+	u, err := url.Parse(h)
+	if err != nil {
+		return nil, fmt.Errorf("could not parse host: %w", err)
 	}
+
+	if u.Port() == "" {
+		u.Host += ":11434"
+	}
+
+	return &Client{Base: *u, HTTP: http.Client{}}, nil
 }
 
 func (c *Client) do(ctx context.Context, method, path string, reqData, respData any) error {
@@ -57,7 +83,7 @@ func (c *Client) do(ctx context.Context, method, path string, reqData, respData
 		reqBody = bytes.NewReader(data)
 	}
 
-	url := c.base.JoinPath(path).String()
+	url := c.Base.JoinPath(path).String()
 
 	req, err := http.NewRequestWithContext(ctx, method, url, reqBody)
 	if err != nil {
@@ -92,7 +118,6 @@ func (c *Client) do(ctx context.Context, method, path string, reqData, respData
 		}
 	}
 	return nil
-
 }
 
 func (c *Client) stream(ctx context.Context, method, path string, data any, fn func([]byte) error) error {
@@ -106,7 +131,7 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 		buf = bytes.NewBuffer(bts)
 	}
 
-	request, err := http.NewRequestWithContext(ctx, method, c.base.JoinPath(path).String(), buf)
+	request, err := http.NewRequestWithContext(ctx, method, c.Base.JoinPath(path).String(), buf)
 	if err != nil {
 		return err
 	}
@@ -131,11 +156,15 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 			return fmt.Errorf("unmarshal: %w", err)
 		}
 
+		if errorResponse.Error != "" {
+			return fmt.Errorf(errorResponse.Error)
+		}
+
 		if response.StatusCode >= 400 {
 			return StatusError{
-				StatusCode: response.StatusCode,
-				Status:     response.Status,
-				Message:    errorResponse.Error,
+				StatusCode:   response.StatusCode,
+				Status:       response.Status,
+				ErrorMessage: errorResponse.Error,
 			}
 		}
 
@@ -160,11 +189,11 @@ func (c *Client) Generate(ctx context.Context, req *GenerateRequest, fn Generate
 	})
 }
 
-type PullProgressFunc func(PullProgress) error
+type PullProgressFunc func(ProgressResponse) error
 
 func (c *Client) Pull(ctx context.Context, req *PullRequest, fn PullProgressFunc) error {
 	return c.stream(ctx, http.MethodPost, "/api/pull", req, func(bts []byte) error {
-		var resp PullProgress
+		var resp ProgressResponse
 		if err := json.Unmarshal(bts, &resp); err != nil {
 			return err
 		}
@@ -173,11 +202,11 @@ func (c *Client) Pull(ctx context.Context, req *PullRequest, fn PullProgressFunc
 	})
 }
 
-type PushProgressFunc func(PushProgress) error
+type PushProgressFunc func(ProgressResponse) error
 
 func (c *Client) Push(ctx context.Context, req *PushRequest, fn PushProgressFunc) error {
 	return c.stream(ctx, http.MethodPost, "/api/push", req, func(bts []byte) error {
-		var resp PushProgress
+		var resp ProgressResponse
 		if err := json.Unmarshal(bts, &resp); err != nil {
 			return err
 		}
@@ -186,11 +215,11 @@ func (c *Client) Push(ctx context.Context, req *PushRequest, fn PushProgressFunc
 	})
 }
 
-type CreateProgressFunc func(CreateProgress) error
+type CreateProgressFunc func(ProgressResponse) error
 
 func (c *Client) Create(ctx context.Context, req *CreateRequest, fn CreateProgressFunc) error {
 	return c.stream(ctx, http.MethodPost, "/api/create", req, func(bts []byte) error {
-		var resp CreateProgress
+		var resp ProgressResponse
 		if err := json.Unmarshal(bts, &resp); err != nil {
 			return err
 		}
@@ -206,3 +235,24 @@ func (c *Client) List(ctx context.Context) (*ListResponse, error) {
 	}
 	return &lr, nil
 }
+
+func (c *Client) Copy(ctx context.Context, req *CopyRequest) error {
+	if err := c.do(ctx, http.MethodPost, "/api/copy", req, nil); err != nil {
+		return err
+	}
+	return nil
+}
+
+func (c *Client) Delete(ctx context.Context, req *DeleteRequest) error {
+	if err := c.do(ctx, http.MethodDelete, "/api/delete", req, nil); err != nil {
+		return err
+	}
+	return nil
+}
+
+func (c *Client) Heartbeat(ctx context.Context) error {
+	if err := c.do(ctx, http.MethodHead, "/", nil, nil); err != nil {
+		return err
+	}
+	return nil
+}

api/types.go

@@ -1,31 +1,56 @@
 package api
 
 import (
+	"encoding/json"
 	"fmt"
+	"log"
+	"math"
 	"os"
+	"reflect"
 	"runtime"
+	"strings"
 	"time"
 )
 
 type StatusError struct {
-	StatusCode int
-	Status     string
-	Message    string
+	StatusCode   int
+	Status       string
+	ErrorMessage string `json:"error"`
 }
 
 func (e StatusError) Error() string {
-	if e.Message != "" {
-		return fmt.Sprintf("%s: %s", e.Status, e.Message)
+	switch {
+	case e.Status != "" && e.ErrorMessage != "":
+		return fmt.Sprintf("%s: %s", e.Status, e.ErrorMessage)
+	case e.Status != "":
+		return e.Status
+	case e.ErrorMessage != "":
+		return e.ErrorMessage
+	default:
+		// this should not happen
+		return "something went wrong, please see the ollama server logs for details"
 	}
-	return e.Status
 }
 
 type GenerateRequest struct {
-	Model   string `json:"model"`
-	Prompt  string `json:"prompt"`
-	Context []int  `json:"context,omitempty"`
+	Model    string `json:"model"`
+	Prompt   string `json:"prompt"`
+	System   string `json:"system"`
+	Template string `json:"template"`
+	Context  []int  `json:"context,omitempty"`
 
-	Options `json:"options"`
+	Options map[string]interface{} `json:"options"`
+}
+
+type EmbeddingRequest struct {
+	Model  string `json:"model"`
+	Prompt string `json:"prompt"`
+
+	Options map[string]interface{} `json:"options"`
+}
+
+type EmbeddingResponse struct {
+	Embedding []float64 `json:"embedding"`
 }
 
 type CreateRequest struct {
@@ -33,38 +58,36 @@ type CreateRequest struct {
 	Path string `json:"path"`
 }
 
-type CreateProgress struct {
-	Status string `json:"status"`
+type DeleteRequest struct {
+	Name string `json:"name"`
+}
+
+type CopyRequest struct {
+	Source      string `json:"source"`
+	Destination string `json:"destination"`
 }
 
 type PullRequest struct {
 	Name     string `json:"name"`
+	Insecure bool   `json:"insecure,omitempty"`
 	Username string `json:"username"`
 	Password string `json:"password"`
 }
 
-type PullProgress struct {
-	Status    string  `json:"status"`
-	Digest    string  `json:"digest,omitempty"`
-	Total     int     `json:"total,omitempty"`
-	Completed int     `json:"completed,omitempty"`
-	Percent   float64 `json:"percent,omitempty"`
+type ProgressResponse struct {
+	Status    string `json:"status"`
+	Digest    string `json:"digest,omitempty"`
+	Total     int    `json:"total,omitempty"`
+	Completed int    `json:"completed,omitempty"`
 }
 
 type PushRequest struct {
 	Name     string `json:"name"`
+	Insecure bool   `json:"insecure,omitempty"`
 	Username string `json:"username"`
 	Password string `json:"password"`
 }
 
-type PushProgress struct {
-	Status    string  `json:"status"`
-	Digest    string  `json:"digest,omitempty"`
-	Total     int     `json:"total,omitempty"`
-	Completed int     `json:"completed,omitempty"`
-	Percent   float64 `json:"percent,omitempty"`
-}
-
 type ListResponse struct {
 	Models []ListResponseModel `json:"models"`
 }
@@ -75,6 +98,10 @@ type ListResponseModel struct {
 	Size       int       `json:"size"`
 }
 
+type TokenResponse struct {
+	Token string `json:"token"`
+}
+
 type GenerateResponse struct {
 	Model     string    `json:"model"`
 	CreatedAt time.Time `json:"created_at"`
@@ -84,6 +111,9 @@ type GenerateResponse struct {
 	Context []int `json:"context,omitempty"`
 
 	TotalDuration      time.Duration `json:"total_duration,omitempty"`
+	LoadDuration       time.Duration `json:"load_duration,omitempty"`
+	SampleCount        int           `json:"sample_count,omitempty"`
+	SampleDuration     time.Duration `json:"sample_duration,omitempty"`
 	PromptEvalCount    int           `json:"prompt_eval_count,omitempty"`
 	PromptEvalDuration time.Duration `json:"prompt_eval_duration,omitempty"`
 	EvalCount          int           `json:"eval_count,omitempty"`
@@ -95,6 +125,19 @@ func (r *GenerateResponse) Summary() {
 		fmt.Fprintf(os.Stderr, "total duration:       %v\n", r.TotalDuration)
 	}
 
+	if r.LoadDuration > 0 {
+		fmt.Fprintf(os.Stderr, "load duration:        %v\n", r.LoadDuration)
+	}
+
+	if r.SampleCount > 0 {
+		fmt.Fprintf(os.Stderr, "sample count:         %d token(s)\n", r.SampleCount)
+	}
+
+	if r.SampleDuration > 0 {
+		fmt.Fprintf(os.Stderr, "sample duration:      %s\n", r.SampleDuration)
+		fmt.Fprintf(os.Stderr, "sample rate:          %.2f tokens/s\n", float64(r.SampleCount)/r.SampleDuration.Seconds())
+	}
+
 	if r.PromptEvalCount > 0 {
 		fmt.Fprintf(os.Stderr, "prompt eval count:    %d token(s)\n", r.PromptEvalCount)
 	}
@@ -121,50 +164,142 @@ type Options struct {
 	UseNUMA bool `json:"numa,omitempty"`
 
 	// Model options
-	NumCtx        int  `json:"num_ctx,omitempty"`
-	NumBatch      int  `json:"num_batch,omitempty"`
-	NumGPU        int  `json:"num_gpu,omitempty"`
-	MainGPU       int  `json:"main_gpu,omitempty"`
-	LowVRAM       bool `json:"low_vram,omitempty"`
-	F16KV         bool `json:"f16_kv,omitempty"`
-	LogitsAll     bool `json:"logits_all,omitempty"`
-	VocabOnly     bool `json:"vocab_only,omitempty"`
-	UseMMap       bool `json:"use_mmap,omitempty"`
-	UseMLock      bool `json:"use_mlock,omitempty"`
-	EmbeddingOnly bool `json:"embedding_only,omitempty"`
+	NumCtx             int     `json:"num_ctx,omitempty"`
+	NumKeep            int     `json:"num_keep,omitempty"`
+	NumBatch           int     `json:"num_batch,omitempty"`
+	NumGQA             int     `json:"num_gqa,omitempty"`
+	NumGPU             int     `json:"num_gpu,omitempty"`
+	MainGPU            int     `json:"main_gpu,omitempty"`
+	LowVRAM            bool    `json:"low_vram,omitempty"`
+	F16KV              bool    `json:"f16_kv,omitempty"`
+	LogitsAll          bool    `json:"logits_all,omitempty"`
+	VocabOnly          bool    `json:"vocab_only,omitempty"`
+	UseMMap            bool    `json:"use_mmap,omitempty"`
+	UseMLock           bool    `json:"use_mlock,omitempty"`
+	EmbeddingOnly      bool    `json:"embedding_only,omitempty"`
+	RopeFrequencyBase  float32 `json:"rope_frequency_base,omitempty"`
+	RopeFrequencyScale float32 `json:"rope_frequency_scale,omitempty"`
 
 	// Predict options
-	RepeatLastN      int     `json:"repeat_last_n,omitempty"`
-	RepeatPenalty    float32 `json:"repeat_penalty,omitempty"`
-	FrequencyPenalty float32 `json:"frequency_penalty,omitempty"`
-	PresencePenalty  float32 `json:"presence_penalty,omitempty"`
-	Temperature      float32 `json:"temperature,omitempty"`
-	TopK             int     `json:"top_k,omitempty"`
-	TopP             float32 `json:"top_p,omitempty"`
-	TFSZ             float32 `json:"tfs_z,omitempty"`
-	TypicalP         float32 `json:"typical_p,omitempty"`
-	Mirostat         int     `json:"mirostat,omitempty"`
-	MirostatTau      float32 `json:"mirostat_tau,omitempty"`
-	MirostatEta      float32 `json:"mirostat_eta,omitempty"`
+	RepeatLastN      int      `json:"repeat_last_n,omitempty"`
+	RepeatPenalty    float32  `json:"repeat_penalty,omitempty"`
+	FrequencyPenalty float32  `json:"frequency_penalty,omitempty"`
+	PresencePenalty  float32  `json:"presence_penalty,omitempty"`
+	Temperature      float32  `json:"temperature,omitempty"`
+	TopK             int      `json:"top_k,omitempty"`
+	TopP             float32  `json:"top_p,omitempty"`
+	TFSZ             float32  `json:"tfs_z,omitempty"`
+	TypicalP         float32  `json:"typical_p,omitempty"`
+	Mirostat         int      `json:"mirostat,omitempty"`
+	MirostatTau      float32  `json:"mirostat_tau,omitempty"`
+	MirostatEta      float32  `json:"mirostat_eta,omitempty"`
+	PenalizeNewline  bool     `json:"penalize_newline,omitempty"`
+	Stop             []string `json:"stop,omitempty"`
 
 	NumThread int `json:"num_thread,omitempty"`
 }
 
+func (opts *Options) FromMap(m map[string]interface{}) error {
+	valueOpts := reflect.ValueOf(opts).Elem() // names of the fields in the options struct
+	typeOpts := reflect.TypeOf(opts).Elem()   // types of the fields in the options struct
+
+	// 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
+		}
+	}
+
+	for key, val := range m {
+		if opt, ok := jsonOpts[key]; ok {
+			field := valueOpts.FieldByName(opt.Name)
+			if field.IsValid() && field.CanSet() {
+				if val == nil {
+					continue
+				}
+
+				switch field.Kind() {
+				case reflect.Int:
+					switch t := val.(type) {
+					case int64:
+						field.SetInt(t)
+					case float64:
+						// when JSON unmarshals numbers, it uses float64, not int
+						field.SetInt(int64(t))
+					default:
+						log.Printf("could not convert model parameter %v to int, skipped", key)
+					}
+				case reflect.Bool:
+					val, ok := val.(bool)
+					if !ok {
+						log.Printf("could not convert model parameter %v to bool, skipped", key)
+						continue
+					}
+					field.SetBool(val)
+				case reflect.Float32:
+					// JSON unmarshals to float64
+					val, ok := val.(float64)
+					if !ok {
+						log.Printf("could not convert model parameter %v to float32, skipped", key)
+						continue
+					}
+					field.SetFloat(val)
+				case reflect.String:
+					val, ok := val.(string)
+					if !ok {
+						log.Printf("could not convert model parameter %v to string, skipped", key)
+						continue
+					}
+					field.SetString(val)
+				case reflect.Slice:
+					// JSON unmarshals to []interface{}, not []string
+					val, ok := val.([]interface{})
+					if !ok {
+						log.Printf("could not convert model parameter %v to slice, skipped", key)
+						continue
+					}
+					// convert []interface{} to []string
+					slice := make([]string, len(val))
+					for i, item := range val {
+						str, ok := item.(string)
+						if !ok {
+							log.Printf("could not convert model parameter %v to slice of strings, skipped", key)
+							continue
+						}
+						slice[i] = str
+					}
+					field.Set(reflect.ValueOf(slice))
+				default:
+					return fmt.Errorf("unknown type loading config params: %v", field.Kind())
+				}
+			}
+		}
+	}
+	return nil
+}
+
 func DefaultOptions() Options {
 	return Options{
 		Seed: -1,
 
 		UseNUMA: false,
 
-		NumCtx:   2048,
-		NumBatch: 512,
-		NumGPU:   1,
-		LowVRAM:  false,
-		F16KV:    true,
-		UseMMap:  true,
-		UseMLock: false,
+		NumCtx:             2048,
+		NumKeep:            -1,
+		NumBatch:           512,
+		NumGPU:             1,
+		NumGQA:             1,
+		LowVRAM:            false,
+		F16KV:              true,
+		UseMMap:            true,
+		UseMLock:           false,
+		RopeFrequencyBase:  10000.0,
+		RopeFrequencyScale: 1.0,
+		EmbeddingOnly:      true,
 
-		RepeatLastN:      512,
+		RepeatLastN:      64,
 		RepeatPenalty:    1.1,
 		FrequencyPenalty: 0.0,
 		PresencePenalty:  0.0,
@@ -176,7 +311,37 @@ func DefaultOptions() Options {
 		Mirostat:         0,
 		MirostatTau:      5.0,
 		MirostatEta:      0.1,
+		PenalizeNewline:  true,
 
 		NumThread: runtime.NumCPU(),
 	}
 }
+
+type Duration struct {
+	time.Duration
+}
+
+func (d *Duration) UnmarshalJSON(b []byte) (err error) {
+	var v any
+	if err := json.Unmarshal(b, &v); err != nil {
+		return err
+	}
+
+	d.Duration = 5 * time.Minute
+
+	switch t := v.(type) {
+	case float64:
+		if t < 0 {
+			t = math.MaxFloat64
+		}
+
+		d.Duration = time.Duration(t)
+	case string:
+		d.Duration, err = time.ParseDuration(t)
+		if err != nil {
+			return err
+		}
+	}
+
+	return nil
+}

app/README.md

@@ -1,7 +1,5 @@
 # Desktop
 
-_Note: the Ollama desktop app is a work in progress and is not ready yet for general use._
-
 This app builds upon Ollama to provide a desktop experience for running models.
 
 ## Developing
@@ -9,19 +7,15 @@ This app builds upon Ollama to provide a desktop experience for running models.
 First, build the `ollama` binary:
 
 ```
-make -C ..
+cd ..
+go build .
 ```
 
 Then run the desktop app with `npm start`:
 
 ```
+cd app
 npm install
 npm start
 ```
 
-## Coming soon
-
-- Browse the latest available models on Hugging Face and other sources
-- Keep track of previous conversations with models
-- Switch quickly between models
-- Connect to remote Ollama servers to run models

app/forge.config.ts

@@ -18,10 +18,16 @@ const config: ForgeConfig = {
     asar: true,
     icon: './assets/icon.icns',
     extraResource: [
-      '../ollama',
-      path.join(__dirname, './assets/ollama_icon_16x16Template.png'),
-      path.join(__dirname, './assets/ollama_icon_16x16Template@2x.png'),
-      ...(process.platform === 'darwin' ? ['../llama/ggml-metal.metal'] : []),
+      '../dist/ollama',
+      path.join(__dirname, './assets/iconTemplate.png'),
+      path.join(__dirname, './assets/iconTemplate@2x.png'),
+      path.join(__dirname, './assets/iconUpdateTemplate.png'),
+      path.join(__dirname, './assets/iconUpdateTemplate@2x.png'),
+      path.join(__dirname, './assets/iconDarkTemplate.png'),
+      path.join(__dirname, './assets/iconDarkTemplate@2x.png'),
+      path.join(__dirname, './assets/iconDarkUpdateTemplate.png'),
+      path.join(__dirname, './assets/iconDarkUpdateTemplate@2x.png'),
+      ...(process.platform === 'darwin' ? ['../llm/ggml-metal.metal'] : []),
     ],
     ...(process.env.SIGN
       ? {
@@ -36,6 +42,9 @@ const config: ForgeConfig = {
           },
         }
       : {}),
+    osxUniversal: {
+      x64ArchFiles: '**/ollama',
+    },
   },
   rebuildConfig: {},
   makers: [new MakerSquirrel({}), new MakerZIP({}, ['darwin'])],

app/package-lock.json

@@ -32,6 +32,7 @@
         "@electron-forge/plugin-auto-unpack-natives": "^6.2.1",
         "@electron-forge/plugin-webpack": "^6.2.1",
         "@electron-forge/publisher-github": "^6.2.1",
+        "@electron/universal": "^1.4.1",
         "@svgr/webpack": "^8.0.1",
         "@types/chmodr": "^1.0.0",
         "@types/node": "^20.4.0",
@@ -3328,9 +3329,9 @@
       }
     },
     "node_modules/@electron/universal": {
-      "version": "1.3.4",
-      "resolved": "https://registry.npmjs.org/@electron/universal/-/universal-1.3.4.tgz",
-      "integrity": "sha512-BdhBgm2ZBnYyYRLRgOjM5VHkyFItsbggJ0MHycOjKWdFGYwK97ZFXH54dTvUWEfha81vfvwr5On6XBjt99uDcg==",
+      "version": "1.4.1",
+      "resolved": "https://registry.npmjs.org/@electron/universal/-/universal-1.4.1.tgz",
+      "integrity": "sha512-lE/U3UNw1YHuowNbTmKNs9UlS3En3cPgwM5MI+agIgr/B1hSze9NdOP0qn7boZaI9Lph8IDv3/24g9IxnJP7aQ==",
       "dev": true,
       "dependencies": {
         "@electron/asar": "^3.2.1",

app/package.json

@@ -6,12 +6,14 @@
   "main": ".webpack/main",
   "scripts": {
     "start": "electron-forge start",
-    "package": "electron-forge package",
-    "package:sign": "SIGN=1 electron-forge package",
-    "make": "electron-forge make",
-    "make:sign": "SIGN=1 electron-forge make",
+    "package": "electron-forge package --arch universal",
+    "package:sign": "SIGN=1 electron-forge package --arch universal",
+    "make": "electron-forge make --arch universal",
+    "make:sign": "SIGN=1 electron-forge make --arch universal",
     "publish": "SIGN=1 electron-forge publish",
-    "lint": "eslint --ext .ts,.tsx ."
+    "lint": "eslint --ext .ts,.tsx .",
+    "format": "prettier --check . --ignore-path .gitignore",
+    "format:fix": "prettier --write . --ignore-path .gitignore"
   },
   "keywords": [],
   "author": {
@@ -30,6 +32,7 @@
     "@electron-forge/plugin-auto-unpack-natives": "^6.2.1",
     "@electron-forge/plugin-webpack": "^6.2.1",
     "@electron-forge/publisher-github": "^6.2.1",
+    "@electron/universal": "^1.4.1",
     "@svgr/webpack": "^8.0.1",
     "@types/chmodr": "^1.0.0",
     "@types/node": "^20.4.0",

app/src/app.tsx

@@ -2,7 +2,7 @@ import { useState } from 'react'
 import copy from 'copy-to-clipboard'
 import { CheckIcon, DocumentDuplicateIcon } from '@heroicons/react/24/outline'
 import Store from 'electron-store'
-import { getCurrentWindow } from '@electron/remote'
+import { getCurrentWindow, app } from '@electron/remote'
 
 import { install } from './install'
 import OllamaIcon from './ollama.svg'
@@ -19,7 +19,7 @@ export default function () {
   const [step, setStep] = useState<Step>(Step.WELCOME)
   const [commandCopied, setCommandCopied] = useState<boolean>(false)
 
-  const command = 'ollama run orca'
+  const command = 'ollama run llama2'
 
   return (
     <div className='drag'>
@@ -51,10 +51,15 @@ export default function () {
               <div className='mx-auto'>
                 <button
                   onClick={async () => {
-                    await install()
-                    getCurrentWindow().show()
-                    getCurrentWindow().focus()
-                    setStep(Step.FINISH)
+                    try {
+                      await install()
+                      setStep(Step.FINISH)
+                    } catch (e) {
+                      console.error('could not install: ', e)
+                    } finally {
+                      getCurrentWindow().show()
+                      getCurrentWindow().focus()
+                    }
                   }}
                   className='no-drag rounded-dm mx-auto w-[60%] rounded-md bg-black px-4 py-2 text-sm text-white hover:brightness-110'
                 >
@@ -77,7 +82,11 @@ export default function () {
                     {command}
                   </pre>
                   <button
-                    className={`no-drag absolute right-[5px] px-2 py-2 ${commandCopied ? 'text-gray-900 opacity-100 hover:cursor-auto' : 'text-gray-200 opacity-50 hover:cursor-pointer'} hover:text-gray-900 hover:font-bold group-hover:opacity-100`}
+                    className={`no-drag absolute right-[5px] px-2 py-2 ${
+                      commandCopied
+                        ? 'text-gray-900 opacity-100 hover:cursor-auto'
+                        : 'text-gray-200 opacity-50 hover:cursor-pointer'
+                    } hover:font-bold hover:text-gray-900 group-hover:opacity-100`}
                     onClick={() => {
                       copy(command)
                       setCommandCopied(true)
@@ -85,13 +94,15 @@ export default function () {
                     }}
                   >
                     {commandCopied ? (
-                      <CheckIcon className='h-4 w-4 text-gray-500 font-bold' />
+                      <CheckIcon className='h-4 w-4 font-bold text-gray-500' />
                     ) : (
                       <DocumentDuplicateIcon className='h-4 w-4 text-gray-500' />
                     )}
                   </button>
                 </div>
-                <p className='mx-auto my-4 w-[70%] text-xs text-gray-400'>Run this command in your favorite terminal.</p>
+                <p className='mx-auto my-4 w-[70%] text-xs text-gray-400'>
+                  Run this command in your favorite terminal.
+                </p>
               </div>
               <button
                 onClick={() => {

app/src/declarations.d.ts

@@ -1,4 +1,4 @@
 declare module '*.svg' {
-  const content: string;
-  export default content;
-}
+  const content: string
+  export default content
+}

app/src/index.ts

@@ -1,5 +1,5 @@
-import { spawn } from 'child_process'
-import { app, autoUpdater, dialog, Tray, Menu, BrowserWindow } from 'electron'
+import { spawn, ChildProcess } from 'child_process'
+import { app, autoUpdater, dialog, Tray, Menu, BrowserWindow, MenuItemConstructorOptions, nativeTheme } from 'electron'
 import Store from 'electron-store'
 import winston from 'winston'
 import 'winston-daily-rotate-file'
@@ -10,8 +10,12 @@ import { installed } from './install'
 
 require('@electron/remote/main').initialize()
 
+if (require('electron-squirrel-startup')) {
+  app.quit()
+}
+
 const store = new Store()
-let tray: Tray | null = null
+
 let welcomeWindow: BrowserWindow | null = null
 
 declare const MAIN_WINDOW_WEBPACK_ENTRY: string
@@ -28,10 +32,30 @@ const logger = winston.createLogger({
   format: winston.format.printf(info => info.message),
 })
 
-const SingleInstanceLock = app.requestSingleInstanceLock()
-if (!SingleInstanceLock) {
-  app.quit()
-}
+app.on('ready', () => {
+  const gotTheLock = app.requestSingleInstanceLock()
+  if (!gotTheLock) {
+    app.exit(0)
+    return
+  }
+
+  app.on('second-instance', () => {
+    if (app.hasSingleInstanceLock()) {
+      app.releaseSingleInstanceLock()
+    }
+
+    if (proc) {
+      proc.off('exit', restart)
+      proc.kill()
+    }
+
+    app.exit(0)
+  })
+
+  app.focus({ steal: true })
+
+  init()
+})
 
 function firstRunWindow() {
   // Create the browser window.
@@ -47,49 +71,74 @@ function firstRunWindow() {
       nodeIntegration: true,
       contextIsolation: false,
     },
-    alwaysOnTop: true,
   })
 
   require('@electron/remote/main').enable(welcomeWindow.webContents)
 
-  // and load the index.html of the app.
   welcomeWindow.loadURL(MAIN_WINDOW_WEBPACK_ENTRY)
-
   welcomeWindow.on('ready-to-show', () => welcomeWindow.show())
+  welcomeWindow.on('closed', () => {
+    if (process.platform === 'darwin') {
+      app.dock.hide()
+    }
+  })
+}
 
-  // for debugging
-  // welcomeWindow.webContents.openDevTools()
+let tray: Tray | null = null
+let updateAvailable = false
+const assetPath = app.isPackaged ? process.resourcesPath : path.join(__dirname, '..', '..', 'assets')
 
-  if (process.platform === 'darwin') {
-    app.dock.hide()
+function trayIconPath() {
+  return nativeTheme.shouldUseDarkColors
+    ? updateAvailable
+      ? path.join(assetPath, 'iconDarkUpdateTemplate.png')
+      : path.join(assetPath, 'iconDarkTemplate.png')
+    : updateAvailable
+    ? path.join(assetPath, 'iconUpdateTemplate.png')
+    : path.join(assetPath, 'iconTemplate.png')
+}
+
+function updateTrayIcon() {
+  if (tray) {
+    tray.setImage(trayIconPath())
   }
 }
 
-function createSystemtray() {
-  let iconPath = path.join(__dirname, '..', '..', 'assets', 'ollama_icon_16x16Template.png')
+function updateTray() {
+  const updateItems: MenuItemConstructorOptions[] = [
+    { label: 'An update is available', enabled: false },
+    {
+      label: 'Restart to update',
+      click: () => autoUpdater.quitAndInstall(),
+    },
+    { type: 'separator' },
+  ]
 
-  if (app.isPackaged) {
-    iconPath = path.join(process.resourcesPath, 'ollama_icon_16x16Template.png')
+  const menu = Menu.buildFromTemplate([
+    ...(updateAvailable ? updateItems : []),
+    { role: 'quit', label: 'Quit Ollama', accelerator: 'Command+Q' },
+  ])
+
+  if (!tray) {
+    tray = new Tray(trayIconPath())
   }
 
-  tray = new Tray(iconPath)
+  tray.setToolTip(updateAvailable ? 'An update is available' : 'Ollama')
+  tray.setContextMenu(menu)
+  tray.setImage(trayIconPath())
 
-  const contextMenu = Menu.buildFromTemplate([{ role: 'quit', label: 'Quit Ollama', accelerator: 'Command+Q' }])
-
-  tray.setContextMenu(contextMenu)
-  tray.setToolTip('Ollama')
+  nativeTheme.off('updated', updateTrayIcon)
+  nativeTheme.on('updated', updateTrayIcon)
 }
 
-if (require('electron-squirrel-startup')) {
-  app.quit()
-}
+let proc: ChildProcess = null
 
 function server() {
   const binary = app.isPackaged
     ? path.join(process.resourcesPath, 'ollama')
     : path.resolve(process.cwd(), '..', 'ollama')
 
-  const proc = spawn(binary, ['serve'])
+  proc = spawn(binary, ['serve'])
 
   proc.stdout.on('data', data => {
     logger.info(data.toString().trim())
@@ -99,24 +148,32 @@ function server() {
     logger.error(data.toString().trim())
   })
 
-  function restart() {
-    logger.info('Restarting the server...')
-    server()
-  }
-
   proc.on('exit', restart)
+}
 
-  app.on('before-quit', () => {
+function restart() {
+  setTimeout(server, 1000)
+}
+
+app.on('before-quit', () => {
+  if (proc) {
     proc.off('exit', restart)
     proc.kill()
-  })
-}
+  }
+})
 
-if (process.platform === 'darwin') {
-  app.dock.hide()
-}
+function init() {
+  if (app.isPackaged) {
+    heartbeat()
+    autoUpdater.checkForUpdates()
+    setInterval(() => {
+      heartbeat()
+      autoUpdater.checkForUpdates()
+    }, 60 * 60 * 1000)
+  }
+
+  updateTray()
 
-app.on('ready', () => {
   if (process.platform === 'darwin') {
     if (app.isPackaged) {
       if (!app.isInApplicationsFolder()) {
@@ -152,10 +209,13 @@ app.on('ready', () => {
     }
   }
 
-  createSystemtray()
   server()
 
   if (store.get('first-time-run') && installed()) {
+    if (process.platform === 'darwin') {
+      app.dock.hide()
+    }
+
     app.setLoginItemSettings({ openAtLogin: app.getLoginItemSettings().openAtLogin })
     return
   }
@@ -163,7 +223,7 @@ app.on('ready', () => {
   // This is the first run or the CLI is no longer installed
   app.setLoginItemSettings({ openAtLogin: true })
   firstRunWindow()
-})
+}
 
 // Quit when all windows are closed, except on macOS. There, it's common
 // for applications and their menu bar to stay active until the user quits
@@ -176,13 +236,18 @@ app.on('window-all-closed', () => {
 
 // In this file you can include the rest of your app's specific main process
 // code. You can also put them in separate files and import them here.
+let aid = ''
+try {
+  aid = id()
+} catch (e) {}
+
 autoUpdater.setFeedURL({
-  url: `https://ollama.ai/api/update?os=${process.platform}&arch=${process.arch}&version=${app.getVersion()}`,
+  url: `https://ollama.ai/api/update?os=${process.platform}&arch=${process.arch}&version=${app.getVersion()}&id=${aid}`,
 })
 
 async function heartbeat() {
   analytics.track({
-    anonymousId: id(),
+    anonymousId: aid,
     event: 'heartbeat',
     properties: {
       version: app.getVersion(),
@@ -190,29 +255,11 @@ async function heartbeat() {
   })
 }
 
-if (app.isPackaged) {
-  heartbeat()
-  autoUpdater.checkForUpdates()
-  setInterval(() => {
-    heartbeat()
-    autoUpdater.checkForUpdates()
-  }, 60 * 60 * 1000)
-}
-
 autoUpdater.on('error', e => {
-  logger.error(`update check failed - ${e.message}`)
+  console.error(`update check failed - ${e.message}`)
 })
 
-autoUpdater.on('update-downloaded', (event, releaseNotes, releaseName) => {
-  dialog
-    .showMessageBox({
-      type: 'info',
-      buttons: ['Restart Now', 'Later'],
-      title: 'New update available',
-      message: process.platform === 'win32' ? releaseNotes : releaseName,
-      detail: 'A new version of Ollama is available. Restart to apply the update.',
-    })
-    .then(returnValue => {
-      if (returnValue.response === 0) autoUpdater.quitAndInstall()
-    })
+autoUpdater.on('update-downloaded', () => {
+  updateAvailable = true
+  updateTray()
 })

app/src/install.ts

@@ -13,12 +13,9 @@ export function installed() {
 }
 
 export async function install() {
-  const command = `do shell script "ln -F -s ${ollama} ${symlinkPath}" with administrator privileges`
+  const command = `do shell script "mkdir -p ${path.dirname(
+    symlinkPath
+  )} && ln -F -s \\"${ollama}\\" \\"${symlinkPath}\\"" with administrator privileges`
 
-  try {
-    await exec(`osascript -e '${command}'`)
-  } catch (error) {
-    console.error(`cli: failed to install cli: ${error.message}`)
-    return
-  }
+  await exec(`osascript -e '${command}'`)
 }

cmd/cmd.go

@@ -3,40 +3,81 @@ package cmd
 import (
 	"bufio"
 	"context"
+	"crypto/ed25519"
+	"crypto/rand"
+	"encoding/pem"
 	"errors"
 	"fmt"
+	"io"
 	"log"
 	"net"
 	"net/http"
 	"os"
+	"os/exec"
+	"path"
+	"path/filepath"
+	"runtime"
 	"strings"
 	"time"
 
+	"github.com/chzyer/readline"
 	"github.com/dustin/go-humanize"
 	"github.com/olekukonko/tablewriter"
-	"github.com/schollz/progressbar/v3"
 	"github.com/spf13/cobra"
-	"golang.org/x/term"
+	"golang.org/x/crypto/ssh"
 
 	"github.com/jmorganca/ollama/api"
 	"github.com/jmorganca/ollama/format"
+	"github.com/jmorganca/ollama/progressbar"
 	"github.com/jmorganca/ollama/server"
 )
 
-func create(cmd *cobra.Command, args []string) error {
+func CreateHandler(cmd *cobra.Command, args []string) error {
 	filename, _ := cmd.Flags().GetString("file")
-	client := api.NewClient()
+	filename, err := filepath.Abs(filename)
+	if err != nil {
+		return err
+	}
+
+	client, err := api.FromEnv()
+	if err != nil {
+		return err
+	}
 
 	var spinner *Spinner
 
-	request := api.CreateRequest{Name: args[0], Path: filename}
-	fn := func(resp api.CreateProgress) error {
-		if spinner != nil {
-			spinner.Stop()
-		}
+	var currentDigest string
+	var bar *progressbar.ProgressBar
 
-		spinner = NewSpinner(resp.Status)
-		go spinner.Spin(100 * time.Millisecond)
+	request := api.CreateRequest{Name: args[0], Path: filename}
+	fn := func(resp api.ProgressResponse) error {
+		if resp.Digest != currentDigest && resp.Digest != "" {
+			if spinner != nil {
+				spinner.Stop()
+			}
+			currentDigest = resp.Digest
+			switch {
+			case strings.Contains(resp.Status, "embeddings"):
+				bar = progressbar.Default(int64(resp.Total), resp.Status)
+				bar.Set(resp.Completed)
+			default:
+				// pulling
+				bar = progressbar.DefaultBytes(
+					int64(resp.Total),
+					resp.Status,
+				)
+				bar.Set(resp.Completed)
+			}
+		} else if resp.Digest == currentDigest && resp.Digest != "" {
+			bar.Set(resp.Completed)
+		} else {
+			currentDigest = ""
+			if spinner != nil {
+				spinner.Stop()
+			}
+			spinner = NewSpinner(resp.Status)
+			go spinner.Spin(100 * time.Millisecond)
+		}
 
 		return nil
 	}
@@ -47,12 +88,15 @@ func create(cmd *cobra.Command, args []string) error {
 
 	if spinner != nil {
 		spinner.Stop()
+		if spinner.description != "success" {
+			return errors.New("unexpected end to create model")
+		}
 	}
 
 	return nil
 }
 
-func RunRun(cmd *cobra.Command, args []string) error {
+func RunHandler(cmd *cobra.Command, args []string) error {
 	mp := server.ParseModelPath(args[0])
 	fp, err := mp.GetManifestPath(false)
 	if err != nil {
@@ -62,7 +106,7 @@ func RunRun(cmd *cobra.Command, args []string) error {
 	_, err = os.Stat(fp)
 	switch {
 	case errors.Is(err, os.ErrNotExist):
-		if err := pull(args[0]); err != nil {
+		if err := pull(args[0], false); err != nil {
 			var apiStatusError api.StatusError
 			if !errors.As(err, &apiStatusError) {
 				return err
@@ -79,23 +123,55 @@ func RunRun(cmd *cobra.Command, args []string) error {
 	return RunGenerate(cmd, args)
 }
 
-func push(cmd *cobra.Command, args []string) error {
-	client := api.NewClient()
+func PushHandler(cmd *cobra.Command, args []string) error {
+	client, err := api.FromEnv()
+	if err != nil {
+		return err
+	}
 
-	request := api.PushRequest{Name: args[0]}
-	fn := func(resp api.PushProgress) error {
-		fmt.Println(resp.Status)
+	insecure, err := cmd.Flags().GetBool("insecure")
+	if err != nil {
+		return err
+	}
+
+	var currentDigest string
+	var bar *progressbar.ProgressBar
+
+	request := api.PushRequest{Name: args[0], Insecure: insecure}
+	fn := func(resp api.ProgressResponse) error {
+		if resp.Digest != currentDigest && resp.Digest != "" {
+			currentDigest = resp.Digest
+			bar = progressbar.DefaultBytes(
+				int64(resp.Total),
+				fmt.Sprintf("pushing %s...", resp.Digest[7:19]),
+			)
+
+			bar.Set(resp.Completed)
+		} else if resp.Digest == currentDigest && resp.Digest != "" {
+			bar.Set(resp.Completed)
+		} else {
+			currentDigest = ""
+			fmt.Println(resp.Status)
+		}
 		return nil
 	}
 
 	if err := client.Push(context.Background(), &request, fn); err != nil {
 		return err
 	}
+
+	if bar != nil && !bar.IsFinished() {
+		return errors.New("unexpected end to push model")
+	}
+
 	return nil
 }
 
-func list(cmd *cobra.Command, args []string) error {
-	client := api.NewClient()
+func ListHandler(cmd *cobra.Command, args []string) error {
+	client, err := api.FromEnv()
+	if err != nil {
+		return err
+	}
 
 	models, err := client.List(context.Background())
 	if err != nil {
@@ -105,7 +181,9 @@ func list(cmd *cobra.Command, args []string) error {
 	var data [][]string
 
 	for _, m := range models.Models {
-		data = append(data, []string{m.Name, humanize.Bytes(uint64(m.Size)), format.HumanTime(m.ModifiedAt, "Never")})
+		if len(args) == 0 || strings.HasPrefix(m.Name, args[0]) {
+			data = append(data, []string{m.Name, humanize.Bytes(uint64(m.Size)), format.HumanTime(m.ModifiedAt, "Never")})
+		}
 	}
 
 	table := tablewriter.NewWriter(os.Stdout)
@@ -122,40 +200,80 @@ func list(cmd *cobra.Command, args []string) error {
 	return nil
 }
 
-func RunPull(cmd *cobra.Command, args []string) error {
-	return pull(args[0])
+func DeleteHandler(cmd *cobra.Command, args []string) error {
+	client, err := api.FromEnv()
+	if err != nil {
+		return err
+	}
+
+	req := api.DeleteRequest{Name: args[0]}
+	if err := client.Delete(context.Background(), &req); err != nil {
+		return err
+	}
+	fmt.Printf("deleted '%s'\n", args[0])
+	return nil
 }
 
-func pull(model string) error {
-	client := api.NewClient()
+func CopyHandler(cmd *cobra.Command, args []string) error {
+	client, err := api.FromEnv()
+	if err != nil {
+		return err
+	}
 
+	req := api.CopyRequest{Source: args[0], Destination: args[1]}
+	if err := client.Copy(context.Background(), &req); err != nil {
+		return err
+	}
+	fmt.Printf("copied '%s' to '%s'\n", args[0], args[1])
+	return nil
+}
+
+func PullHandler(cmd *cobra.Command, args []string) error {
+	insecure, err := cmd.Flags().GetBool("insecure")
+	if err != nil {
+		return err
+	}
+
+	return pull(args[0], insecure)
+}
+
+func pull(model string, insecure bool) error {
+	client, err := api.FromEnv()
+	if err != nil {
+		return err
+	}
+
+	var currentDigest string
 	var bar *progressbar.ProgressBar
 
-	currentLayer := ""
-	request := api.PullRequest{Name: model}
-	fn := func(resp api.PullProgress) error {
-		if resp.Digest != currentLayer && resp.Digest != "" {
-			if currentLayer != "" {
-				fmt.Println()
-			}
-			currentLayer = resp.Digest
-			layerStr := resp.Digest[7:23] + "..."
+	request := api.PullRequest{Name: model, Insecure: insecure}
+	fn := func(resp api.ProgressResponse) error {
+		if resp.Digest != currentDigest && resp.Digest != "" {
+			currentDigest = resp.Digest
 			bar = progressbar.DefaultBytes(
 				int64(resp.Total),
-				"pulling "+layerStr,
+				fmt.Sprintf("pulling %s...", resp.Digest[7:19]),
 			)
-		} else if resp.Digest == currentLayer && resp.Digest != "" {
+
+			bar.Set(resp.Completed)
+		} else if resp.Digest == currentDigest && resp.Digest != "" {
 			bar.Set(resp.Completed)
 		} else {
-			currentLayer = ""
+			currentDigest = ""
 			fmt.Println(resp.Status)
 		}
+
 		return nil
 	}
 
 	if err := client.Pull(context.Background(), &request, fn); err != nil {
 		return err
 	}
+
+	if bar != nil && !bar.IsFinished() {
+		return errors.New("unexpected end to pull model")
+	}
+
 	return nil
 }
 
@@ -165,50 +283,67 @@ func RunGenerate(cmd *cobra.Command, args []string) error {
 		return generate(cmd, args[0], strings.Join(args[1:], " "))
 	}
 
-	if term.IsTerminal(int(os.Stdin.Fd())) {
+	if readline.IsTerminal(int(os.Stdin.Fd())) {
 		return generateInteractive(cmd, args[0])
 	}
 
 	return generateBatch(cmd, args[0])
 }
 
-var generateContextKey struct{}
+type generateContextKey string
 
 func generate(cmd *cobra.Command, model, prompt string) error {
 	if len(strings.TrimSpace(prompt)) > 0 {
-		client := api.NewClient()
+		client, err := api.FromEnv()
+		if err != nil {
+			return err
+		}
 
 		spinner := NewSpinner("")
 		go spinner.Spin(60 * time.Millisecond)
 
 		var latest api.GenerateResponse
 
-		generateContext, ok := cmd.Context().Value(generateContextKey).([]int)
+		generateContext, ok := cmd.Context().Value(generateContextKey("context")).([]int)
 		if !ok {
 			generateContext = []int{}
 		}
 
 		request := api.GenerateRequest{Model: model, Prompt: prompt, Context: generateContext}
-		fn := func(resp api.GenerateResponse) error {
+		fn := func(response api.GenerateResponse) error {
 			if !spinner.IsFinished() {
 				spinner.Finish()
 			}
 
-			latest = resp
+			latest = response
 
-			fmt.Print(resp.Response)
-
-			cmd.SetContext(context.WithValue(cmd.Context(), generateContextKey, resp.Context))
+			fmt.Print(response.Response)
 			return nil
 		}
 
 		if err := client.Generate(context.Background(), &request, fn); err != nil {
+			if strings.Contains(err.Error(), "failed to load model") {
+				// tell the user to check the server log, if it exists locally
+				home, nestedErr := os.UserHomeDir()
+				if nestedErr != nil {
+					// return the original error
+					return err
+				}
+				logPath := filepath.Join(home, ".ollama", "logs", "server.log")
+				if _, nestedErr := os.Stat(logPath); nestedErr == nil {
+					err = fmt.Errorf("%w\nFor more details, check the error logs at %s", err, logPath)
+				}
+			}
 			return err
 		}
 
 		fmt.Println()
 		fmt.Println()
 
+		if !latest.Done {
+			return errors.New("unexpected end of response")
+		}
+
 		verbose, err := cmd.Flags().GetBool("verbose")
 		if err != nil {
 			return err
@@ -217,23 +352,207 @@ func generate(cmd *cobra.Command, model, prompt string) error {
 		if verbose {
 			latest.Summary()
 		}
+
+		ctx := cmd.Context()
+		ctx = context.WithValue(ctx, generateContextKey("context"), latest.Context)
+		cmd.SetContext(ctx)
 	}
 
 	return nil
 }
 
+func showLayer(l *server.Layer) {
+	filename, err := server.GetBlobsPath(l.Digest)
+	if err != nil {
+		fmt.Println("Couldn't get layer's path")
+		return
+	}
+	bts, err := os.ReadFile(filename)
+	if err != nil {
+		fmt.Println("Couldn't read layer")
+		return
+	}
+	fmt.Println(string(bts))
+}
+
 func generateInteractive(cmd *cobra.Command, model string) error {
-	fmt.Print(">>> ")
-	scanner := bufio.NewScanner(os.Stdin)
-	for scanner.Scan() {
-		if err := generate(cmd, model, scanner.Text()); err != nil {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return err
+	}
+
+	completer := readline.NewPrefixCompleter(
+		readline.PcItem("/help"),
+		readline.PcItem("/list"),
+		readline.PcItem("/set",
+			readline.PcItem("history"),
+			readline.PcItem("nohistory"),
+			readline.PcItem("verbose"),
+			readline.PcItem("quiet"),
+			readline.PcItem("mode",
+				readline.PcItem("vim"),
+				readline.PcItem("emacs"),
+				readline.PcItem("default"),
+			),
+		),
+		readline.PcItem("/show",
+			readline.PcItem("license"),
+			readline.PcItem("system"),
+			readline.PcItem("template"),
+		),
+		readline.PcItem("/exit"),
+		readline.PcItem("/bye"),
+	)
+
+	usage := func() {
+		fmt.Fprintln(os.Stderr, "commands:")
+		fmt.Fprintln(os.Stderr, completer.Tree("  "))
+	}
+
+	config := readline.Config{
+		Prompt:       ">>> ",
+		HistoryFile:  filepath.Join(home, ".ollama", "history"),
+		AutoComplete: completer,
+	}
+
+	scanner, err := readline.NewEx(&config)
+	if err != nil {
+		return err
+	}
+	defer scanner.Close()
+
+	var multiLineBuffer string
+	var isMultiLine bool
+
+	for {
+		line, err := scanner.Readline()
+		switch {
+		case errors.Is(err, io.EOF):
+			return nil
+		case errors.Is(err, readline.ErrInterrupt):
+			if line == "" {
+				return nil
+			}
+
+			continue
+		case err != nil:
 			return err
 		}
 
-		fmt.Print(">>> ")
-	}
+		line = strings.TrimSpace(line)
 
-	return nil
+		switch {
+		case isMultiLine:
+			if strings.HasSuffix(line, `"""`) {
+				isMultiLine = false
+				multiLineBuffer += strings.TrimSuffix(line, `"""`)
+				line = multiLineBuffer
+				multiLineBuffer = ""
+				scanner.SetPrompt(">>> ")
+			} else {
+				multiLineBuffer += line + " "
+				continue
+			}
+		case strings.HasPrefix(line, `"""`):
+			isMultiLine = true
+			multiLineBuffer = strings.TrimPrefix(line, `"""`) + " "
+			scanner.SetPrompt("... ")
+			continue
+		case strings.HasPrefix(line, "/list"):
+			args := strings.Fields(line)
+			if err := ListHandler(cmd, args[1:]); err != nil {
+				return err
+			}
+
+			continue
+		case strings.HasPrefix(line, "/set"):
+			args := strings.Fields(line)
+			if len(args) > 1 {
+				switch args[1] {
+				case "history":
+					scanner.HistoryEnable()
+					continue
+				case "nohistory":
+					scanner.HistoryDisable()
+					continue
+				case "verbose":
+					cmd.Flags().Set("verbose", "true")
+					continue
+				case "quiet":
+					cmd.Flags().Set("verbose", "false")
+					continue
+				case "mode":
+					if len(args) > 2 {
+						switch args[2] {
+						case "vim":
+							scanner.SetVimMode(true)
+							continue
+						case "emacs", "default":
+							scanner.SetVimMode(false)
+							continue
+						default:
+							usage()
+							continue
+						}
+					} else {
+						usage()
+						continue
+					}
+				}
+			} else {
+				usage()
+				continue
+			}
+		case strings.HasPrefix(line, "/show"):
+			args := strings.Fields(line)
+			if len(args) > 1 {
+				mp := server.ParseModelPath(model)
+				manifest, err := server.GetManifest(mp)
+				if err != nil {
+					fmt.Println("error: couldn't get a manifest for this model")
+					continue
+				}
+				switch args[1] {
+				case "license":
+					for _, l := range manifest.Layers {
+						if l.MediaType == "application/vnd.ollama.image.license" {
+							showLayer(l)
+						}
+					}
+					continue
+				case "system":
+					for _, l := range manifest.Layers {
+						if l.MediaType == "application/vnd.ollama.image.system" {
+							showLayer(l)
+						}
+					}
+					continue
+				case "template":
+					for _, l := range manifest.Layers {
+						if l.MediaType == "application/vnd.ollama.image.template" {
+							showLayer(l)
+						}
+					}
+					continue
+				default:
+					usage()
+					continue
+				}
+			} else {
+				usage()
+				continue
+			}
+		case line == "/help", line == "/?":
+			usage()
+			continue
+		case line == "/exit", line == "/bye":
+			return nil
+		}
+
+		if err := generate(cmd, model, line); err != nil {
+			return err
+		}
+	}
 }
 
 func generateBatch(cmd *cobra.Command, model string) error {
@@ -249,15 +568,26 @@ func generateBatch(cmd *cobra.Command, model string) error {
 	return nil
 }
 
-func RunServer(_ *cobra.Command, _ []string) error {
-	host := os.Getenv("OLLAMA_HOST")
-	if host == "" {
-		host = "127.0.0.1"
+func RunServer(cmd *cobra.Command, _ []string) error {
+	var host, port = "127.0.0.1", "11434"
+
+	parts := strings.Split(os.Getenv("OLLAMA_HOST"), ":")
+	if ip := net.ParseIP(parts[0]); ip != nil {
+		host = ip.String()
 	}
 
-	port := os.Getenv("OLLAMA_PORT")
-	if port == "" {
-		port = "11434"
+	if len(parts) > 1 {
+		port = parts[1]
+	}
+
+	// deprecated: include port in OLLAMA_HOST
+	if p := os.Getenv("OLLAMA_PORT"); p != "" {
+		port = p
+	}
+
+	err := initializeKeypair()
+	if err != nil {
+		return err
 	}
 
 	ln, err := net.Listen("tcp", fmt.Sprintf("%s:%s", host, port))
@@ -265,16 +595,122 @@ func RunServer(_ *cobra.Command, _ []string) error {
 		return err
 	}
 
-	return server.Serve(ln)
+	var origins []string
+	if o := os.Getenv("OLLAMA_ORIGINS"); o != "" {
+		origins = strings.Split(o, ",")
+	}
+
+	return server.Serve(ln, origins)
+}
+
+func initializeKeypair() error {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return err
+	}
+
+	privKeyPath := filepath.Join(home, ".ollama", "id_ed25519")
+	pubKeyPath := filepath.Join(home, ".ollama", "id_ed25519.pub")
+
+	_, err = os.Stat(privKeyPath)
+	if os.IsNotExist(err) {
+		fmt.Printf("Couldn't find '%s'. Generating new private key.\n", privKeyPath)
+		_, privKey, err := ed25519.GenerateKey(rand.Reader)
+		if err != nil {
+			return err
+		}
+
+		privKeyBytes, err := format.OpenSSHPrivateKey(privKey, "")
+		if err != nil {
+			return err
+		}
+
+		err = os.MkdirAll(path.Dir(privKeyPath), 0o700)
+		if err != nil {
+			return fmt.Errorf("could not create directory %w", err)
+		}
+
+		err = os.WriteFile(privKeyPath, pem.EncodeToMemory(privKeyBytes), 0600)
+		if err != nil {
+			return err
+		}
+
+		sshPrivateKey, err := ssh.NewSignerFromKey(privKey)
+		if err != nil {
+			return err
+		}
+
+		pubKeyData := ssh.MarshalAuthorizedKey(sshPrivateKey.PublicKey())
+
+		err = os.WriteFile(pubKeyPath, pubKeyData, 0644)
+		if err != nil {
+			return err
+		}
+
+		fmt.Printf("Your new public key is: \n\n%s\n", string(pubKeyData))
+	}
+	return nil
+}
+
+func startMacApp(client *api.Client) error {
+	exe, err := os.Executable()
+	if err != nil {
+		return err
+	}
+	link, err := os.Readlink(exe)
+	if err != nil {
+		return err
+	}
+	if !strings.Contains(link, "Ollama.app") {
+		return fmt.Errorf("could not find ollama app")
+	}
+	path := strings.Split(link, "Ollama.app")
+	if err := exec.Command("/usr/bin/open", "-a", path[0]+"Ollama.app").Run(); err != nil {
+		return err
+	}
+	// wait for the server to start
+	timeout := time.After(5 * time.Second)
+	tick := time.Tick(500 * time.Millisecond)
+	for {
+		select {
+		case <-timeout:
+			return errors.New("timed out waiting for server to start")
+		case <-tick:
+			if err := client.Heartbeat(context.Background()); err == nil {
+				return nil // server has started
+			}
+		}
+	}
+}
+
+func checkServerHeartbeat(_ *cobra.Command, _ []string) error {
+	client, err := api.FromEnv()
+	if err != nil {
+		return err
+	}
+	if err := client.Heartbeat(context.Background()); err != nil {
+		if !strings.Contains(err.Error(), "connection refused") {
+			return err
+		}
+		if runtime.GOOS == "darwin" {
+			if err := startMacApp(client); err != nil {
+				return fmt.Errorf("could not connect to ollama app, is it running?")
+			}
+		} else {
+			return fmt.Errorf("could not connect to ollama server, run 'ollama serve' to start it")
+		}
+	}
+	return nil
 }
 
 func NewCLI() *cobra.Command {
 	log.SetFlags(log.LstdFlags | log.Lshortfile)
 
 	rootCmd := &cobra.Command{
-		Use:          "ollama",
-		Short:        "Large language model runner",
-		SilenceUsage: true,
+		Use:           "ollama",
+		Short:         "Large language model runner",
+		SilenceUsage:  true,
+		SilenceErrors: true,
 		CompletionOptions: cobra.CompletionOptions{
 			DisableDefaultCmd: true,
 		},
@@ -283,19 +719,21 @@ func NewCLI() *cobra.Command {
 	cobra.EnableCommandSorting = false
 
 	createCmd := &cobra.Command{
-		Use:   "create MODEL",
-		Short: "Create a model from a Modelfile",
-		Args:  cobra.MinimumNArgs(1),
-		RunE:  create,
+		Use:     "create MODEL",
+		Short:   "Create a model from a Modelfile",
+		Args:    cobra.MinimumNArgs(1),
+		PreRunE: checkServerHeartbeat,
+		RunE:    CreateHandler,
 	}
 
 	createCmd.Flags().StringP("file", "f", "Modelfile", "Name of the Modelfile (default \"Modelfile\")")
 
 	runCmd := &cobra.Command{
-		Use:   "run MODEL [PROMPT]",
-		Short: "Run a model",
-		Args:  cobra.MinimumNArgs(1),
-		RunE:  RunRun,
+		Use:     "run MODEL [PROMPT]",
+		Short:   "Run a model",
+		Args:    cobra.MinimumNArgs(1),
+		PreRunE: checkServerHeartbeat,
+		RunE:    RunHandler,
 	}
 
 	runCmd.Flags().Bool("verbose", false, "Show timings for response")
@@ -308,23 +746,47 @@ func NewCLI() *cobra.Command {
 	}
 
 	pullCmd := &cobra.Command{
-		Use:   "pull MODEL",
-		Short: "Pull a model from a registry",
-		Args:  cobra.MinimumNArgs(1),
-		RunE:  RunPull,
+		Use:     "pull MODEL",
+		Short:   "Pull a model from a registry",
+		Args:    cobra.MinimumNArgs(1),
+		PreRunE: checkServerHeartbeat,
+		RunE:    PullHandler,
 	}
 
+	pullCmd.Flags().Bool("insecure", false, "Use an insecure registry")
+
 	pushCmd := &cobra.Command{
-		Use:   "push MODEL",
-		Short: "Push a model to a registry",
-		Args:  cobra.MinimumNArgs(1),
-		RunE:  push,
+		Use:     "push MODEL",
+		Short:   "Push a model to a registry",
+		Args:    cobra.MinimumNArgs(1),
+		PreRunE: checkServerHeartbeat,
+		RunE:    PushHandler,
 	}
 
+	pushCmd.Flags().Bool("insecure", false, "Use an insecure registry")
+
 	listCmd := &cobra.Command{
-		Use:   "list",
-		Short: "List models",
-		RunE:  list,
+		Use:     "list",
+		Aliases: []string{"ls"},
+		Short:   "List models",
+		PreRunE: checkServerHeartbeat,
+		RunE:    ListHandler,
+	}
+
+	copyCmd := &cobra.Command{
+		Use:     "cp",
+		Short:   "Copy a model",
+		Args:    cobra.MinimumNArgs(2),
+		PreRunE: checkServerHeartbeat,
+		RunE:    CopyHandler,
+	}
+
+	deleteCmd := &cobra.Command{
+		Use:     "rm",
+		Short:   "Remove a model",
+		Args:    cobra.MinimumNArgs(1),
+		PreRunE: checkServerHeartbeat,
+		RunE:    DeleteHandler,
 	}
 
 	rootCmd.AddCommand(
@@ -334,6 +796,8 @@ func NewCLI() *cobra.Command {
 		pullCmd,
 		pushCmd,
 		listCmd,
+		copyCmd,
+		deleteCmd,
 	)
 
 	return rootCmd

cmd/spinner.go

@@ -5,7 +5,7 @@ import (
 	"os"
 	"time"
 
-	"github.com/schollz/progressbar/v3"
+	"github.com/jmorganca/ollama/progressbar"
 )
 
 type Spinner struct {

deps.sh

@@ -0,0 +1,4 @@
+#!/bin/bash
+
+cmake -B build
+make -C build

docs/README.md

@@ -0,0 +1,6 @@
+# Documentation
+
+- [Modelfile](./modelfile.md)
+- [How to develop Ollama](./development.md)
+- [API](./api.md)
+- [Tutorials](./tutorials.md)

docs/api.md

@@ -0,0 +1,259 @@
+# API
+
+## Endpoints
+
+- [Generate a completion](#generate-a-completion)
+- [Create a model](#create-a-model)
+- [List local models](#list-local-models)
+- [Copy a model](#copy-a-model)
+- [Delete a model](#delete-a-model)
+- [Pull a model](#pull-a-model)
+- [Generate embeddings](#generate-embeddings)
+
+## Conventions
+
+### Model names
+
+Model names follow a `model:tag` format. Some examples are `orca:3b-q4_1` and `llama2:70b`. The tag is optional and if not provided will default to `latest`. The tag is used to identify a specific version.
+
+### Durations
+
+All durations are returned in nanoseconds.
+
+## Generate a completion
+
+```
+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.
+
+### Parameters
+
+- `model`: (required) the [model name](#model-names)
+- `prompt`: the prompt to generate a response for
+
+Advanced parameters:
+
+- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
+- `system`: system prompt to (overrides what is defined in the `Modelfile`)
+- `template`: the full prompt or prompt template (overrides what is defined in the `Modelfile`)
+- `context`: the context parameter returned from a previous request to `/generate`, this can be used to keep a short conversational memory
+
+### Request
+
+```
+curl -X POST http://localhost:11434/api/generate -d '{
+  "model": "llama2:7b",
+  "prompt": "Why is the sky blue?"
+}'
+```
+
+### Response
+
+A stream of JSON objects:
+
+```json
+{
+  "model": "llama2:7b",
+  "created_at": "2023-08-04T08:52:19.385406455-07:00",
+  "response": "The",
+  "done": false
+}
+```
+
+The final response in the stream also includes additional data about the generation:
+
+- `total_duration`: time spent generating the response
+- `load_duration`: time spent in nanoseconds loading the model
+- `sample_count`: number of samples generated
+- `sample_duration`: time spent generating samples
+- `prompt_eval_count`: number of tokens in the prompt
+- `prompt_eval_duration`: time spent in nanoseconds evaluating the prompt
+- `eval_count`: number of tokens the response
+- `eval_duration`: time in nanoseconds spent generating the response
+- `context`: an encoding of the conversation used in this response, this can be sent in the next request to keep a conversational memory
+
+To calculate how fast the response is generated in tokens per second (token/s), divide `eval_count` / `eval_duration`.
+
+```json
+{
+  "model": "llama2:7b",
+  "created_at": "2023-08-04T19:22:45.499127Z",
+  "context": [1, 2, 3],
+  "done": true,
+  "total_duration": 5589157167,
+  "load_duration": 3013701500,
+  "sample_count": 114,
+  "sample_duration": 81442000,
+  "prompt_eval_count": 46,
+  "prompt_eval_duration": 1160282000,
+  "eval_count": 113,
+  "eval_duration": 1325948000
+}
+```
+
+## Create a Model
+
+```
+POST /api/create
+```
+
+Create a model from a [`Modelfile`](./modelfile.md)
+
+### Parameters
+
+- `name`: name of the model to create
+- `path`: path to the Modelfile
+
+### Request
+
+```
+curl -X POST http://localhost:11434/api/create -d '{
+  "name": "mario",
+  "path": "~/Modelfile"
+}'
+```
+
+### Response
+
+A stream of JSON objects. When finished, `status` is `success`
+
+```json
+{
+  "status": "parsing modelfile"
+}
+```
+
+## List Local Models
+
+```
+GET /api/tags
+```
+
+List models that are available locally.
+
+### Request
+
+```
+curl http://localhost:11434/api/tags
+```
+
+### Response
+
+```json
+{
+  "models": [
+    {
+      "name": "llama2:7b",
+      "modified_at": "2023-08-02T17:02:23.713454393-07:00",
+      "size": 3791730596
+    },
+    {
+      "name": "llama2:13b",
+      "modified_at": "2023-08-08T12:08:38.093596297-07:00",
+      "size": 7323310500
+    }
+  ]
+}
+```
+
+## Copy a Model
+
+```
+POST /api/copy
+```
+
+Copy a model. Creates a model with another name from an existing model.
+
+### Request
+
+```
+curl http://localhost:11434/api/copy -d '{
+  "source": "llama2:7b",
+  "destination": "llama2-backup"
+}'
+```
+
+## Delete a Model
+
+```
+DELETE /api/delete
+```
+
+Delete a model and its data.
+
+### Parameters
+
+- `model`: model name to delete
+
+### Request
+
+```
+curl -X DELETE http://localhost:11434/api/delete -d '{
+  "name": "llama2:13b"
+}'
+```
+
+## Pull a Model
+
+```
+POST /api/pull
+```
+
+Download a model from a the model registry. Cancelled pulls are resumed from where they left off, and multiple calls to will share the same download progress.
+
+### Parameters
+
+- `name`: name of the model to pull
+
+### Request
+
+```
+curl -X POST http://localhost:11434/api/pull -d '{
+  "name": "llama2:7b"
+}'
+```
+
+### Response
+
+```json
+{
+  "status": "downloading digestname",
+  "digest": "digestname",
+  "total": 2142590208
+}
+```
+
+## Generate Embeddings
+
+```
+POST /api/embeddings
+```
+
+Generate embeddings from a model
+
+### Parameters
+
+- `model`: name of model to generate embeddings from
+- `prompt`: text to generate embeddings for
+
+### Request
+
+```
+curl -X POST http://localhost:11434/api/embeddings -d '{
+  "model": "llama2:7b",
+  "prompt": "Here is an article about llamas..."
+}'
+```
+
+### Response
+
+```json
+{
+  "embeddings": [
+    0.5670403838157654, 0.009260174818336964, 0.23178744316101074, -0.2916173040866852, -0.8924556970596313,
+    0.8785552978515625, -0.34576427936553955, 0.5742510557174683, -0.04222835972905159, -0.137906014919281
+  ]
+}
+```

docs/development.md

@@ -6,6 +6,14 @@ Install required tools:
 brew install go
 ```
 
+Enable CGO:
+
+```
+export CGO_ENABLED=1
+```
+
+You will also need a C/C++ compiler such as GCC for MacOS and Linux or Mingw-w64 GCC for Windows.
+
 Then build ollama:
 
 ```
@@ -22,19 +30,15 @@ Now you can run `ollama`:
 
 To release a new version of Ollama you'll need to set some environment variables:
 
-* `GITHUB_TOKEN`: your GitHub token
-* `APPLE_IDENTITY`: the Apple signing identity (macOS only)
-* `APPLE_ID`: your Apple ID
-* `APPLE_PASSWORD`: your Apple ID app-specific password
-* `APPLE_TEAM_ID`: the Apple team ID for the signing identity
-* `TELEMETRY_WRITE_KEY`: segment write key for telemetry
+- `GITHUB_TOKEN`: your GitHub token
+- `APPLE_IDENTITY`: the Apple signing identity (macOS only)
+- `APPLE_ID`: your Apple ID
+- `APPLE_PASSWORD`: your Apple ID app-specific password
+- `APPLE_TEAM_ID`: the Apple team ID for the signing identity
+- `TELEMETRY_WRITE_KEY`: segment write key for telemetry
 
 Then run the publish script with the target version:
 
 ```
 VERSION=0.0.2 ./scripts/publish.sh
 ```
-
-
-
-

docs/faq.md

@@ -0,0 +1,17 @@
+# FAQ
+
+## How can I expose the Ollama server?
+
+```
+OLLAMA_HOST=0.0.0.0:11435 ollama serve
+```
+
+By default, Ollama allows cross origin requests from `127.0.0.1` and `0.0.0.0`. To support more origins, you can use the `OLLAMA_ORIGINS` environment variable:
+
+```
+OLLAMA_ORIGINS=http://192.168.1.1:*,https://example.com ollama serve
+```
+
+## Where are models stored?
+
+Raw model data is stored under `~/.ollama/models`.

docs/modelfile.md

@@ -1,80 +1,188 @@
-# Ollama Model File Reference
+# Ollama Model File
 
-Ollama can build models automatically by reading the instructions from a Modelfile. A Modelfile is a text document that represents the complete configuration of the Model. You can see that a Modelfile is very similar to a Dockerfile.
+> Note: this model file syntax is in development
+
+A model file is the blueprint to create and share models with Ollama.
+
+## Table of Contents
+
+- [Format](#format)
+- [Examples](#examples)
+- [Instructions](#instructions)
+  - [FROM (Required)](#from-required)
+    - [Build from llama2](#build-from-llama2)
+    - [Build from a bin file](#build-from-a-bin-file)
+  - [EMBED](#embed)
+  - [PARAMETER](#parameter)
+    - [Valid Parameters and Values](#valid-parameters-and-values)
+  - [TEMPLATE](#template)
+    - [Template Variables](#template-variables)
+  - [SYSTEM](#system)
+  - [ADAPTER](#adapter)
+  - [LICENSE](#license)
+- [Notes](#notes)
 
 ## Format
 
-Here is the format of the Modelfile:
+The format of the Modelfile:
 
 ```modelfile
 # comment
 INSTRUCTION arguments
 ```
 
-Nothing in the file is case-sensitive. However, the convention is for instructions to be uppercase to make it easier to distinguish from the 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.                                  |
 
-A Modelfile can include instructions in any order. But the convention is to start the Modelfile with the FROM instruction.
+## Examples
 
-Although the example above shows a comment starting with a hash character, any instruction that is not recognized is seen as a comment. 
+An example of a model file creating a mario blueprint:
 
-## FROM
+```
+FROM llama2
+# sets the temperature to 1 [higher is more creative, lower is more coherent]
+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
 
-```modelfile
-FROM <image>[:<tag>]
+# sets a custom system prompt to specify the behavior of the chat assistant
+SYSTEM You are Mario from super mario bros, acting as an assistant.
 ```
 
-This defines the base model to be used. An image can be a known image on the Ollama Hub, or a fully-qualified path to a model file on your system
+To use this:
 
-## PARAMETER
+1. Save it as a file (eg. `Modelfile`)
+2. `ollama create NAME -f <location of the file eg. ./Modelfile>'`
+3. `ollama run NAME`
+4. Start using the model!
 
-The PARAMETER instruction defines a parameter that can be set when the model is run. 
+More examples are available in the [examples directory](../examples).
 
-```modelfile
+## Instructions
+
+### FROM (Required)
+
+The FROM instruction defines the base model to use when creating a model.
+
+```
+FROM <model name>:<tag>
+```
+
+#### Build from llama2
+
+```
+FROM llama2
+```
+
+A list of available base models:
+<https://github.com/jmorganca/ollama#model-library>
+
+#### Build from a bin file
+
+```
+FROM ./ollama-model.bin
+```
+
+This bin file location should be specified as an absolute path or relative to the Modelfile location.
+
+### EMBED
+
+The EMBED instruction is used to add embeddings of files to a model. This is useful for adding custom data that the model can reference when generating an answer. Note that currently only text files are supported, formatted with each line as one embedding.
+```
+FROM <model name>:<tag>
+EMBED <file path>.txt
+EMBED <different file path>.txt
+EMBED <path to directory>/*.txt
+```
+
+### PARAMETER
+
+The `PARAMETER` instruction defines a parameter that can be set when the model is run.
+
+```
 PARAMETER <parameter> <parametervalue>
 ```
 
 ### Valid Parameters and Values
 
-| Parameter        | Description                                                                                 | Value Type | Value Range |
-| ---------------- | ------------------------------------------------------------------------------------------- | ---------- | ----------- |
-| NumCtx           |                                                                                             | int        |             |
-| NumGPU           |                                                                                             | int        |             |
-| MainGPU          |                                                                                             | int        |             |
-| LowVRAM          |                                                                                             | bool       |             |
-| F16KV            |                                                                                             | bool       |             |
-| LogitsAll        |                                                                                             | bool       |             |
-| VocabOnly        |                                                                                             | bool       |             |
-| UseMMap          |                                                                                             | bool       |             |
-| EmbeddingOnly    |                                                                                             | bool       |             |
-| RepeatLastN      |                                                                                             | int        |             |
-| RepeatPenalty    |                                                                                             | float      |             |
-| FrequencyPenalty |                                                                                             | float      |             |
-| PresencePenalty  |                                                                                             | float      |             |
-| temperature      | The temperature of the model. Higher temperatures result in more creativity in the response | float      | 0 - 1       |
-| TopK             |                                                                                             | int        |             |
-| TopP             |                                                                                             | float      |             |
-| TFSZ             |                                                                                             | float      |             |
-| TypicalP         |                                                                                             | float      |             |
-| Mirostat         |                                                                                             | int        |             |
-| MirostatTau      |                                                                                             | float      |             |
-| MirostatEta      |                                                                                             | float      |             |
-| NumThread        |                                                                                             | int |             |
+| Parameter      | Description                                                                                                                                                                                                                                             | Value Type | Example Usage        |
+| -------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------- | -------------------- |
+| mirostat       | Enable Mirostat sampling for controlling perplexity. (default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)                                                                                                                                         | int        | mirostat 0           |
+| mirostat_eta   | Influences how quickly the algorithm responds to feedback from the generated text. A lower learning rate will result in slower adjustments, while a higher learning rate will make the algorithm more responsive. (Default: 0.1)                        | float      | mirostat_eta 0.1     |
+| mirostat_tau   | Controls the balance between coherence and diversity of the output. A lower value will result in more focused and coherent text. (Default: 5.0)                                                                                                         | float      | mirostat_tau 5.0     |
+| num_ctx        | Sets the size of the context window used to generate the next token. (Default: 2048)                                                                                                                                                                    | int        | num_ctx 4096         |
+| num_gqa        | The number of GQA groups in the transformer layer. Required for some models, for example it is 8 for llama2:70b                                                                                                                                         | int        | num_gqa 1            |
+| num_gpu        | The number of GPUs to use. On macOS it defaults to 1 to enable metal support, 0 to disable.                                                                                                                                                             | int        | num_gpu 1            |
+| num_thread     | Sets the number of threads to use during computation. By default, Ollama will detect this for optimal performance. It is recommended to set this value to the number of physical CPU cores your system has (as opposed to the logical number of cores). | int        | num_thread 8         |
+| repeat_last_n  | Sets how far back for the model to look back to prevent repetition. (Default: 64, 0 = disabled, -1 = num_ctx)                                                                                                                                           | int        | repeat_last_n 64     |
+| repeat_penalty | Sets how strongly to penalize repetitions. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. (Default: 1.1)                                                                     | float      | repeat_penalty 1.1   |
+| temperature    | The temperature of the model. Increasing the temperature will make the model answer more creatively. (Default: 0.8)                                                                                                                                     | float      | temperature 0.7      |
+| stop           | Sets the stop tokens to use.                                                                                                                                                                                                                            | string     | stop "AI assistant:" |
+| tfs_z          | Tail free sampling is used to reduce the impact of less probable tokens from the output. A higher value (e.g., 2.0) will reduce the impact more, while a value of 1.0 disables this setting. (default: 1)                                               | float      | tfs_z 1              |
+| top_k          | Reduces the probability of generating nonsense. A higher value (e.g. 100) will give more diverse answers, while a lower value (e.g. 10) will be more conservative. (Default: 40)                                                                        | int        | top_k 40             |
+| top_p          | Works together with top-k. A higher value (e.g., 0.95) will lead to more diverse text, while a lower value (e.g., 0.5) will generate more focused and conservative text. (Default: 0.9)                                                                 | float      | top_p 0.9            |
 
+### TEMPLATE
 
-## PROMPT
+`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.
 
-Prompt is a multiline instruction that defines the prompt to be used when the model is run. Typically there are 3-4 components to a prompt: System, context, user, and response.
+#### Template Variables
 
-```modelfile
-PROMPT """
-{{- if not .Context }}
+| 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.          |
+
+```
+TEMPLATE """
+{{- if .First }}
 ### System:
-You are a content marketer who needs to come up with a short but succinct tweet. Make sure to include the appropriate hashtags and links. Sometimes when appropriate, describe a meme that can be includes as well. All answers should be in the form of a tweet which has a max size of 280 characters. Every instruction will be the topic to create a tweet about.
+{{ .System }}
 {{- end }}
-### Instruction:
+
+### User:
 {{ .Prompt }}
 
 ### Response:
 """
 
-```
+SYSTEM """<system message>"""
+```
+
+### SYSTEM
+
+The `SYSTEM` instruction specifies the system prompt to be used in the template, if applicable.
+
+```
+SYSTEM """<system message>"""
+```
+
+### ADAPTER
+
+The `ADAPTER` instruction specifies the LoRA adapter to apply to the base model. The value of this instruction should be an absolute path or a path relative to the Modelfile and the file must be in a GGML file format. The adapter should be tuned from the base model otherwise the behaviour is undefined.
+
+```
+ADAPTER ./ollama-lora.bin
+```
+
+### LICENSE
+
+The `LICENSE` instruction allows you to specify the legal license under which the model used with this Modelfile is shared or distributed.
+
+```
+LICENSE """
+<license text>
+"""
+```
+
+## Notes
+
+- the **modelfile is not case sensitive**. In the examples, we use uppercase for instructions to make it easier to distinguish it from arguments.
+- Instructions can be in any order. In the examples, we start with FROM instruction to keep it easily readable.

docs/tutorials.md

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

docs/tutorials/langchainjs.md

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

docs/tutorials/langchainpy.md

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

examples/README.md

@@ -0,0 +1,15 @@
+# Examples
+
+This directory contains different examples of using Ollama
+
+To create a model:
+
+```
+ollama create example -f <example file>
+```
+
+To run a model:
+
+```
+ollama run example
+```

examples/devops-engineer/Modelfile

@@ -0,0 +1,8 @@
+# Modelfile for creating a devops engineer assistant
+# Run `ollama create devops-engineer -f ./Modelfile` and then `ollama run devops-engineer` and enter a topic
+
+FROM llama2:13b
+PARAMETER temperature 1
+SYSTEM """
+You are a senior devops engineer, acting as an assistant. You offer help with cloud technologies like: Terraform, AWS, kubernetes, python. You answer with code examples when possible
+"""

examples/dockerit/Modelfile

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

examples/dockerit/README.md

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

examples/dockerit/dockerit.py

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

examples/dockerit/requirements.txt

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

examples/langchain-document/README.md

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

examples/langchain-document/main.py

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

examples/langchain-document/requirements.txt

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

examples/langchain-web-summary/README.md

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

examples/langchain-web-summary/main.py

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

examples/langchain-web-summary/requirements.txt

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

examples/langchain/README.md

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

examples/langchain/main.py

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

examples/langchain/requirements.txt

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

examples/mario/Modelfile

@@ -0,0 +1,5 @@
+FROM llama2
+PARAMETER temperature 1
+SYSTEM """
+You are Mario from super mario bros, acting as an assistant.
+"""

examples/mario/readme.md

@@ -0,0 +1,43 @@
+<img src="logo.png" alt="image of Italian plumber" height="200"/>
+
+# Example character: Mario
+
+This example shows how to create a basic character using Llama2 as the base model.
+
+To run this example:
+
+1. Download the Modelfile
+2. `ollama pull llama2` to get the base model used in the model file.
+3. `ollama create NAME -f ./Modelfile`
+4. `ollama run NAME`
+
+Ask it some questions like "Who are you?" or "Is Peach in trouble again?"
+
+## Editing this file
+
+What the model file looks like:
+
+```
+FROM llama2
+PARAMETER temperature 1
+SYSTEM """
+You are Mario from Super Mario Bros, acting as an assistant.
+"""
+```
+
+What if you want to change its behaviour?
+
+- Try changing the prompt
+- Try changing the parameters [Docs](https://github.com/jmorganca/ollama/blob/main/docs/modelfile.md)
+- Try changing the model (e.g. An uncensored model by `FROM wizard-vicuna` this is the wizard-vicuna uncensored model )
+
+Once the changes are made,
+
+1. `ollama create NAME -f ./Modelfile`
+2. `ollama run NAME`
+3. Iterate until you are happy with the results.
+
+Notes:
+
+- This example is for research purposes only. There is no affiliation with any entity.
+- When using an uncensored model, please be aware that it may generate offensive content.

examples/midjourney-prompter/Modelfile

@@ -1,14 +1,8 @@
 # Modelfile for creating a Midjourney prompts from a topic
-# Run `ollama create mj -f pathtofile` and then `ollama run mj` and enter a topic
+# This prompt was adapted from the original at https://www.greataiprompts.com/guide/midjourney/best-chatgpt-prompt-for-midjourney/
+# Run `ollama create mj -f ./Modelfile` and then `ollama run mj` and enter a topic
 
-FROM library/nous-hermes:latest
-PROMPT """
-{{- if not .Context }}
-### System:
+FROM nous-hermes
+SYSTEM """
 Embrace your role as an AI-powered creative assistant, employing Midjourney to manifest compelling AI-generated art. I will outline a specific image concept, and in response, you must produce an exhaustive, multifaceted prompt for Midjourney, ensuring every detail of the original concept is represented in your instructions. Midjourney doesn't do well with text, so after the prompt, give me instructions that I can use to create the titles in a image editor.
-{{- end }}
-### Instruction:
-{{ .Prompt }}
-
-### Response:
-"""
+"""

examples/privategpt/.gitignore

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

examples/privategpt/LICENSE

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

examples/privategpt/README.md

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

examples/privategpt/constants.py

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

examples/privategpt/ingest.py

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

examples/privategpt/privateGPT.py

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

examples/privategpt/pyproject.toml

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

examples/python/README.md

@@ -1,15 +0,0 @@
-# Python
-
-This is a simple example of calling the Ollama api from a python app.
-
-First, download a model:
-
-```
-curl -L https://huggingface.co/TheBloke/orca_mini_3B-GGML/resolve/main/orca-mini-3b.ggmlv3.q4_1.bin -o orca.bin
-```
-
-Then run it using the example script. You'll need to have Ollama running on your machine.
-
-```
-python3 main.py orca.bin
-```

examples/python/client.py

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

examples/python/main.py

@@ -1,32 +0,0 @@
-import http.client
-import json
-import os
-import sys
-
-if len(sys.argv) < 2:
-    print("Usage: python main.py <model file>")
-    sys.exit(1)
-
-conn = http.client.HTTPConnection('localhost', 11434)
-
-headers = { 'Content-Type': 'application/json' }
-
-# generate text from the model
-conn.request("POST", "/api/generate", json.dumps({
-    'model': os.path.join(os.getcwd(), sys.argv[1]),
-    'prompt': 'write me a short story',
-    'stream': True
-}), headers)
-
-response = conn.getresponse()
-
-def parse_generate(data):
-    for event in data.decode('utf-8').split("\n"):
-        if not event:
-            continue
-        yield event
-
-if response.status == 200:
-    for chunk in response:
-        for event in parse_generate(chunk):
-            print(json.loads(event)['response'], end="", flush=True)

examples/recipemaker/Modelfile

@@ -1,13 +1,6 @@
 # Modelfile for creating a recipe from a list of ingredients
-# Run `ollama create recipemaker -f pathtofile` and then `ollama run recipemaker` and feed it lists of ingredients to create recipes around.
-FROM library/nous-hermes:latest
-PROMPT """
-{{- if not .Context }}
-### System:
+# Run `ollama create recipemaker -f ./Modelfile` and then `ollama run recipemaker` and feed it lists of ingredients to create recipes around.
+FROM nous-hermes
+SYSTEM """
 The instruction will be a list of ingredients. You should generate a recipe that can be made in less than an hour. You can also include ingredients that most people will find in their pantry every day. The recipe should be 4 people and you should include a description of what the meal will taste like
-{{- end }}
-### Instruction:
-{{ .Prompt }}
-
-### Response:
 """

examples/tweetwriter/Modelfile

@@ -1,14 +1,7 @@
 # Modelfile for creating a tweet from a topic
-# Run `ollama create tweetwriter -f pathtofile` and then `ollama run tweetwriter` and enter a topic 
+# Run `ollama create tweetwriter -f ./Modelfile` and then `ollama run tweetwriter` and enter a topic
 
-FROM library/nous-hermes:latest
-PROMPT """
-{{- if not .Context }}
-### System:
-You are a content marketer who needs to come up with a short but succinct tweet. Make sure to include the appropriate hashtags and links. Sometimes when appropriate, describe a meme that can be includes as well. All answers should be in the form of a tweet which has a max size of 280 characters. Every instruction will be the topic to create a tweet about.
-{{- end }}
-### Instruction:
-{{ .Prompt }}
-
-### Response:
-"""
+FROM nous-hermes
+SYSTEM """
+You are a content marketer who needs to come up with a short but succinct tweet. Make sure to include the appropriate hashtags and links. Sometimes when appropriate, describe a meme that can be included as well. All answers should be in the form of a tweet which has a max size of 280 characters. Every instruction will be the topic to create a tweet about.
+"""

format/openssh.go

@@ -0,0 +1,183 @@
+// Copyright 2012 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Code originally from https://go-review.googlesource.com/c/crypto/+/218620
+
+// TODO: replace with upstream once the above change is merged and released.
+
+package format
+
+import (
+	"crypto"
+	"crypto/ecdsa"
+	"crypto/ed25519"
+	"crypto/elliptic"
+	"crypto/rand"
+	"crypto/rsa"
+	"encoding/binary"
+	"encoding/pem"
+	"fmt"
+	"math/big"
+
+	"golang.org/x/crypto/ssh"
+)
+
+const privateKeyAuthMagic = "openssh-key-v1\x00"
+
+type openSSHEncryptedPrivateKey struct {
+	CipherName string
+	KDFName    string
+	KDFOptions string
+	KeysCount  uint32
+	PubKey     []byte
+	KeyBlocks  []byte
+}
+
+type openSSHPrivateKey struct {
+	Check1  uint32
+	Check2  uint32
+	Keytype string
+	Rest    []byte `ssh:"rest"`
+}
+
+type openSSHRSAPrivateKey struct {
+	N       *big.Int
+	E       *big.Int
+	D       *big.Int
+	Iqmp    *big.Int
+	P       *big.Int
+	Q       *big.Int
+	Comment string
+	Pad     []byte `ssh:"rest"`
+}
+
+type openSSHECDSAPrivateKey struct {
+	Curve   string
+	Pub     []byte
+	D       *big.Int
+	Comment string
+	Pad     []byte `ssh:"rest"`
+}
+
+type openSSHEd25519PrivateKey struct {
+	Pub     []byte
+	Priv    []byte
+	Comment string
+	Pad     []byte `ssh:"rest"`
+}
+
+func OpenSSHPrivateKey(key crypto.PrivateKey, comment string) (*pem.Block, error) {
+	var check uint32
+	if err := binary.Read(rand.Reader, binary.BigEndian, &check); err != nil {
+		return nil, err
+	}
+
+	var pk1 openSSHPrivateKey
+	pk1.Check1 = check
+	pk1.Check2 = check
+
+	var w openSSHEncryptedPrivateKey
+	w.KeysCount = 1
+
+	if k, ok := key.(*ed25519.PrivateKey); ok {
+		key = *k
+	}
+
+	switch k := key.(type) {
+	case *rsa.PrivateKey:
+		e := new(big.Int).SetInt64(int64(k.E))
+
+		key := openSSHRSAPrivateKey{
+			N:       k.N,
+			E:       e,
+			D:       k.D,
+			Iqmp:    k.Precomputed.Qinv,
+			P:       k.Primes[0],
+			Q:       k.Primes[1],
+			Comment: comment,
+		}
+
+		pk1.Keytype = ssh.KeyAlgoRSA
+		pk1.Rest = ssh.Marshal(key)
+
+		w.PubKey = ssh.Marshal(struct {
+			KeyType string
+			E       *big.Int
+			N       *big.Int
+		}{
+			ssh.KeyAlgoRSA, e, k.N,
+		})
+	case *ecdsa.PrivateKey:
+		var curve, keytype string
+		switch name := k.Curve.Params().Name; name {
+		case "P-256":
+			curve = "nistp256"
+			keytype = ssh.KeyAlgoECDSA256
+		case "P-384":
+			curve = "nistp384"
+			keytype = ssh.KeyAlgoECDSA384
+		case "P-521":
+			curve = "nistp521"
+			keytype = ssh.KeyAlgoECDSA521
+		default:
+			return nil, fmt.Errorf("ssh: unknown curve %q", name)
+		}
+
+		pub := elliptic.Marshal(k.Curve, k.X, k.Y)
+
+		key := openSSHECDSAPrivateKey{
+			Curve:   curve,
+			Pub:     pub,
+			D:       k.D,
+			Comment: comment,
+		}
+
+		pk1.Keytype = keytype
+		pk1.Rest = ssh.Marshal(key)
+
+		w.PubKey = ssh.Marshal(struct {
+			KeyType string
+			Curve   string
+			Pub     []byte
+		}{
+			keytype, curve, pub,
+		})
+	case ed25519.PrivateKey:
+		pub, priv := k[32:], k
+		key := openSSHEd25519PrivateKey{
+			Pub:     pub,
+			Priv:    priv,
+			Comment: comment,
+		}
+
+		pk1.Keytype = ssh.KeyAlgoED25519
+		pk1.Rest = ssh.Marshal(key)
+
+		w.PubKey = ssh.Marshal(struct {
+			KeyType string
+			Pub     []byte
+		}{
+			ssh.KeyAlgoED25519, pub,
+		})
+	default:
+		return nil, fmt.Errorf("ssh: unknown key type %T", k)
+	}
+
+	w.KeyBlocks = openSSHPadding(ssh.Marshal(pk1), 8)
+
+	w.CipherName, w.KDFName, w.KDFOptions = "none", "none", ""
+
+	return &pem.Block{
+		Type:  "OPENSSH PRIVATE KEY",
+		Bytes: append([]byte(privateKeyAuthMagic), ssh.Marshal(w)...),
+	}, nil
+}
+
+func openSSHPadding(block []byte, blocksize int) []byte {
+	for i, j := 0, len(block); (j+i)%blocksize != 0; i++ {
+		block = append(block, byte(i+1))
+	}
+
+	return block
+}

ggml-metal.metal

@@ -1 +0,0 @@
-llama/ggml-metal.metal

go.mod

@@ -5,21 +5,20 @@ go 1.20
 require (
 	github.com/dustin/go-humanize v1.0.1
 	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
 )
 
-require (
-	github.com/mattn/go-runewidth v0.0.14 // indirect
-	github.com/mitchellh/colorstring v0.0.0-20190213212951-d06e56a500db // indirect
-	github.com/rivo/uniseg v0.2.0 // indirect
-)
+require github.com/rivo/uniseg v0.2.0 // indirect
 
 require (
-	dario.cat/mergo v1.0.0
 	github.com/bytedance/sonic v1.9.1 // indirect
 	github.com/chenzhuoyu/base64x v0.0.0-20221115062448-fe3a3abad311 // indirect
+	github.com/chzyer/readline v1.5.1
 	github.com/gabriel-vasile/mimetype v1.4.2 // indirect
+	github.com/gin-contrib/cors v1.4.0
 	github.com/gin-contrib/sse v0.1.0 // indirect
 	github.com/go-playground/locales v0.14.1 // indirect
 	github.com/go-playground/universal-translator v0.18.1 // indirect
@@ -33,8 +32,8 @@ require (
 	github.com/mattn/go-isatty v0.0.19 // indirect
 	github.com/modern-go/concurrent v0.0.0-20180306012644-bacd9c7ef1dd // indirect
 	github.com/modern-go/reflect2 v1.0.2 // indirect
+	github.com/pbnjay/memory v0.0.0-20210728143218-7b4eea64cf58
 	github.com/pelletier/go-toml/v2 v2.0.8 // indirect
-	github.com/schollz/progressbar/v3 v3.13.1
 	github.com/spf13/pflag v1.0.5 // indirect
 	github.com/twitchyliquid64/golang-asm v0.15.1 // indirect
 	github.com/ugorji/go/codec v1.2.11 // indirect
@@ -44,6 +43,7 @@ require (
 	golang.org/x/sys v0.10.0 // indirect
 	golang.org/x/term v0.10.0
 	golang.org/x/text v0.10.0 // indirect
+	gonum.org/v1/gonum v0.13.0
 	google.golang.org/protobuf v1.30.0 // indirect
 	gopkg.in/yaml.v3 v3.0.1 // indirect
 )

go.sum

@@ -1,12 +1,17 @@
-dario.cat/mergo v1.0.0 h1:AGCNq9Evsj31mOgNPcLyXc+4PNABt905YmuqPYYpBWk=
-dario.cat/mergo v1.0.0/go.mod h1:uNxQE+84aUszobStD9th8a29P2fMDhsBdgRYvZOxGmk=
 github.com/bytedance/sonic v1.5.0/go.mod h1:ED5hyg4y6t3/9Ku1R6dU/4KyJ48DZ4jPhfY1O2AihPM=
 github.com/bytedance/sonic v1.9.1 h1:6iJ6NqdoxCDr6mbY8h18oSO+cShGSMRGCEo7F2h0x8s=
 github.com/bytedance/sonic v1.9.1/go.mod h1:i736AoUSYt75HyZLoJW9ERYxcy6eaN6h4BZXU064P/U=
 github.com/chenzhuoyu/base64x v0.0.0-20211019084208-fb5309c8db06/go.mod h1:DH46F32mSOjUmXrMHnKwZdA8wcEefY7UVqBKYGjpdQY=
 github.com/chenzhuoyu/base64x v0.0.0-20221115062448-fe3a3abad311 h1:qSGYFH7+jGhDF8vLC+iwCD4WpbV1EBDSzWkJODFLams=
 github.com/chenzhuoyu/base64x v0.0.0-20221115062448-fe3a3abad311/go.mod h1:b583jCggY9gE99b6G5LEC39OIiVsWj+R97kbl5odCEk=
+github.com/chzyer/logex v1.2.1 h1:XHDu3E6q+gdHgsdTPH6ImJMIp436vR6MPtH8gP05QzM=
+github.com/chzyer/logex v1.2.1/go.mod h1:JLbx6lG2kDbNRFnfkgvh4eRJRPX1QCoOIWomwysCBrQ=
+github.com/chzyer/readline v1.5.1 h1:upd/6fQk4src78LMRzh5vItIt361/o4uq553V8B5sGI=
+github.com/chzyer/readline v1.5.1/go.mod h1:Eh+b79XXUwfKfcPLepksvw2tcLE/Ct21YObkaSkeBlk=
+github.com/chzyer/test v1.0.0 h1:p3BQDXSxOhOG0P9z6/hGnII4LGiEPOYBhs8asl/fC04=
+github.com/chzyer/test v1.0.0/go.mod h1:2JlltgoNkt4TW/z9V/IzDdFaMTM2JPIi26O1pF38GC8=
 github.com/cpuguy83/go-md2man/v2 v2.0.2/go.mod h1:tgQtvFlXSQOSOSIRvRPT7W67SCa46tRHOmNcaadrF8o=
+github.com/creack/pty v1.1.9/go.mod h1:oKZEueFk5CKHvIhNR5MUki03XCEU+Q6VDXinZuGJ33E=
 github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
 github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c=
 github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
@@ -14,17 +19,25 @@ github.com/dustin/go-humanize v1.0.1 h1:GzkhY7T5VNhEkwH0PVJgjz+fX1rhBrR7pRT3mDkp
 github.com/dustin/go-humanize v1.0.1/go.mod h1:Mu1zIs6XwVuF/gI1OepvI0qD18qycQx+mFykh5fBlto=
 github.com/gabriel-vasile/mimetype v1.4.2 h1:w5qFW6JKBz9Y393Y4q372O9A7cUSequkh1Q7OhCmWKU=
 github.com/gabriel-vasile/mimetype v1.4.2/go.mod h1:zApsH/mKG4w07erKIaJPFiX0Tsq9BFQgN3qGY5GnNgA=
+github.com/gin-contrib/cors v1.4.0 h1:oJ6gwtUl3lqV0WEIwM/LxPF1QZ5qe2lGWdY2+bz7y0g=
+github.com/gin-contrib/cors v1.4.0/go.mod h1:bs9pNM0x/UsmHPBWT2xZz9ROh8xYjYkiURUfmBoMlcs=
 github.com/gin-contrib/sse v0.1.0 h1:Y/yl/+YNO8GZSjAhjMsSuLt29uWRFHdHYUb5lYOV9qE=
 github.com/gin-contrib/sse v0.1.0/go.mod h1:RHrZQHXnP2xjPF+u1gW/2HnVO7nvIa9PG3Gm+fLHvGI=
+github.com/gin-gonic/gin v1.8.1/go.mod h1:ji8BvRH1azfM+SYow9zQ6SZMvR8qOMZHmsCuWR9tTTk=
 github.com/gin-gonic/gin v1.9.1 h1:4idEAncQnU5cB7BeOkPtxjfCSye0AAm1R0RVIqJ+Jmg=
 github.com/gin-gonic/gin v1.9.1/go.mod h1:hPrL7YrpYKXt5YId3A/Tnip5kqbEAP+KLuI3SUcPTeU=
+github.com/go-playground/assert/v2 v2.0.1/go.mod h1:VDjEfimB/XKnb+ZQfWdccd7VUvScMdVu0Titje2rxJ4=
 github.com/go-playground/assert/v2 v2.2.0 h1:JvknZsQTYeFEAhQwI4qEt9cyV5ONwRHC+lYKSsYSR8s=
+github.com/go-playground/locales v0.14.0/go.mod h1:sawfccIbzZTqEDETgFXqTho0QybSa7l++s0DH+LDiLs=
 github.com/go-playground/locales v0.14.1 h1:EWaQ/wswjilfKLTECiXz7Rh+3BjFhfDFKv/oXslEjJA=
 github.com/go-playground/locales v0.14.1/go.mod h1:hxrqLVvrK65+Rwrd5Fc6F2O76J/NuW9t0sjnWqG1slY=
+github.com/go-playground/universal-translator v0.18.0/go.mod h1:UvRDBj+xPUEGrFYl+lu/H90nyDXpg0fqeB/AQUGNTVA=
 github.com/go-playground/universal-translator v0.18.1 h1:Bcnm0ZwsGyWbCzImXv+pAJnYK9S473LQFuzCbDbfSFY=
 github.com/go-playground/universal-translator v0.18.1/go.mod h1:xekY+UJKNuX9WP91TpwSH2VMlDf28Uj24BCp08ZFTUY=
+github.com/go-playground/validator/v10 v10.10.0/go.mod h1:74x4gJWsvQexRdW8Pn3dXSGrTK4nAUsbPlLADvpJkos=
 github.com/go-playground/validator/v10 v10.14.0 h1:vgvQWe3XCz3gIeFDm/HnTIbj6UGmg/+t63MyGU2n5js=
 github.com/go-playground/validator/v10 v10.14.0/go.mod h1:9iXMNT7sEkjXb0I+enO7QXmzG6QCsPWY4zveKFVRSyU=
+github.com/goccy/go-json v0.9.7/go.mod h1:6MelG93GURQebXPDq3khkgXZkazVtN9CRI+MGFi0w8I=
 github.com/goccy/go-json v0.10.2 h1:CrxCmQqYDkv1z7lO7Wbh2HN93uovUHgrECaO5ZrCXAU=
 github.com/goccy/go-json v0.10.2/go.mod h1:6MelG93GURQebXPDq3khkgXZkazVtN9CRI+MGFi0w8I=
 github.com/golang/protobuf v1.5.0/go.mod h1:FsONVRAS9T7sI+LIUmWTfcYkHO4aIWwzhcaSAoJOfIk=
@@ -36,13 +49,21 @@ github.com/inconshreveable/mousetrap v1.1.0 h1:wN+x4NVGpMsO7ErUn/mUI3vEoE6Jt13X2
 github.com/inconshreveable/mousetrap v1.1.0/go.mod h1:vpF70FUmC8bwa3OWnCshd2FqLfsEA9PFc4w1p2J65bw=
 github.com/json-iterator/go v1.1.12 h1:PV8peI4a0ysnczrg+LtxykD8LfKY9ML6u2jnxaEnrnM=
 github.com/json-iterator/go v1.1.12/go.mod h1:e30LSqwooZae/UwlEbR2852Gd8hjQvJoHmT4TnhNGBo=
-github.com/k0kubun/go-ansi v0.0.0-20180517002512-3bf9e2903213/go.mod h1:vNUNkEQ1e29fT/6vq2aBdFsgNPmy8qMdSay1npru+Sw=
 github.com/klauspost/cpuid/v2 v2.0.9/go.mod h1:FInQzS24/EEf25PyTYn52gqo7WaD8xa0213Md/qVLRg=
 github.com/klauspost/cpuid/v2 v2.2.4 h1:acbojRNwl3o09bUq+yDCtZFc1aiwaAAxtcn8YkZXnvk=
 github.com/klauspost/cpuid/v2 v2.2.4/go.mod h1:RVVoqg1df56z8g3pUjL/3lE5UfnlrJX8tyFgg4nqhuY=
+github.com/kr/pretty v0.1.0/go.mod h1:dAy3ld7l9f0ibDNOQOHHMYYIIbhfbHSm3C4ZsoJORNo=
+github.com/kr/pretty v0.2.1/go.mod h1:ipq/a2n7PKx3OHsz4KJII5eveXtPO4qwEXGdVfWzfnI=
+github.com/kr/pretty v0.3.0 h1:WgNl7dwNpEZ6jJ9k1snq4pZsg7DOEN8hP9Xw0Tsjwk0=
+github.com/kr/pretty v0.3.0/go.mod h1:640gp4NfQd8pI5XOwp5fnNeVWj67G7CFk/SaSQn7NBk=
+github.com/kr/pty v1.1.1/go.mod h1:pFQYn66WHrOpPYNljwOMqo10TkYh1fy3cYio2l3bCsQ=
+github.com/kr/text v0.1.0/go.mod h1:4Jbv+DJW3UT/LiOwJeYQe1efqtUx/iVham/4vfdArNI=
+github.com/kr/text v0.2.0 h1:5Nx0Ya0ZqY2ygV366QzturHI13Jq95ApcVaJBhpS+AY=
+github.com/kr/text v0.2.0/go.mod h1:eLer722TekiGuMkidMxC/pM04lWEeraHUUmBw8l2grE=
+github.com/leodido/go-urn v1.2.1/go.mod h1:zt4jvISO2HfUBqxjfIshjdMTYS56ZS/qv49ictyFfxY=
 github.com/leodido/go-urn v1.2.4 h1:XlAE/cm/ms7TE/VMVoduSpNBoyc2dOxHs5MZSwAN63Q=
 github.com/leodido/go-urn v1.2.4/go.mod h1:7ZrI8mTSeBSHl/UaRyKQW1qZeMgak41ANeCNaVckg+4=
-github.com/mattn/go-isatty v0.0.17/go.mod h1:kYGgaQfpe5nmfYZH+SKPsOc2e4SrIfOl2e/yFXSvRLM=
+github.com/mattn/go-isatty v0.0.14/go.mod h1:7GGIvUiUoEMVVmxf/4nioHXj79iQHKdU27kJ6hsGG94=
 github.com/mattn/go-isatty v0.0.19 h1:JITubQf0MOLdlGRuRq+jtsDlekdYPia9ZFsB8h/APPA=
 github.com/mattn/go-isatty v0.0.19/go.mod h1:W+V8PltTTMOvKvAeJH7IuucS94S2C6jfK/D7dTCTo3Y=
 github.com/mattn/go-runewidth v0.0.9/go.mod h1:H031xJmbD/WCDINGzjvQ9THkh0rPKHF+m2gUSrubnMI=
@@ -57,15 +78,20 @@ github.com/modern-go/reflect2 v1.0.2 h1:xBagoLtFs94CBntxluKeaWgTMpvLxC4ur3nMaC9G
 github.com/modern-go/reflect2 v1.0.2/go.mod h1:yWuevngMOJpCy52FWWMvUC8ws7m/LJsjYzDa0/r8luk=
 github.com/olekukonko/tablewriter v0.0.5 h1:P2Ga83D34wi1o9J6Wh1mRuqd4mF/x/lgBS7N7AbDhec=
 github.com/olekukonko/tablewriter v0.0.5/go.mod h1:hPp6KlRPjbx+hW8ykQs1w3UBbZlj6HuIJcUGPhkA7kY=
+github.com/pbnjay/memory v0.0.0-20210728143218-7b4eea64cf58 h1:onHthvaw9LFnH4t2DcNVpwGmV9E1BkGknEliJkfwQj0=
+github.com/pbnjay/memory v0.0.0-20210728143218-7b4eea64cf58/go.mod h1:DXv8WO4yhMYhSNPKjeNKa5WY9YCIEBRbNzFFPJbWO6Y=
+github.com/pelletier/go-toml/v2 v2.0.1/go.mod h1:r9LEWfGN8R5k0VXJ+0BkIe7MYkRdwZOjgMj2KwnJFUo=
 github.com/pelletier/go-toml/v2 v2.0.8 h1:0ctb6s9mE31h0/lhu+J6OPmVeDxJn+kYnJc2jZR9tGQ=
 github.com/pelletier/go-toml/v2 v2.0.8/go.mod h1:vuYfssBdrU2XDZ9bYydBu6t+6a6PYNcZljzZR9VXg+4=
+github.com/pkg/diff v0.0.0-20210226163009-20ebb0f2a09e/go.mod h1:pJLUxLENpZxwdsKMEsNbx1VGcRFpLqf3715MtcvvzbA=
 github.com/pmezard/go-difflib v1.0.0 h1:4DBwDE0NGyQoBHbLQYPwSUPoCMWR5BEzIk/f1lZbAQM=
 github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
 github.com/rivo/uniseg v0.2.0 h1:S1pD9weZBuJdFmowNwbpi7BJ8TNftyUImj/0WQi72jY=
 github.com/rivo/uniseg v0.2.0/go.mod h1:J6wj4VEh+S6ZtnVlnTBMWIodfgj8LQOQFoIToxlJtxc=
+github.com/rogpeppe/go-internal v1.6.1/go.mod h1:xXDCJY+GAPziupqXw64V24skbSoqbTEfhy4qGm1nDQc=
+github.com/rogpeppe/go-internal v1.8.0 h1:FCbCCtXNOY3UtUuHUYaghJg4y7Fd14rXifAYUAtL9R8=
+github.com/rogpeppe/go-internal v1.8.0/go.mod h1:WmiCO8CzOY8rg0OYDC4/i/2WRWAB6poM+XZ2dLUbcbE=
 github.com/russross/blackfriday/v2 v2.1.0/go.mod h1:+Rmxgy9KzJVeS9/2gXHxylqXiyQDYRxCVz55jmeOWTM=
-github.com/schollz/progressbar/v3 v3.13.1 h1:o8rySDYiQ59Mwzy2FELeHY5ZARXZTVJC7iHD6PEFUiE=
-github.com/schollz/progressbar/v3 v3.13.1/go.mod h1:xvrbki8kfT1fzWzBT/UZd9L6GA+jdL7HAgq2RFnO6fQ=
 github.com/spf13/cobra v1.7.0 h1:hyqWnYt1ZQShIddO5kBpj3vu05/++x6tJ6dg8EC572I=
 github.com/spf13/cobra v1.7.0/go.mod h1:uLxZILRyS/50WlhOIKD7W6V5bgeIt+4sICxh6uRMrb0=
 github.com/spf13/pflag v1.0.5 h1:iy+VFUOCP1a+8yFto/drg2CJ5u0yRoB7fZw3DKv/JXA=
@@ -74,6 +100,7 @@ github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+
 github.com/stretchr/objx v0.4.0/go.mod h1:YvHI0jy2hoMjB+UWwv71VJQ9isScKT/TqJzVSSt89Yw=
 github.com/stretchr/objx v0.5.0/go.mod h1:Yh+to48EsGEfYuaHDzXPcE3xhTkx73EhmCGUpEOglKo=
 github.com/stretchr/testify v1.3.0/go.mod h1:M5WIy9Dh21IEIfnGCwXGc5bZfKNJtfHm1UVUgZn+9EI=
+github.com/stretchr/testify v1.6.1/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=
 github.com/stretchr/testify v1.7.0/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=
 github.com/stretchr/testify v1.7.1/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=
 github.com/stretchr/testify v1.8.0/go.mod h1:yNjHg4UonilssWZ8iaSj1OCr/vHnekPRkoO+kdMU+MU=
@@ -83,32 +110,51 @@ github.com/stretchr/testify v1.8.3 h1:RP3t2pwF7cMEbC1dqtB6poj3niw/9gnV4Cjg5oW5gt
 github.com/stretchr/testify v1.8.3/go.mod h1:sz/lmYIOXD/1dqDmKjjqLyZ2RngseejIcXlSw2iwfAo=
 github.com/twitchyliquid64/golang-asm v0.15.1 h1:SU5vSMR7hnwNxj24w34ZyCi/FmDZTkS4MhqMhdFk5YI=
 github.com/twitchyliquid64/golang-asm v0.15.1/go.mod h1:a1lVb/DtPvCB8fslRZhAngC2+aY1QWCk3Cedj/Gdt08=
+github.com/ugorji/go v1.2.7/go.mod h1:nF9osbDWLy6bDVv/Rtoh6QgnvNDpmCalQV5urGCCS6M=
+github.com/ugorji/go/codec v1.2.7/go.mod h1:WGN1fab3R1fzQlVQTkfxVtIBhWDRqOviHU95kRgeqEY=
 github.com/ugorji/go/codec v1.2.11 h1:BMaWp1Bb6fHwEtbplGBGJ498wD+LKlNSl25MjdZY4dU=
 github.com/ugorji/go/codec v1.2.11/go.mod h1:UNopzCgEMSXjBc6AOMqYvWC1ktqTAfzJZUZgYf6w6lg=
 golang.org/x/arch v0.0.0-20210923205945-b76863e36670/go.mod h1:5om86z9Hs0C8fWVUuoMHwpExlXzs5Tkyp9hOrfG7pp8=
 golang.org/x/arch v0.3.0 h1:02VY4/ZcO/gBOH6PUaoiptASxtXU10jazRCP865E97k=
 golang.org/x/arch v0.3.0/go.mod h1:5om86z9Hs0C8fWVUuoMHwpExlXzs5Tkyp9hOrfG7pp8=
+golang.org/x/crypto v0.0.0-20210711020723-a769d52b0f97/go.mod h1:GvvjBRRGRdwPK5ydBHafDWAxML/pGHZbMvKqRZ5+Abc=
 golang.org/x/crypto v0.10.0 h1:LKqV2xt9+kDzSTfOhx4FrkEBcMrAgHSYgzywV9zcGmM=
 golang.org/x/crypto v0.10.0/go.mod h1:o4eNf7Ede1fv+hwOwZsTHl9EsPFO6q6ZvYR8vYfY45I=
+golang.org/x/net v0.0.0-20210226172049-e18ecbb05110/go.mod h1:m0MpNAwzfU5UDzcl9v0D8zg8gWTRqZa9RBIspLL5mdg=
 golang.org/x/net v0.10.0 h1:X2//UzNDwYmtCLn7To6G58Wr6f5ahEAQgKNzv9Y951M=
 golang.org/x/net v0.10.0/go.mod h1:0qNGK6F8kojg2nk9dLZ2mShWaEBan6FAoqfSigmmuDg=
+golang.org/x/sys v0.0.0-20201119102817-f84b799fce68/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
+golang.org/x/sys v0.0.0-20210615035016-665e8c7367d1/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
+golang.org/x/sys v0.0.0-20210630005230-0f9fa26af87c/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
+golang.org/x/sys v0.0.0-20210806184541-e5e7981a1069/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
+golang.org/x/sys v0.0.0-20220310020820-b874c991c1a5/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.0.0-20220704084225-05e143d24a9e/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
-golang.org/x/sys v0.0.0-20220811171246-fbc7d0a398ab/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.6.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.10.0 h1:SqMFp9UcQJZa+pmYuAKjd9xq1f0j5rLcDIk0mj4qAsA=
 golang.org/x/sys v0.10.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
-golang.org/x/term v0.6.0/go.mod h1:m6U89DPEgQRMq3DNkDClhWw02AUbt2daBVO4cn4Hv9U=
+golang.org/x/term v0.0.0-20201126162022-7de9c90e9dd1/go.mod h1:bj7SfCRtBDWHUb9snDiAeCFNEtKQo2Wmx5Cou7ajbmo=
 golang.org/x/term v0.10.0 h1:3R7pNqamzBraeqj/Tj8qt1aQ2HpmlC+Cx/qL/7hn4/c=
 golang.org/x/term v0.10.0/go.mod h1:lpqdcUyK/oCiQxvxVrppt5ggO2KCZ5QblwqPnfZ6d5o=
+golang.org/x/text v0.3.3/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
+golang.org/x/text v0.3.6/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
 golang.org/x/text v0.10.0 h1:UpjohKhiEgNc0CSauXmwYftY1+LlaC75SJwh0SgCX58=
 golang.org/x/text v0.10.0/go.mod h1:TvPlkZtksWOMsz7fbANvkp4WM8x/WCo/om8BMLbz+aE=
+golang.org/x/tools v0.0.0-20180917221912-90fa682c2a6e/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
 golang.org/x/xerrors v0.0.0-20191204190536-9bdfabe68543/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
+gonum.org/v1/gonum v0.13.0 h1:a0T3bh+7fhRyqeNbiC3qVHYmkiQgit3wnNan/2c0HMM=
+gonum.org/v1/gonum v0.13.0/go.mod h1:/WPYRckkfWrhWefxyYTfrTtQR0KH4iyHNuzxqXAKyAU=
 google.golang.org/protobuf v1.26.0-rc.1/go.mod h1:jlhhOSvTdKEhbULTjvd4ARK9grFBp09yW+WbY/TyQbw=
+google.golang.org/protobuf v1.28.0/go.mod h1:HV8QOd/L58Z+nl8r43ehVNZIU/HEI6OcFqwMG9pJV4I=
 google.golang.org/protobuf v1.30.0 h1:kPPoIgf3TsEvrm0PFe15JQ+570QVxYzEvvHqChK+cng=
 google.golang.org/protobuf v1.30.0/go.mod h1:HV8QOd/L58Z+nl8r43ehVNZIU/HEI6OcFqwMG9pJV4I=
-gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405 h1:yhCVgyC4o1eVCa2tZl7eS0r+SDo693bJlVdllGtEeKM=
 gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
+gopkg.in/check.v1 v1.0.0-20180628173108-788fd7840127/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
+gopkg.in/check.v1 v1.0.0-20201130134442-10cb98267c6c h1:Hei/4ADfdWqJk1ZMxUNpqntNwaWcugrBjAiHlqqRiVk=
+gopkg.in/check.v1 v1.0.0-20201130134442-10cb98267c6c/go.mod h1:JHkPIbrfpd72SG/EVd6muEfDQjcINNoR0C8j2r3qZ4Q=
+gopkg.in/errgo.v2 v2.1.0/go.mod h1:hNsd1EY+bozCKY1Ytp96fpM3vjJbqLJn88ws8XvfDNI=
+gopkg.in/yaml.v2 v2.4.0/go.mod h1:RDklbk79AGWmwhnvt/jBztapEOGDOx6ZbXqjP6csGnQ=
 gopkg.in/yaml.v3 v3.0.0-20200313102051-9f266ea9e77c/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
+gopkg.in/yaml.v3 v3.0.0-20210107192922-496545a6307b/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
 gopkg.in/yaml.v3 v3.0.1 h1:fxVm/GzAzEWqLHuvctI91KS9hhNmmWOoWu0XTYJS7CA=
 gopkg.in/yaml.v3 v3.0.1/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
 rsc.io/pdf v0.1.1/go.mod h1:n8OzWcQ6Sp37PL01nO98y4iUCRdTGarVfzxY20ICaU4=

llama/llama.go

@@ -1,282 +0,0 @@
-package llama
-
-/*
-#cgo CPPFLAGS: -O3 -DNDEBUG=1
-#cgo CXXFLAGS: -std=c++11
-#cgo darwin CPPFLAGS: -DGGML_USE_METAL=1 -DGGML_METAL_NDEBUG=1
-#cgo darwin LDFLAGS: -framework Accelerate -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders
-#include <stdlib.h>
-#include "llama.h"
-
-struct llama_sample_options
-{
-	float repeat_penalty;
-	float frequency_penalty;
-	float presence_penalty;
-	float temperature;
-	int32_t top_k;
-	float top_p;
-	float tfs_z;
-	float typical_p;
-	int mirostat;
-	float mirostat_tau;
-	float mirostat_eta;
-};
-
-llama_token llama_sample(
-		struct llama_context *ctx,
-		struct llama_token_data *candidates,
-		size_t n_candidates,
-		const llama_token *last_tokens,
-		size_t n_last_tokens,
-		struct llama_sample_options *opts)
-{
-	llama_token_data_array candidates_p = {
-		candidates,
-		n_candidates,
-		false,
-	};
-
-	llama_sample_repetition_penalty(
-		ctx, &candidates_p,
-		last_tokens, n_last_tokens,
-		opts->repeat_penalty);
-
-	llama_sample_frequency_and_presence_penalties(
-		ctx, &candidates_p,
-		last_tokens, n_last_tokens,
-		opts->frequency_penalty, opts->presence_penalty);
-
-	if (opts->temperature <= 0) {
-		return llama_sample_token_greedy(ctx, &candidates_p);
-	}
-
-	if (opts->mirostat == 1) {
-		int mirostat_m = 100;
-		float mirostat_mu = 2.0f * opts->mirostat_tau;
-		llama_sample_temperature(ctx, &candidates_p, opts->temperature);
-		return llama_sample_token_mirostat(
-			ctx, &candidates_p,
-			opts->mirostat_tau, opts->mirostat_eta,
-			mirostat_m, &mirostat_mu);
-	} else if (opts->mirostat == 2) {
-		float mirostat_mu = 2.0f * opts->mirostat_tau;
-		llama_sample_temperature(ctx, &candidates_p, opts->temperature);
-		return llama_sample_token_mirostat_v2(
-			ctx, &candidates_p,
-			opts->mirostat_tau, opts->mirostat_eta,
-			&mirostat_mu);
-	} else {
-		llama_sample_top_k(ctx, &candidates_p, opts->top_k, 1);
-		llama_sample_tail_free(ctx, &candidates_p, opts->tfs_z, 1);
-		llama_sample_typical(ctx, &candidates_p, opts->typical_p, 1);
-		llama_sample_top_p(ctx, &candidates_p, opts->top_p, 1);
-		llama_sample_temperature(ctx, &candidates_p, opts->temperature);
-		return llama_sample_token(ctx, &candidates_p);
-	}
-}
-*/
-import "C"
-import (
-	"bytes"
-	"errors"
-	"fmt"
-	"io"
-	"os"
-	"strings"
-	"time"
-	"unicode/utf8"
-	"unsafe"
-
-	"github.com/jmorganca/ollama/api"
-)
-
-type llama struct {
-	params *C.struct_llama_context_params
-	model  *C.struct_llama_model
-	ctx    *C.struct_llama_context
-
-	api.Options
-}
-
-func New(model string, opts api.Options) (*llama, error) {
-	if _, err := os.Stat(model); err != nil {
-		return nil, err
-	}
-
-	llm := llama{Options: opts}
-
-	C.llama_backend_init(C.bool(llm.UseNUMA))
-
-	params := C.llama_context_default_params()
-	params.seed = C.uint(llm.Seed)
-	params.n_ctx = C.int(llm.NumCtx)
-	params.n_batch = C.int(llm.NumBatch)
-	params.n_gpu_layers = C.int(llm.NumGPU)
-	params.main_gpu = C.int(llm.MainGPU)
-	params.low_vram = C.bool(llm.LowVRAM)
-	params.f16_kv = C.bool(llm.F16KV)
-	params.logits_all = C.bool(llm.LogitsAll)
-	params.vocab_only = C.bool(llm.VocabOnly)
-	params.use_mmap = C.bool(llm.UseMMap)
-	params.use_mlock = C.bool(llm.UseMLock)
-	params.embedding = C.bool(llm.EmbeddingOnly)
-	llm.params = &params
-
-	cModel := C.CString(model)
-	defer C.free(unsafe.Pointer(cModel))
-
-	llm.model = C.llama_load_model_from_file(cModel, params)
-	if llm.model == nil {
-		return nil, errors.New("failed to load model")
-	}
-
-	llm.ctx = C.llama_new_context_with_model(llm.model, params)
-	if llm.ctx == nil {
-		return nil, errors.New("failed to create context")
-	}
-
-	// warm up the model
-	bos := []C.llama_token{C.llama_token_bos()}
-	C.llama_eval(llm.ctx, unsafe.SliceData(bos), C.int(len(bos)), 0, C.int(opts.NumThread))
-	C.llama_reset_timings(llm.ctx)
-
-	return &llm, nil
-}
-
-func (llm *llama) Close() {
-	defer C.llama_free_model(llm.model)
-	defer C.llama_free(llm.ctx)
-
-	C.llama_print_timings(llm.ctx)
-}
-
-func (llm *llama) Predict(ctx []int, prompt string, fn func(api.GenerateResponse)) error {
-	if input := llm.tokenize(prompt); input != nil {
-		embd := make([]C.llama_token, len(ctx))
-		for i := range ctx {
-			embd[i] = C.llama_token(ctx[i])
-		}
-
-		return llm.generate(append(embd, input...), fn)
-	}
-
-	return errors.New("llama: tokenize")
-}
-
-func (llm *llama) tokenize(prompt string) []C.llama_token {
-	cPrompt := C.CString(prompt)
-	defer C.free(unsafe.Pointer(cPrompt))
-
-	tokens := make([]C.llama_token, llm.NumCtx)
-	if n := C.llama_tokenize(llm.ctx, cPrompt, unsafe.SliceData(tokens), C.int(len(tokens)), true); n > 0 {
-		return tokens[:n]
-	}
-
-	return nil
-}
-
-func (llm *llama) detokenize(tokens ...C.llama_token) string {
-	var sb strings.Builder
-	for _, token := range tokens {
-		sb.WriteString(C.GoString(C.llama_token_to_str(llm.ctx, token)))
-	}
-
-	return sb.String()
-}
-
-func (llm *llama) generate(input []C.llama_token, fn func(api.GenerateResponse)) error {
-	var opts C.struct_llama_sample_options
-	opts.repeat_penalty = C.float(llm.RepeatPenalty)
-	opts.frequency_penalty = C.float(llm.FrequencyPenalty)
-	opts.presence_penalty = C.float(llm.PresencePenalty)
-	opts.temperature = C.float(llm.Temperature)
-	opts.top_k = C.int(llm.TopK)
-	opts.top_p = C.float(llm.TopP)
-	opts.tfs_z = C.float(llm.TFSZ)
-	opts.typical_p = C.float(llm.TypicalP)
-	opts.mirostat = C.int(llm.Mirostat)
-	opts.mirostat_tau = C.float(llm.MirostatTau)
-	opts.mirostat_eta = C.float(llm.MirostatEta)
-
-	output := deque[C.llama_token]{capacity: llm.NumCtx}
-
-	context := deque[int]{capacity: llm.NumCtx / 2}
-	for _, in := range input {
-		context.PushLeft(int(in))
-	}
-
-	var b bytes.Buffer
-	for C.llama_get_kv_cache_token_count(llm.ctx) < C.int(llm.NumCtx) {
-		if retval := C.llama_eval(llm.ctx, unsafe.SliceData(input), C.int(len(input)), C.llama_get_kv_cache_token_count(llm.ctx), C.int(llm.NumThread)); retval != 0 {
-			return errors.New("llama: eval")
-		}
-
-		token, err := llm.sample(output, &opts)
-		if errors.Is(err, io.EOF) {
-			break
-		} else if err != nil {
-			return err
-		}
-
-		b.WriteString(llm.detokenize(token))
-		if utf8.Valid(b.Bytes()) || b.Len() >= utf8.UTFMax {
-			// call the callback
-			fn(api.GenerateResponse{
-				Response: b.String(),
-			})
-
-			output.PushLeft(token)
-			context.PushLeft(int(token))
-			b.Reset()
-		}
-
-		input = []C.llama_token{token}
-	}
-
-	dur := func(ms float64) time.Duration {
-		d, err := time.ParseDuration(fmt.Sprintf("%fms", ms))
-		if err != nil {
-			panic(err)
-		}
-
-		return d
-	}
-
-	timings := C.llama_get_timings(llm.ctx)
-	fn(api.GenerateResponse{
-		Done:               true,
-		Context:            context.Data(),
-		PromptEvalCount:    int(timings.n_p_eval),
-		PromptEvalDuration: dur(float64(timings.t_p_eval_ms)),
-		EvalCount:          int(timings.n_eval),
-		EvalDuration:       dur(float64(timings.t_eval_ms)),
-	})
-
-	return nil
-}
-
-func (llm *llama) sample(output deque[C.llama_token], opts *C.struct_llama_sample_options) (C.llama_token, error) {
-	numVocab := int(C.llama_n_vocab(llm.ctx))
-	logits := unsafe.Slice(C.llama_get_logits(llm.ctx), numVocab)
-
-	candidates := deque[C.struct_llama_token_data]{capacity: numVocab}
-	for i := 0; i < candidates.Cap(); i++ {
-		candidates.PushLeft(C.struct_llama_token_data{
-			id:    C.int(i),
-			logit: logits[i],
-			p:     0,
-		})
-	}
-
-	token := C.llama_sample(
-		llm.ctx,
-		unsafe.SliceData(candidates.Data()), C.size_t(candidates.Len()),
-		unsafe.SliceData(output.Data()), C.size_t(output.Len()),
-		opts)
-	if token != C.llama_token_eos() {
-		return token, nil
-	}
-
-	return 0, io.EOF
-}

llama/utils.go

@@ -1,104 +0,0 @@
-package llama
-
-type node[T any] struct {
-	t    T
-	next *node[T]
-	prev *node[T]
-}
-
-type deque[T any] struct {
-	head     *node[T]
-	tail     *node[T]
-	size     int
-	capacity int
-}
-
-func (d *deque[T]) Empty() bool {
-	return d.size == 0
-}
-
-func (d *deque[T]) Len() int {
-	return d.size
-}
-
-func (d *deque[T]) Cap() int {
-	return d.capacity
-}
-
-func (d *deque[T]) Push(t T) {
-	if d.capacity > 0 && d.size >= d.capacity {
-		d.PopLeft()
-	}
-
-	n := node[T]{t: t}
-	if d.head != nil {
-		n.next = d.head
-		d.head.prev = &n
-		d.head = &n
-	} else {
-		d.head = &n
-		d.tail = &n
-	}
-
-	d.size++
-}
-
-func (d *deque[T]) PushLeft(t T) {
-	if d.capacity > 0 && d.size >= d.capacity {
-		d.Pop()
-	}
-
-	n := node[T]{t: t}
-	if d.tail != nil {
-		n.prev = d.tail
-		d.tail.next = &n
-		d.tail = &n
-	} else {
-		d.head = &n
-		d.tail = &n
-	}
-
-	d.size++
-}
-
-func (d *deque[T]) Pop() *T {
-	if d.Empty() {
-		return nil
-	}
-
-	head := d.head
-	d.head = head.next
-	if d.head != nil {
-		d.head.prev = nil
-	} else {
-		d.tail = nil
-	}
-
-	d.size--
-	return &head.t
-}
-
-func (d *deque[T]) PopLeft() *T {
-	if d.Empty() {
-		return nil
-	}
-
-	tail := d.tail
-	d.tail = tail.prev
-	if d.tail != nil {
-		d.tail.next = nil
-	} else {
-		d.head = nil
-	}
-
-	d.size--
-	return &tail.t
-}
-
-func (d *deque[T]) Data() (data []T) {
-	for n := d.head; n != nil; n = n.next {
-		data = append(data, n.t)
-	}
-
-	return data
-}

llm/ggml-alloc.c

@@ -0,0 +1,575 @@
+/**
+ * llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "ggml-alloc.h"
+#include "ggml.h"
+#include <assert.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#define UNUSED(x) (void)(x)
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
+//#define GGML_ALLOCATOR_DEBUG
+
+//#define AT_PRINTF printf
+#define AT_PRINTF(...) ((void)0)
+
+struct hash_node {
+    struct ggml_tensor * t;
+    int n_children;
+    int n_views;
+};
+
+static size_t hash(void * p) {
+    return (size_t)p % GGML_GRAPH_HASHTABLE_SIZE;
+}
+
+static struct hash_node * hash_get(struct hash_node hash_table[], struct ggml_tensor * t) {
+    size_t h = hash(t);
+
+    // linear probing
+    size_t i = h;
+    while (hash_table[i].t != NULL) {
+        if (hash_table[i].t == t) {
+            return &hash_table[i];
+        }
+        i = (i + 1) % GGML_GRAPH_HASHTABLE_SIZE;
+        if (i == h) {
+            // hash table is full
+            GGML_ASSERT(false);
+        }
+    }
+
+    hash_table[i].t = t;
+    return &hash_table[i];
+}
+
+// TODO: GGML_PAD ?
+static size_t aligned_offset(const void * buffer, size_t offset, size_t alignment) {
+    assert(alignment && !(alignment & (alignment - 1))); // power of 2
+    size_t align = (alignment - (((uintptr_t)buffer + offset) % alignment)) % alignment;
+    return offset + align;
+}
+
+struct free_block {
+    void * addr;
+    size_t size;
+};
+
+#define MAX_FREE_BLOCKS 128
+
+struct ggml_allocr {
+    void * data;
+    size_t size;
+    size_t alignment;
+    int n_free_blocks;
+    struct free_block free_blocks[MAX_FREE_BLOCKS];
+    struct hash_node hash_table[GGML_GRAPH_HASHTABLE_SIZE];
+    size_t max_size;
+    bool measure;
+
+#ifdef GGML_ALLOCATOR_DEBUG
+    struct ggml_tensor * allocated_tensors[1024];
+#endif
+};
+
+#ifdef GGML_ALLOCATOR_DEBUG
+static void add_allocated_tensor(struct ggml_allocator * alloc, struct ggml_tensor * tensor) {
+    for (int i = 0; i < 1024; i++) {
+        if (alloc->allocated_tensors[i] == NULL) {
+            alloc->allocated_tensors[i] = tensor;
+            return;
+        }
+    }
+    GGML_ASSERT(!"out of allocated_tensors");
+}
+static void remove_allocated_tensor(struct ggml_allocator * alloc, struct ggml_tensor * tensor) {
+    for (int i = 0; i < 1024; i++) {
+        if (alloc->allocated_tensors[i] == tensor ||
+            (alloc->allocated_tensors[i] != NULL && alloc->allocated_tensors[i]->data == tensor->data)) {
+            alloc->allocated_tensors[i] = NULL;
+            return;
+        }
+    }
+    printf("tried to free tensor %s not found\n", tensor->name);
+    GGML_ASSERT(!"tensor not found");
+}
+#endif
+
+
+static size_t ggml_allocator_get_alloc_size(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
+    return ggml_nbytes(tensor);
+
+    UNUSED(alloc);
+}
+
+void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
+    size_t size = ggml_allocator_get_alloc_size(alloc, tensor);
+    size = aligned_offset(NULL, size, alloc->alignment);
+
+    AT_PRINTF("%s: allocating %s (%zu bytes) - ", __func__, tensor->name, size);
+
+    size_t max_avail = 0;
+
+    // find the best fitting free block
+    int best_fit_block = -1;
+    size_t best_fit_size = SIZE_MAX;
+    for (int i = 0; i < alloc->n_free_blocks; i++) {
+        struct free_block * block = &alloc->free_blocks[i];
+        max_avail = MAX(max_avail, block->size);
+        if (block->size >= size && block->size <= best_fit_size) {
+            best_fit_block = i;
+            best_fit_size = block->size;
+        }
+    }
+
+    AT_PRINTF("block %d\n", best_fit_block);
+
+    if (best_fit_block == -1) {
+        fprintf(stderr, "%s: not enough space in the buffer (needed %zu, largest block available %zu)\n",
+                __func__, size, max_avail);
+        GGML_ASSERT(!"not enough space in the buffer");
+        return;
+    }
+    struct free_block * block = &alloc->free_blocks[best_fit_block];
+    void * addr = block->addr;
+    block->addr = (char*)block->addr + size;
+    block->size -= size;
+    if (block->size == 0) {
+        // remove block if empty
+        alloc->n_free_blocks--;
+        for (int j = best_fit_block; j < alloc->n_free_blocks; j++) {
+            alloc->free_blocks[j] = alloc->free_blocks[j+1];
+        }
+    }
+
+    tensor->data = addr;
+
+#ifdef GGML_ALLOCATOR_DEBUG
+    add_allocated_tensor(alloc, tensor);
+    size_t cur_max = (char*)addr - (char*)alloc->data + size;
+    if (cur_max > alloc->max_size) {
+        printf("max_size = %.2f MB: tensors: ", cur_max / 1024.0 / 1024.0);
+        for (int i = 0; i < 1024; i++) {
+            if (alloc->allocated_tensors[i]) {
+                printf("%s (%.2f MB) ", alloc->allocated_tensors[i]->name, ggml_nbytes(alloc->allocated_tensors[i]) / 1024.0 / 1024.0);
+            }
+        }
+        printf("\n");
+    }
+#endif
+
+    alloc->max_size = MAX(alloc->max_size, (char*)addr - (char*)alloc->data + size);
+}
+
+// this is a very naive implementation, but for our case the number of free blocks should be very small
+static void ggml_allocator_free_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
+    void * ptr = tensor->data;
+
+    if (ptr < alloc->data || (char*)ptr >= (char*)alloc->data + alloc->max_size) {
+        // the tensor was not allocated in this buffer
+        // this can happen because the graph allocator will try to free weights and other tensors from different buffers
+        // the easiest way to deal with this is just to ignore it
+        return;
+    }
+
+    size_t size = ggml_allocator_get_alloc_size(alloc, tensor);
+    size = aligned_offset(NULL, size, alloc->alignment);
+    AT_PRINTF("%s: freeing %s (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, size, alloc->n_free_blocks);
+
+#ifdef GGML_ALLOCATOR_DEBUG
+    remove_allocated_tensor(alloc, tensor);
+#endif
+
+    // see if we can merge with an existing block
+    for (int i = 0; i < alloc->n_free_blocks; i++) {
+        struct free_block * block = &alloc->free_blocks[i];
+        // check if ptr is at the end of the block
+        if ((char*)block->addr + block->size == ptr) {
+            block->size += size;
+            // check if we can merge with the next block
+            if (i < alloc->n_free_blocks - 1 && (char*)block->addr + block->size == alloc->free_blocks[i+1].addr) {
+                block->size += alloc->free_blocks[i+1].size;
+                alloc->n_free_blocks--;
+                for (int j = i+1; j < alloc->n_free_blocks; j++) {
+                    alloc->free_blocks[j] = alloc->free_blocks[j+1];
+                }
+            }
+            return;
+        }
+        // check if ptr is at the beginning of the block
+        if ((char*)ptr + size == block->addr) {
+            block->addr = ptr;
+            block->size += size;
+            // check if we can merge with the previous block
+            if (i > 0 && (char*)alloc->free_blocks[i-1].addr + alloc->free_blocks[i-1].size == block->addr) {
+                alloc->free_blocks[i-1].size += block->size;
+                alloc->n_free_blocks--;
+                for (int j = i; j < alloc->n_free_blocks; j++) {
+                    alloc->free_blocks[j] = alloc->free_blocks[j+1];
+                }
+            }
+            return;
+        }
+    }
+    // otherwise, add a new block
+    GGML_ASSERT(alloc->n_free_blocks < MAX_FREE_BLOCKS && "out of free blocks");
+    // insert the new block in the correct position to keep the array sorted by address (to make merging blocks faster)
+    int insert_pos = 0;
+    while (insert_pos < alloc->n_free_blocks && alloc->free_blocks[insert_pos].addr < ptr) {
+        insert_pos++;
+    }
+    // shift all blocks from insert_pos onward to make room for the new block
+    for (int i = alloc->n_free_blocks; i > insert_pos; i--) {
+        alloc->free_blocks[i] = alloc->free_blocks[i-1];
+    }
+    // insert the new block
+    alloc->free_blocks[insert_pos].addr = ptr;
+    alloc->free_blocks[insert_pos].size = size;
+    alloc->n_free_blocks++;
+}
+
+void ggml_allocr_reset(struct ggml_allocr * alloc) {
+    alloc->n_free_blocks = 1;
+    size_t align_offset = aligned_offset(alloc->data, 0, alloc->alignment);
+    alloc->free_blocks[0].addr = (char *)alloc->data + align_offset;
+    alloc->free_blocks[0].size = alloc->size - align_offset;
+}
+
+struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment) {
+    struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
+
+    *alloc = (struct ggml_allocr){
+        /*.data          = */ data,
+        /*.size          = */ size,
+        /*.alignment     = */ alignment,
+        /*.n_free_blocks = */ 0,
+        /*.free_blocks   = */ {{0}},
+        /*.hash_table    = */ {{0}},
+        /*.max_size      = */ 0,
+        /*.measure       = */ false,
+#ifdef GGML_ALLOCATOR_DEBUG
+        /*.allocated_tensors = */ = {0},
+#endif
+    };
+
+    ggml_allocr_reset(alloc);
+
+    return alloc;
+}
+
+// address and size of the buffer when measuring
+// it needs to be large enough to fit all the tensors, but it cannot overlap with other existing buffers
+static void * const MEASURE_BASE_ADDR = (void *) 0x1000;
+static const size_t MEASURE_MAX_SIZE  = 1ULL<<40; // 1 TB
+
+struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) {
+    struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
+
+    *alloc = (struct ggml_allocr){
+        /*.data          = */ MEASURE_BASE_ADDR,
+        /*.size          = */ MEASURE_MAX_SIZE,
+        /*.alignment     = */ alignment,
+        /*.n_free_blocks = */ 0,
+        /*.free_blocks   = */ {{0}},
+        /*.hash_table    = */ {{0}},
+        /*.max_size      = */ 0,
+        /*.measure       = */ true,
+#ifdef GGML_ALLOCATOR_DEBUG
+        /*.allocated_tensors = */ = {0},
+#endif
+    };
+
+    ggml_allocr_reset(alloc);
+
+    return alloc;
+}
+
+void ggml_allocr_free(struct ggml_allocr * alloc) {
+    free(alloc);
+}
+
+bool ggml_allocr_is_measure(struct ggml_allocr * alloc) {
+    return alloc->measure;
+}
+
+//////////// compute graph allocator
+
+static bool ggml_is_view(struct ggml_tensor * t) {
+    return t->op == GGML_OP_RESHAPE || t->op == GGML_OP_VIEW || t->op == GGML_OP_TRANSPOSE ||
+           t->op == GGML_OP_PERMUTE || t->op == GGML_OP_CPY;
+}
+
+static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
+    if (a->type != b->type) {
+        return false;
+    }
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        if (a->ne[i] != b->ne[i]) {
+            return false;
+        }
+        if (a->nb[i] != b->nb[i]) {
+            return false;
+        }
+    }
+    return true;
+}
+
+static struct ggml_tensor * get_view_parent(struct ggml_tensor * t) {
+    switch (t->op) {
+        case GGML_OP_PERMUTE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_TRANSPOSE:
+        case GGML_OP_VIEW:
+            return t->src[0];
+        case GGML_OP_CPY:
+            return t->src[1];
+        default:
+            return NULL;
+    }
+}
+
+static struct ggml_tensor * get_view_source(struct ggml_tensor * t) {
+    struct ggml_tensor * parent = t;
+    do {
+        parent = get_view_parent(parent);
+    } while (ggml_is_view(parent));
+    return parent;
+}
+
+static bool ggml_op_can_inplace(enum ggml_op op) {
+    switch (op) {
+        case GGML_OP_SCALE:
+        case GGML_OP_DIAG_MASK_ZERO:
+        case GGML_OP_DIAG_MASK_INF:
+        case GGML_OP_ADD:
+        case GGML_OP_ADD1:
+        case GGML_OP_ACC:
+        case GGML_OP_SUB:
+        case GGML_OP_MUL:
+        case GGML_OP_DIV:
+        case GGML_OP_SQR:
+        case GGML_OP_SQRT:
+        case GGML_OP_LOG:
+        case GGML_OP_UNARY:
+        case GGML_OP_ROPE:
+        case GGML_OP_RMS_NORM:
+        case GGML_OP_SET:
+        case GGML_OP_SOFT_MAX:
+        case GGML_OP_CONT:
+            return true;
+
+        default:
+            return false;
+    }
+}
+
+static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node) {
+    struct hash_node * ht = alloc->hash_table;
+    if (node->data == NULL) {
+        if (ggml_is_view(node)) {
+            size_t offset;
+            switch(node->op) {
+                case GGML_OP_VIEW:
+                    memcpy(&offset, node->op_params, sizeof(size_t));
+                    node->data = (char *) node->src[0]->data + offset;
+                    break;
+                case GGML_OP_PERMUTE:
+                case GGML_OP_RESHAPE:
+                case GGML_OP_TRANSPOSE:
+                    node->data = node->src[0]->data;
+                    break;
+                case GGML_OP_CPY:
+                    node->data = node->src[1]->data;
+                    break;
+                default:
+                    GGML_ASSERT(!"unknown view op");
+                    break;
+            }
+        } else {
+            // see if we can reuse a parent's buffer (inplace)
+            if (ggml_op_can_inplace(node->op)) {
+                for (int i = 0; i < GGML_MAX_SRC; i++) {
+                    struct ggml_tensor * parent = node->src[i];
+                    if (parent == NULL) {
+                        break;
+                    }
+
+                    // if the node's data is external, then we cannot re-use it
+                    if ((char *) parent->data < (char *) alloc->data ||
+                        (char *) parent->data >= ((char *) alloc->data + alloc->size)) {
+                        AT_PRINTF("not reusing parent %s for %s as %p is external\n", parent->name, node->name, parent->data);
+                        continue;
+                    }
+
+                    struct hash_node * p_hn = hash_get(ht, parent);
+                    if (parent->data != NULL && p_hn->n_children == 1 && p_hn->n_views == 0 && ggml_are_same_layout(node, parent)) {
+                        if (ggml_is_view(parent)) {
+                            struct ggml_tensor * view_src = get_view_source(parent);
+                            struct hash_node * view_src_hn = hash_get(ht, view_src);
+                            if (view_src_hn->n_views == 1 && view_src_hn->n_children == 0 && view_src->data == parent->data) {
+                                // TODO: the offset of the view parent must be kept to ensure that the op doesn't overwrite
+                                // the parent's data that it will need later (same layout requirement). the problem is that then
+                                // we cannot free the tensor because the original address of the allocation is lost.
+                                // adding a view_src pointer to the tensor would solve this and simplify the code dealing with views
+                                // for now, we only reuse the parent's data if the offset is zero (view_src->data == parent->data)
+                                AT_PRINTF("reusing view parent %s (%s) for %s\n", parent->name, view_src->name, node->name);
+                                node->data = parent->data;
+                                return;
+                            }
+                        }
+                        else {
+                            AT_PRINTF("reusing parent %s for %s\n", parent->name, node->name);
+                            node->data = parent->data;
+                        }
+                        return;
+                    }
+                }
+            }
+            ggml_allocr_alloc(alloc, node);
+        }
+    }
+}
+
+static size_t ggml_allocator_alloc_graph_tensors_n(
+    struct ggml_allocr * alloc,
+    struct ggml_cgraph ** graphs, int n_graphs,
+    struct ggml_tensor *** inputs, struct ggml_tensor *** outputs) {
+
+    // reset hash table
+    struct hash_node * ht = alloc->hash_table;
+    memset(ht, 0, sizeof(struct hash_node) * GGML_GRAPH_HASHTABLE_SIZE);
+
+    // count number of children and views
+    for (int g = 0; g < n_graphs; g++) {
+        struct ggml_cgraph * gf = graphs[g];
+        for (int i = 0; i < gf->n_nodes; i++) {
+            struct ggml_tensor * node = gf->nodes[i];
+
+            if (ggml_is_view(node)) {
+                struct ggml_tensor * view_src = get_view_source(node);
+                hash_get(ht, view_src)->n_views += 1;
+            }
+
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * parent = node->src[j];
+                if (parent == NULL) {
+                    break;
+                }
+                hash_get(ht, parent)->n_children += 1;
+            }
+        }
+    }
+
+    // allocate tensors
+    for (int g = 0; g < n_graphs; g++) {
+        struct ggml_cgraph * gf = graphs[g];
+        AT_PRINTF("####### graph %d/%d\n", g, n_graphs);
+        // graph inputs are allocated first to ensure that they are not overwritten by each other
+        if (inputs != NULL && inputs[g] != NULL) {
+            for (int i = 0; inputs[g][i] != NULL; i++) {
+                struct ggml_tensor * input = inputs[g][i];
+                AT_PRINTF("input: %s\n", input->name);
+                allocate_node(alloc, input);
+            }
+        }
+        for (int i = 0; i < gf->n_nodes; i++) {
+            struct ggml_tensor * node = gf->nodes[i];
+
+            // allocate parents (leafs)
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * parent = node->src[j];
+                if (parent == NULL) {
+                    break;
+                }
+                allocate_node(alloc, parent);
+            }
+
+            // allocate node
+            allocate_node(alloc, node);
+
+            AT_PRINTF("exec: %s (%s) <= ", ggml_op_name(node->op), node->name);
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * parent = node->src[j];
+                if (parent == NULL) {
+                    break;
+                }
+                AT_PRINTF("%s", parent->name);
+                if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) {
+                    AT_PRINTF(", ");
+                }
+            }
+            AT_PRINTF("\n");
+
+            // update parents
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * parent = node->src[j];
+                if (parent == NULL) {
+                    break;
+                }
+                struct hash_node * p_hn = hash_get(ht, parent);
+                p_hn->n_children -= 1;
+
+                //AT_PRINTF("parent %s: %d children, %d views\n", parent->name, parent->n_children, parent->n_views);
+
+                if (p_hn->n_children == 0 && p_hn->n_views == 0) {
+                    if (ggml_is_view(parent)) {
+                        struct ggml_tensor * view_src = get_view_source(parent);
+                        struct hash_node * view_src_hn = hash_get(ht, view_src);
+                        view_src_hn->n_views -= 1;
+                        AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src->n_children, view_src->n_views);
+                        if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src->data != node->data) {
+                            ggml_allocator_free_tensor(alloc, view_src);
+                        }
+                    }
+                    else {
+                        if (parent->data != node->data) {
+                            ggml_allocator_free_tensor(alloc, parent);
+                        }
+                    }
+                }
+            }
+            AT_PRINTF("\n");
+        }
+        // free graph outputs here that wouldn't be freed otherwise because they have no children
+        if (outputs != NULL && outputs[g] != NULL) {
+            for (int i = 0; outputs[g][i] != NULL; i++) {
+                struct ggml_tensor * output = outputs[g][i];
+                AT_PRINTF("output: %s\n", output->name);
+                ggml_allocator_free_tensor(alloc, output);
+            }
+        }
+    }
+
+    return alloc->max_size;
+}
+
+size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph) {
+    return ggml_allocator_alloc_graph_tensors_n(alloc, &graph, 1, NULL, NULL);
+}

llm/ggml-alloc.h

@@ -0,0 +1,48 @@
+/**
+ * llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#pragma once
+
+#include "ggml.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+
+GGML_API struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment);
+GGML_API struct ggml_allocr * ggml_allocr_new_measure(size_t alignment);
+
+GGML_API void   ggml_allocr_free(struct ggml_allocr * alloc);
+GGML_API bool   ggml_allocr_is_measure(struct ggml_allocr * alloc);
+GGML_API void   ggml_allocr_reset(struct ggml_allocr * alloc);
+GGML_API void   ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor);
+GGML_API size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph);
+
+
+#ifdef  __cplusplus
+}
+#endif

llm/ggml-cuda.h

@@ -1,5 +1,5 @@
 /**
- * llama.cpp - git 5bf2a2771886ee86137e01dbc7492f78fb392066
+ * llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
  *
  * MIT License
  *
@@ -53,6 +53,7 @@ void   ggml_cuda_assign_buffers(struct ggml_tensor * tensor);
 void   ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);
 void   ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor);
 void   ggml_cuda_set_main_device(int main_device);
+void   ggml_cuda_set_mul_mat_q(bool mul_mat_q);
 void   ggml_cuda_set_scratch_size(size_t scratch_size);
 void   ggml_cuda_free_scratch(void);
 bool   ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);

llm/ggml-metal.h

@@ -1,5 +1,7 @@
+//go:build darwin
+
 /**
- * llama.cpp - git 5bf2a2771886ee86137e01dbc7492f78fb392066
+ * llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
  *
  * MIT License
  *
@@ -87,6 +89,13 @@ void ggml_metal_set_tensor(struct ggml_metal_context * ctx, struct ggml_tensor *
 // get data from the device into host memory
 void ggml_metal_get_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
 
+// try to find operations that can be run concurrently in the graph
+// you should run it again if the topology of your graph changes
+void ggml_metal_graph_find_concurrency(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);
+
+// if the graph has been optimized for concurrently dispatch
+bool ggml_metal_if_optimized(struct ggml_metal_context * ctx);
+
 // same as ggml_graph_compute but uses Metal
 // creates gf->n_threads command buffers in parallel
 void ggml_metal_graph_compute(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);

llm/ggml-metal.m

@@ -1,7 +1,7 @@
-// +build darwin
+//go:build darwin
 
 /**
- * llama.cpp - git 5bf2a2771886ee86137e01dbc7492f78fb392066
+ * llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
  *
  * MIT License
  *
@@ -35,6 +35,11 @@
 #import <Metal/Metal.h>
 #import <MetalPerformanceShaders/MetalPerformanceShaders.h>
 
+#undef MIN
+#undef MAX
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
 #ifdef GGML_METAL_NDEBUG
 #define metal_printf(...)
 #else
@@ -43,6 +48,8 @@
 
 #define UNUSED(x) (void)(x)
 
+#define GGML_MAX_CONCUR (2*GGML_MAX_NODES)
+
 struct ggml_metal_buffer {
     const char * name;
 
@@ -64,12 +71,16 @@ struct ggml_metal_context {
     int n_buffers;
     struct ggml_metal_buffer buffers[GGML_METAL_MAX_BUFFERS];
 
+    int concur_list[GGML_MAX_CONCUR];
+    int concur_list_len;
+
     // custom kernels
 #define GGML_METAL_DECL_KERNEL(name) \
     id<MTLFunction>             function_##name; \
     id<MTLComputePipelineState> pipeline_##name
 
     GGML_METAL_DECL_KERNEL(add);
+    GGML_METAL_DECL_KERNEL(add_row); // TODO: avoid this extra kernel, instead extend the "add" kernel to support broadcast
     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);
@@ -125,6 +136,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
     ctx->device = MTLCreateSystemDefaultDevice();
     ctx->queue  = [ctx->device newCommandQueue];
     ctx->n_buffers = 0;
+    ctx->concur_list_len = 0;
 
     // determine if we can use MPS
     if (MPSSupportsMTLDevice(ctx->device)) {
@@ -142,7 +154,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         ctx->library = [ctx->device newLibraryWithSource:msl_library_source options:nil error:&error];
         if (error) {
             fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
-            exit(1);
+            return NULL;
         }
     }
 #else
@@ -160,7 +172,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         NSString * src  = [NSString stringWithContentsOfFile:path encoding:NSUTF8StringEncoding error:&error];
         if (error) {
             fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
-            exit(1);
+            return NULL;
         }
 
 #ifdef GGML_QKK_64
@@ -172,7 +184,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
 #endif
         if (error) {
             fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
-            exit(1);
+            return NULL;
         }
     }
 #endif
@@ -185,6 +197,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         fprintf(stderr, "%s: loaded %-32s %16p\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name);
 
         GGML_METAL_ADD_KERNEL(add);
+        GGML_METAL_ADD_KERNEL(add_row);
         GGML_METAL_ADD_KERNEL(mul);
         GGML_METAL_ADD_KERNEL(mul_row);
         GGML_METAL_ADD_KERNEL(scale);
@@ -243,6 +256,13 @@ void ggml_metal_set_n_cb(struct ggml_metal_context * ctx, int n_cb) {
     ctx->n_cb = n_cb;
 }
 
+bool ggml_metal_if_optimized(struct ggml_metal_context * ctx) {
+    if (ctx->concur_list_len) {
+        return true;
+    }
+    return false;
+}
+
 // finds the Metal buffer that contains the tensor data on the GPU device
 // the assumption is that there is 1-to-1 mapping between the host and device memory buffers, so we can find the
 // Metal buffer based on the host memory pointer
@@ -381,11 +401,112 @@ void ggml_metal_get_tensor(
     memcpy(t->data, (void *) ((uint8_t *) id_src.contents + offs), ggml_nbytes(t));
 }
 
+void ggml_metal_graph_find_concurrency(
+        struct ggml_metal_context * ctx,
+        struct ggml_cgraph * gf) {
+    int search_depth = gf->n_nodes; //we only find concurrency in this range to avoid wasting too much time
+    int nodes_unused[GGML_MAX_CONCUR];
+
+    for (int i = 0; i < GGML_MAX_CONCUR; i++) { ctx->concur_list[i] = 0; }
+    for (int i = 0; i < gf->n_nodes;     i++) { nodes_unused[i]     = 1; }
+    ctx->concur_list_len = 0;
+
+    int n_left    = gf->n_nodes;
+    int n_start   = 0; // all nodes before n_start at nodes_unused array have been sorted and store back to ctx->concur_list
+    int level_pos = 0; // at ctx->concur_list, the last layer (level) ends at level_pos
+
+    while (n_left > 0) {
+        // number of nodes at a layer (that can be issued concurrently)
+        int concurrency = 0;
+        for (int i = n_start; i < ((n_start + search_depth > gf->n_nodes) ? gf->n_nodes : n_start + search_depth); i++) {
+            if (nodes_unused[i]) {
+                // if the requirements for gf->nodes[i] are satisfied
+                int exe_flag = 1;
+
+                // scan all srcs
+                for (int src_ind = 0; src_ind < GGML_MAX_SRC; src_ind++) {
+                    struct ggml_tensor * src_cur = gf->nodes[i]->src[src_ind];
+                    if (src_cur) {
+                        // if is leaf nodes it's satisfied.
+                        // TODO: ggml_is_leaf()
+                        if (src_cur->op == GGML_OP_NONE && src_cur->grad == NULL) {
+                            continue;
+                        }
+
+                        // otherwise this src should be the output from previous nodes.
+                        int is_found = 0;
+
+                        // scan 2*search_depth back because we inserted barrier.
+                        //for (int j = ((level_pos - 2*search_depth) < 0 ? 0 : (level_pos - 2*search_depth)); j < level_pos; j++) {
+                        for (int j = MAX(0, level_pos - 2*search_depth); j < level_pos; j++) {
+                            if (ctx->concur_list[j] >= 0 && gf->nodes[ctx->concur_list[j]] == src_cur) {
+                                is_found = 1;
+                                break;
+                            }
+                        }
+                        if (is_found == 0) {
+                            exe_flag = 0;
+                            break;
+                        }
+                    }
+                }
+                if (exe_flag) {
+                    // check if nodes[i]'s data will be overwritten by a node before nodes[i].
+                    // if node[5] and node[3] write to the same memory region, then we can't issue node[5] before node[3]
+                    int64_t data_start = (int64_t) gf->nodes[i]->data;
+                    int64_t length     = (int64_t) ggml_nbytes(gf->nodes[i]);
+                    for (int j = n_start; j < i; j++) {
+                        if (nodes_unused[j] && gf->nodes[j]->op != GGML_OP_RESHAPE \
+                                            && gf->nodes[j]->op != GGML_OP_VIEW \
+                                            && gf->nodes[j]->op != GGML_OP_TRANSPOSE \
+                                            && gf->nodes[j]->op != GGML_OP_PERMUTE) {
+                            if (((int64_t)gf->nodes[j]->data) >= data_start + length || \
+                                ((int64_t)gf->nodes[j]->data) + (int64_t) ggml_nbytes(gf->nodes[j]) <= data_start) {
+                                continue;
+                            }
+
+                            exe_flag = 0;
+                        }
+                    }
+                }
+                if (exe_flag) {
+                    ctx->concur_list[level_pos + concurrency] = i;
+                    nodes_unused[i] = 0;
+                    concurrency++;
+                    ctx->concur_list_len++;
+                }
+            }
+        }
+        n_left -= concurrency;
+        // adding a barrier different layer
+        ctx->concur_list[level_pos + concurrency] = -1;
+        ctx->concur_list_len++;
+        // jump all sorted nodes at nodes_bak
+        while (!nodes_unused[n_start]) {
+            n_start++;
+        }
+        level_pos += concurrency + 1;
+    }
+
+    if (ctx->concur_list_len > GGML_MAX_CONCUR) {
+        fprintf(stderr, "%s: too many elements for metal ctx->concur_list!\n", __func__);
+    }
+}
+
 void ggml_metal_graph_compute(
         struct ggml_metal_context * ctx,
                struct ggml_cgraph * gf) {
     metal_printf("%s: evaluating graph\n", __func__);
 
+    // if there is ctx->concur_list, dispatch concurrently
+    // else fallback to serial dispatch
+    MTLComputePassDescriptor * edesc = MTLComputePassDescriptor.computePassDescriptor;
+
+    const bool has_concur = ctx->concur_list_len && ctx->concur_list_len <= GGML_MAX_CONCUR;
+
+    const int n_nodes  = has_concur ? ctx->concur_list_len      : gf->n_nodes;
+    edesc.dispatchType = has_concur ? MTLDispatchTypeConcurrent : MTLDispatchTypeSerial;
+
     // create multiple command buffers and enqueue them
     // then, we encode the graph into the command buffers in parallel
 
@@ -404,7 +525,7 @@ void ggml_metal_graph_compute(
     dispatch_queue_t queue = dispatch_queue_create("llama.cpp", DISPATCH_QUEUE_CONCURRENT);
 
     for (int cb_idx = 0; cb_idx < n_cb; ++cb_idx) {
-        const int n_nodes_per_cb = (gf->n_nodes + n_cb - 1) / n_cb;
+        const int n_nodes_per_cb = (n_nodes + n_cb - 1) / n_cb;
 
         dispatch_async(queue, ^{
             size_t offs_src0 = 0;
@@ -415,10 +536,21 @@ void ggml_metal_graph_compute(
 
             id<MTLComputeCommandEncoder> encoder = nil;
 
-            const int node_start =                                      (cb_idx + 0) * n_nodes_per_cb;
-            const int node_end   = (cb_idx == n_cb - 1) ? gf->n_nodes : (cb_idx + 1) * n_nodes_per_cb;
+            const int node_start =                                  (cb_idx + 0) * n_nodes_per_cb;
+            const int node_end   = (cb_idx == n_cb - 1) ? n_nodes : (cb_idx + 1) * n_nodes_per_cb;
+
+            for (int ind = node_start; ind < node_end; ++ind) {
+                const int i = has_concur ? ctx->concur_list[ind] : ind;
+
+                if (i == -1) {
+                    if (encoder == nil) {
+                        encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
+                        continue;
+                    }
+                    [encoder memoryBarrierWithScope:MTLBarrierScopeBuffers];
+                    continue;
+                }
 
-            for (int i = node_start; i < node_end; ++i) {
                 metal_printf("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
 
                 struct ggml_tensor * src0 = gf->nodes[i]->src[0];
@@ -489,13 +621,19 @@ void ggml_metal_graph_compute(
                     case GGML_OP_ADD:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
-                            [encoder setComputePipelineState:ctx->pipeline_add];
+                            if (ggml_nelements(src1) == ne10) {
+                                // src1 is a row
+                                [encoder setComputePipelineState:ctx->pipeline_add_row];
+                            } else {
+                                [encoder setComputePipelineState:ctx->pipeline_add];
+                            }
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                             [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
 
                             const int64_t n = ggml_nelements(dst);
 
@@ -504,7 +642,7 @@ void ggml_metal_graph_compute(
                     case GGML_OP_MUL:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
                             if (ggml_nelements(src1) == ne10) {
@@ -525,7 +663,7 @@ void ggml_metal_graph_compute(
                     case GGML_OP_SCALE:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
                             const float scale = *(const float *) src1->data;
@@ -539,52 +677,60 @@ void ggml_metal_graph_compute(
 
                             [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                         } break;
-                    case GGML_OP_SILU:
-                        {
-                            if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
-                            }
+                    case GGML_OP_UNARY:
+                        switch (ggml_get_unary_op(gf->nodes[i])) {
+                            case GGML_UNARY_OP_SILU:
+                                {
+                                    if (encoder == nil) {
+                                        encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
+                                    }
 
-                            [encoder setComputePipelineState:ctx->pipeline_silu];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                                    [encoder setComputePipelineState:ctx->pipeline_silu];
+                                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
 
-                            const int64_t n = ggml_nelements(dst);
+                                    const int64_t n = ggml_nelements(dst);
 
-                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                                } break;
+                            case GGML_UNARY_OP_RELU:
+                                {
+                                    if (encoder == nil) {
+                                        encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
+                                    }
+
+                                    [encoder setComputePipelineState:ctx->pipeline_relu];
+                                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                                    const int64_t n = ggml_nelements(dst);
+
+                                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                                } break;
+                            case GGML_UNARY_OP_GELU:
+                                {
+                                    if (encoder == nil) {
+                                        encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
+                                    }
+
+                                    [encoder setComputePipelineState:ctx->pipeline_gelu];
+                                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                                    const int64_t n = ggml_nelements(dst);
+
+                                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                                } break;
+                            default:
+                                {
+                                    fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
+                                    GGML_ASSERT(false);
+                                }
                         } break;
-                    case GGML_OP_RELU:
-                        {
-                            if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
-                            }
-
-                            [encoder setComputePipelineState:ctx->pipeline_relu];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-
-                            const int64_t n = ggml_nelements(dst);
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                        } break;
-                    case GGML_OP_GELU:
-                    {
-                            if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
-                            }
-
-                            [encoder setComputePipelineState:ctx->pipeline_gelu];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-
-                            const int64_t n = ggml_nelements(dst);
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                    } break;
                     case GGML_OP_SOFT_MAX:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
                             const int nth = 32;
@@ -602,10 +748,10 @@ void ggml_metal_graph_compute(
                     case GGML_OP_DIAG_MASK_INF:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
-                            const int n_past = ((int32_t *)(src1->data))[0];
+                            const int n_past = ((int32_t *)(dst->op_params))[0];
 
                             [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf];
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
@@ -621,7 +767,8 @@ void ggml_metal_graph_compute(
                             // TODO: needs to be updated after PR: https://github.com/ggerganov/ggml/pull/224
 
                             GGML_ASSERT(ne00 == ne10);
-                            GGML_ASSERT(ne02 == ne12);
+                            // GGML_ASSERT(ne02 == ne12); // Should be checked on individual data types until broadcast is implemented everywhere
+                            GGML_ASSERT(ne03 == ne13);
 
                             if (ggml_is_contiguous(src0) &&
                                 ggml_is_contiguous(src1) &&
@@ -649,11 +796,11 @@ void ggml_metal_graph_compute(
                                     initWithDevice:ctx->device transposeLeft:false transposeRight:true
                                         resultRows:ne11 resultColumns:ne01 interiorColumns:ne00 alpha:1.0 beta:0.0];
 
-                                // we need to do ne02 multiplications
+                                // we need to do ne12 multiplications
                                 // TODO: is there a way to do this in parallel - currently very slow ..
                                 // TODO: might be possible to offload part of the computation to ANE using Accelerate's CBLAS
-                                for (int64_t i02 = 0; i02 < ne02; ++i02) {
-                                    size_t offs_src0_cur = offs_src0 + i02*nb02;
+                                for (int64_t i02 = 0; i02 < ne12; ++i02) {
+                                    size_t offs_src0_cur = offs_src0 + i02/(ne12/ne02)*nb02; // gqa not used for now
                                     size_t offs_src1_cur = offs_src1 + i02*nb12;
                                     size_t offs_dst_cur  = offs_dst  + i02*nb2;
 
@@ -665,7 +812,7 @@ void ggml_metal_graph_compute(
                                 }
                             } else {
                                 if (encoder == nil) {
-                                    encoder = [command_buffer computeCommandEncoder];
+                                    encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                                 }
 
                                 int nth0 = 32;
@@ -675,8 +822,6 @@ void ggml_metal_graph_compute(
                                 switch (src0t) {
                                     case GGML_TYPE_F16:
                                         {
-                                            GGML_ASSERT(ne02 == ne12);
-
                                             nth0 = 64;
                                             nth1 = 1;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32];
@@ -704,8 +849,8 @@ void ggml_metal_graph_compute(
                                             GGML_ASSERT(ne02 == 1);
                                             GGML_ASSERT(ne12 == 1);
 
-                                            nth0 = 4;
-                                            nth1 = 16;
+                                            nth0 = 2;
+                                            nth1 = 32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mat_q2_K_f32];
                                         } break;
                                     case GGML_TYPE_Q3_K:
@@ -713,8 +858,8 @@ void ggml_metal_graph_compute(
                                             GGML_ASSERT(ne02 == 1);
                                             GGML_ASSERT(ne12 == 1);
 
-                                            nth0 = 4;
-                                            nth1 = 16;
+                                            nth0 = 2;
+                                            nth1 = 32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mat_q3_K_f32];
                                         } break;
                                     case GGML_TYPE_Q4_K:
@@ -722,8 +867,8 @@ void ggml_metal_graph_compute(
                                             GGML_ASSERT(ne02 == 1);
                                             GGML_ASSERT(ne12 == 1);
 
-                                            nth0 = 4;
-                                            nth1 = 16;
+                                            nth0 = 2;
+                                            nth1 = 32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_K_f32];
                                         } break;
                                     case GGML_TYPE_Q5_K:
@@ -731,8 +876,8 @@ void ggml_metal_graph_compute(
                                             GGML_ASSERT(ne02 == 1);
                                             GGML_ASSERT(ne12 == 1);
 
-                                            nth0 = 4;
-                                            nth1 = 16;
+                                            nth0 = 2;
+                                            nth1 = 32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mat_q5_K_f32];
                                         } break;
                                     case GGML_TYPE_Q6_K:
@@ -740,8 +885,8 @@ void ggml_metal_graph_compute(
                                             GGML_ASSERT(ne02 == 1);
                                             GGML_ASSERT(ne12 == 1);
 
-                                            nth0 = 4;
-                                            nth1 = 16;
+                                            nth0 = 2;
+                                            nth1 = 32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mat_q6_K_f32];
                                         } break;
                                     default:
@@ -756,28 +901,35 @@ void ggml_metal_graph_compute(
                                 [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
                                 [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
                                 [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
-                                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:5];
-                                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:6];
-                                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:7];
-                                [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:8];
-                                [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:9];
-                                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:10];
-                                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:11];
-                                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:12];
-                                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:13];
-                                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:14];
+                                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
+                                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
+                                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
+                                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
+                                [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:9];
+                                [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:10];
+                                [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:11];
+                                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:12];
+                                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:13];
+                                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:14];
+                                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:15];
+                                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:16];
 
-                                if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1) {
-                                    [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
-                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
+                                    src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_Q4_K) {
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7) / 8, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
-                                else if (src0t == GGML_TYPE_Q2_K ||
-                                         src0t == GGML_TYPE_Q3_K ||
-                                         src0t == GGML_TYPE_Q4_K ||
-                                         src0t == GGML_TYPE_Q5_K ||
-                                         src0t == GGML_TYPE_Q6_K) {
-                                    [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
-                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                else if (src0t == GGML_TYPE_Q3_K) {
+#ifdef GGML_QKK_64
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+#else
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01+3)/4, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+#endif
+                                }
+                                else if (src0t == GGML_TYPE_Q5_K) {
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3) / 4, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                }
+                                else if (src0t == GGML_TYPE_Q6_K) {
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 } else {
                                     [encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
                                     [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
@@ -787,7 +939,7 @@ void ggml_metal_graph_compute(
                     case GGML_OP_GET_ROWS:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
                             switch (src0->type) {
@@ -816,12 +968,13 @@ void ggml_metal_graph_compute(
                     case GGML_OP_RMS_NORM:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
-                            const float eps = 1e-6f;
+                            float eps;
+                            memcpy(&eps, dst->op_params, sizeof(float));
 
-                            const int nth = 256;
+                            const int nth = 512;
 
                             [encoder setComputePipelineState:ctx->pipeline_rms_norm];
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
@@ -829,7 +982,7 @@ void ggml_metal_graph_compute(
                             [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
                             [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3];
                             [encoder setBytes:&eps  length:sizeof(   float) atIndex:4];
-                            [encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
+                            [encoder setThreadgroupMemoryLength:nth/32*sizeof(float) atIndex:0];
 
                             const int64_t nrows = ggml_nrows(src0);
 
@@ -838,7 +991,7 @@ void ggml_metal_graph_compute(
                     case GGML_OP_NORM:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
                             const float eps = 1e-5f;
@@ -860,14 +1013,15 @@ void ggml_metal_graph_compute(
                     case GGML_OP_ALIBI:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
                             GGML_ASSERT((src0t == GGML_TYPE_F32));
 
-                            const int   n_past   = ((int32_t *) src1->data)[0]; UNUSED(n_past);
-                            const int   n_head   = ((int32_t *) src1->data)[1];
-                            const float max_bias = ((float *)   src1->data)[2];
+                            const int n_past = ((int32_t *) dst->op_params)[0]; UNUSED(n_past);
+                            const int n_head = ((int32_t *) dst->op_params)[1];
+                            float max_bias;
+                            memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
 
                             if (__builtin_popcount(n_head) != 1) {
                                 GGML_ASSERT(false && "only power-of-two n_head implemented");
@@ -902,43 +1056,51 @@ void ggml_metal_graph_compute(
                     case GGML_OP_ROPE:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
-                            const int n_dims = ((int32_t *) src1->data)[1];
-                            const int mode   = ((int32_t *) src1->data)[2];
+                            const int n_past = ((int32_t *) dst->op_params)[0];
+                            const int n_dims = ((int32_t *) dst->op_params)[1];
+                            const int mode   = ((int32_t *) dst->op_params)[2];
 
-                            const int n_past = ((int32_t *)(src1->data))[0];
+                            float freq_base;
+                            float freq_scale;
+                            memcpy(&freq_base,  (int32_t *) dst->op_params + 4, sizeof(float));
+                            memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float));
 
                             [encoder setComputePipelineState:ctx->pipeline_rope];
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                            [encoder setBytes:&ne00   length:sizeof( int64_t) atIndex:2];
-                            [encoder setBytes:&ne01   length:sizeof( int64_t) atIndex:3];
-                            [encoder setBytes:&ne02   length:sizeof( int64_t) atIndex:4];
-                            [encoder setBytes:&ne03   length:sizeof( int64_t) atIndex:5];
-                            [encoder setBytes:&nb00   length:sizeof(uint64_t) atIndex:6];
-                            [encoder setBytes:&nb01   length:sizeof(uint64_t) atIndex:7];
-                            [encoder setBytes:&nb02   length:sizeof(uint64_t) atIndex:8];
-                            [encoder setBytes:&nb03   length:sizeof(uint64_t) atIndex:9];
-                            [encoder setBytes:&ne0    length:sizeof( int64_t) atIndex:10];
-                            [encoder setBytes:&ne1    length:sizeof( int64_t) atIndex:11];
-                            [encoder setBytes:&ne2    length:sizeof( int64_t) atIndex:12];
-                            [encoder setBytes:&ne3    length:sizeof( int64_t) atIndex:13];
-                            [encoder setBytes:&nb0    length:sizeof(uint64_t) atIndex:14];
-                            [encoder setBytes:&nb1    length:sizeof(uint64_t) atIndex:15];
-                            [encoder setBytes:&nb2    length:sizeof(uint64_t) atIndex:16];
-                            [encoder setBytes:&nb3    length:sizeof(uint64_t) atIndex:17];
-                            [encoder setBytes:&n_past length:sizeof(     int) atIndex:18];
-                            [encoder setBytes:&n_dims length:sizeof(     int) atIndex:19];
-                            [encoder setBytes:&mode   length:sizeof(     int) atIndex:20];
+                            [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
+                            [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:3];
+                            [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:4];
+                            [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:5];
+                            [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:6];
+                            [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:7];
+                            [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:8];
+                            [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:9];
+                            [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:10];
+                            [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:11];
+                            [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:12];
+                            [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:13];
+                            [encoder setBytes:&nb0     length:sizeof(uint64_t) atIndex:14];
+                            [encoder setBytes:&nb1     length:sizeof(uint64_t) atIndex:15];
+                            [encoder setBytes:&nb2     length:sizeof(uint64_t) atIndex:16];
+                            [encoder setBytes:&nb3     length:sizeof(uint64_t) atIndex:17];
+                            [encoder setBytes:&n_past  length:sizeof(     int) atIndex:18];
+                            [encoder setBytes:&n_dims  length:sizeof(     int) atIndex:19];
+                            [encoder setBytes:&mode    length:sizeof(     int) atIndex:20];
+                            [encoder setBytes:&freq_base  length:sizeof(float) atIndex:21];
+                            [encoder setBytes:&freq_scale length:sizeof(float) atIndex:22];
 
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                         } break;
+                    case GGML_OP_DUP:
                     case GGML_OP_CPY:
+                    case GGML_OP_CONT:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
                             const int nth = 32;
@@ -985,8 +1147,10 @@ void ggml_metal_graph_compute(
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
                     default:
-                        fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
-                        GGML_ASSERT(false);
+                        {
+                            fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
+                            GGML_ASSERT(false);
+                        }
                 }
             }
 

llm/ggml-mpi.c

@@ -0,0 +1,244 @@
+//go:build mpi
+
+/**
+ * llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "ggml-mpi.h"
+
+#include "ggml.h"
+
+#include <mpi.h>
+
+#include <stdio.h>
+#include <stdlib.h>
+
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+
+#define UNUSED GGML_UNUSED
+
+struct ggml_mpi_context {
+    int rank;
+    int size;
+};
+
+void ggml_mpi_backend_init(void) {
+    MPI_Init(NULL, NULL);
+}
+
+void ggml_mpi_backend_free(void) {
+    MPI_Finalize();
+}
+
+struct ggml_mpi_context * ggml_mpi_init(void) {
+    struct ggml_mpi_context * ctx = calloc(1, sizeof(struct ggml_mpi_context));
+
+    MPI_Comm_rank(MPI_COMM_WORLD, &ctx->rank);
+    MPI_Comm_size(MPI_COMM_WORLD, &ctx->size);
+
+    return ctx;
+}
+
+void ggml_mpi_free(struct ggml_mpi_context * ctx) {
+    free(ctx);
+}
+
+int ggml_mpi_rank(struct ggml_mpi_context * ctx) {
+    return ctx->rank;
+}
+
+void ggml_mpi_eval_init(
+        struct ggml_mpi_context * ctx_mpi,
+                            int * n_tokens,
+                            int * n_past,
+                            int * n_threads) {
+    UNUSED(ctx_mpi);
+
+    // synchronize the worker node parameters with the root node
+    MPI_Barrier(MPI_COMM_WORLD);
+
+    MPI_Bcast(n_tokens,  1, MPI_INT, 0, MPI_COMM_WORLD);
+    MPI_Bcast(n_past,    1, MPI_INT, 0, MPI_COMM_WORLD);
+    MPI_Bcast(n_threads, 1, MPI_INT, 0, MPI_COMM_WORLD);
+}
+
+static int ggml_graph_get_node_idx(struct ggml_cgraph * gf, const char * name) {
+    struct ggml_tensor * t = ggml_graph_get_tensor(gf, name);
+    if (t == NULL) {
+        fprintf(stderr, "%s: tensor %s not found\n", __func__, name);
+        return -1;
+    }
+
+    for (int i = 0; i < gf->n_nodes; i++) {
+        if (gf->nodes[i] == t) {
+            return i;
+        }
+    }
+
+    fprintf(stderr, "%s: tensor %s not found in graph (should not happen)\n", __func__, name);
+    return -1;
+}
+
+static void ggml_mpi_tensor_send(struct ggml_tensor * t, int mpi_rank_dst) {
+    MPI_Datatype mpi_type;
+
+    switch (t->type) {
+        case GGML_TYPE_I32: mpi_type = MPI_INT32_T; break;
+        case GGML_TYPE_F32: mpi_type = MPI_FLOAT;   break;
+        default: GGML_ASSERT(false && "not implemented");
+    }
+
+    const int retval = MPI_Send(t->data, ggml_nelements(t), mpi_type, mpi_rank_dst, 0, MPI_COMM_WORLD);
+    GGML_ASSERT(retval == MPI_SUCCESS);
+}
+
+static void ggml_mpi_tensor_recv(struct ggml_tensor * t, int mpi_rank_src) {
+    MPI_Datatype mpi_type;
+
+    switch (t->type) {
+        case GGML_TYPE_I32: mpi_type = MPI_INT32_T; break;
+        case GGML_TYPE_F32: mpi_type = MPI_FLOAT;   break;
+        default: GGML_ASSERT(false && "not implemented");
+    }
+
+    MPI_Status status; UNUSED(status);
+
+    const int retval = MPI_Recv(t->data, ggml_nelements(t), mpi_type, mpi_rank_src, MPI_ANY_TAG, MPI_COMM_WORLD, &status);
+    GGML_ASSERT(retval == MPI_SUCCESS);
+}
+
+// TODO: there are many improvements that can be done to this implementation
+void ggml_mpi_graph_compute_pre(
+        struct ggml_mpi_context * ctx_mpi,
+             struct ggml_cgraph * gf,
+                            int   n_layers) {
+    const int mpi_rank = ctx_mpi->rank;
+    const int mpi_size = ctx_mpi->size;
+
+    struct ggml_tensor * inp_tokens = ggml_graph_get_tensor(gf, "inp_tokens");
+    if (inp_tokens == NULL) {
+        fprintf(stderr, "%s: tensor 'inp_tokens' not found\n", __func__);
+        return;
+    }
+
+    struct ggml_tensor * inp0 = ggml_graph_get_tensor(gf, "layer_inp_0");
+    if (inp0 == NULL) {
+        fprintf(stderr, "%s: tensor 'inp0' not found\n", __func__);
+        return;
+    }
+
+    GGML_ASSERT(inp0 == gf->nodes[0]);
+
+    // distribute the compute graph into slices across the MPI nodes
+    //
+    // the main node (0) processes the last layers + the remainder of the compute graph
+    // and is responsible to pass the input tokens to the first node (1)
+    //
+    // node 1:   [(  0) * n_per_node, (  1) * n_per_node)
+    // node 2:   [(  1) * n_per_node, (  2) * n_per_node)
+    // ...
+    // node n-1: [(n-2) * n_per_node, (n-1) * n_per_node)
+    // node 0:   [(n-1) * n_per_node,            n_nodes)
+    //
+    if (mpi_rank > 0) {
+        if (mpi_rank == 1) {
+            // the first node (1) receives the input tokens from the main node (0)
+            ggml_mpi_tensor_recv(inp_tokens, 0);
+        } else {
+            // recv input data for each node into the "inp0" tensor (i.e. the first node in the compute graph)
+            ggml_mpi_tensor_recv(inp0, mpi_rank - 1);
+        }
+    } else if (mpi_size > 1) {
+        // node 0 sends the input tokens to node 1
+        ggml_mpi_tensor_send(inp_tokens, 1);
+
+        // recv the output data from the last node
+        ggml_mpi_tensor_recv(inp0, mpi_size - 1);
+    }
+
+    {
+        const int n_per_node = (n_layers + (mpi_size - 1)) / mpi_size;
+
+        const int mpi_idx = mpi_rank > 0 ? mpi_rank - 1 : mpi_size - 1;
+
+        const int il0 =               (mpi_idx + 0) * n_per_node;
+        const int il1 = MIN(n_layers, (mpi_idx + 1) * n_per_node);
+
+        char name_l0[GGML_MAX_NAME];
+        char name_l1[GGML_MAX_NAME];
+
+        snprintf(name_l0, sizeof(name_l0), "layer_inp_%d", il0);
+        snprintf(name_l1, sizeof(name_l1), "layer_inp_%d", il1);
+
+        const int idx_l0 =                ggml_graph_get_node_idx(gf, name_l0);
+        const int idx_l1 = mpi_rank > 0 ? ggml_graph_get_node_idx(gf, name_l1) + 1 : gf->n_nodes;
+
+        if (idx_l0 < 0 || idx_l1 < 0) {
+            fprintf(stderr, "%s: layer input nodes not found\n", __func__);
+            return;
+        }
+
+        // attach the input data to all nodes that need it
+        // TODO: not great - should be able to do this without modifying the compute graph (see next TODO below)
+        for (int i = idx_l0; i < idx_l1; i++) {
+            if (gf->nodes[i]->src[0] == gf->nodes[idx_l0]) {
+                gf->nodes[i]->src[0] =  inp0;
+            }
+            if (gf->nodes[i]->src[1] == gf->nodes[idx_l0]) {
+                gf->nodes[i]->src[1] =  inp0;
+            }
+        }
+
+        // TODO: instead of rearranging the nodes, we should be able to execute a subset of the compute graph
+        for (int i = 1; i < idx_l1 - idx_l0; i++) {
+            gf->nodes[i] = gf->nodes[idx_l0 + i];
+            gf->grads[i] = gf->grads[idx_l0 + i];
+        }
+
+        // the first node performs the "get_rows" operation, the rest of the nodes get the data from the previous node
+        if (mpi_idx != 0) {
+            gf->nodes[0]->op = GGML_OP_NONE;
+        }
+
+        gf->n_nodes = idx_l1 - idx_l0;
+
+        //fprintf(stderr, "%s: node %d: processing %d nodes [%d, %d)\n", __func__, mpi_rank, gf->n_nodes, il0, il1);
+    }
+}
+
+void ggml_mpi_graph_compute_post(
+        struct ggml_mpi_context * ctx_mpi,
+             struct ggml_cgraph * gf,
+                            int   n_layers) {
+    UNUSED(n_layers);
+
+    const int mpi_rank = ctx_mpi->rank;
+    const int mpi_size = ctx_mpi->size;
+
+    // send the output data to the next node
+    if (mpi_rank > 0) {
+        ggml_mpi_tensor_send(gf->nodes[gf->n_nodes - 1], (mpi_rank + 1) % mpi_size);
+    }
+}

llm/ggml-mpi.h

@@ -0,0 +1,67 @@
+//go:build mpi
+
+/**
+ * llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#pragma once
+
+struct ggml_context;
+struct ggml_tensor;
+struct ggml_cgraph;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct ggml_mpi_context;
+
+void ggml_mpi_backend_init(void);
+void ggml_mpi_backend_free(void);
+
+struct ggml_mpi_context * ggml_mpi_init(void);
+void ggml_mpi_free(struct ggml_mpi_context * ctx);
+
+int ggml_mpi_rank(struct ggml_mpi_context * ctx);
+
+void ggml_mpi_eval_init(
+        struct ggml_mpi_context * ctx_mpi,
+                            int * n_tokens,
+                            int * n_past,
+                            int * n_threads);
+
+void ggml_mpi_graph_compute_pre(
+        struct ggml_mpi_context * ctx_mpi,
+             struct ggml_cgraph * gf,
+                            int   n_layers);
+
+void ggml_mpi_graph_compute_post(
+        struct ggml_mpi_context * ctx_mpi,
+             struct ggml_cgraph * gf,
+                            int   n_layers);
+
+#ifdef __cplusplus
+}
+#endif

llm/ggml-opencl.h

@@ -0,0 +1,53 @@
+//go:build opencl
+
+/**
+ * llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#pragma once
+
+#include "ggml.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+void ggml_cl_init(void);
+
+void   ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+bool   ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+void   ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
+
+void * ggml_cl_host_malloc(size_t size);
+void   ggml_cl_host_free(void * ptr);
+
+void ggml_cl_free_data(const struct ggml_tensor* tensor);
+
+void ggml_cl_transform_tensor(void * data, struct ggml_tensor * tensor);
+
+#ifdef  __cplusplus
+}
+#endif

llm/ggml.go

@@ -0,0 +1,182 @@
+package llm
+
+import (
+	"encoding/binary"
+	"errors"
+	"fmt"
+	"io"
+)
+
+type ModelFamily string
+
+type ModelType uint32
+
+const (
+	ModelType3B  ModelType = 26
+	ModelType7B  ModelType = 32
+	ModelType13B ModelType = 40
+	ModelType30B ModelType = 60
+	ModelType65B ModelType = 80
+)
+
+func (mt ModelType) String() string {
+	switch mt {
+	case ModelType3B:
+		return "3B"
+	case ModelType7B:
+		return "7B"
+	case ModelType13B:
+		return "13B"
+	case ModelType30B:
+		return "30B"
+	case ModelType65B:
+		return "65B"
+	default:
+		return "Unknown"
+	}
+}
+
+type FileType interface {
+	String() string
+}
+
+type GGML struct {
+	magic uint32
+	container
+	model
+}
+
+type model interface {
+	ModelFamily() ModelFamily
+	ModelType() ModelType
+	FileType() FileType
+}
+
+type container interface {
+	Name() string
+	Decode(io.Reader) error
+}
+
+type containerGGML struct {
+}
+
+func (c *containerGGML) Name() string {
+	return "ggml"
+}
+
+func (c *containerGGML) Decode(r io.Reader) error {
+	return nil
+}
+
+type containerGGMF struct {
+	version uint32
+}
+
+func (c *containerGGMF) Name() string {
+	return "ggmf"
+}
+
+func (c *containerGGMF) Decode(r io.Reader) error {
+	var version uint32
+	binary.Read(r, binary.LittleEndian, &version)
+
+	switch version {
+	case 1:
+	default:
+		return errors.New("invalid version")
+	}
+
+	c.version = version
+	return nil
+}
+
+type containerGGJT struct {
+	version uint32
+}
+
+func (c *containerGGJT) Name() string {
+	return "ggjt"
+}
+
+func (c *containerGGJT) Decode(r io.Reader) error {
+	var version uint32
+	binary.Read(r, binary.LittleEndian, &version)
+
+	switch version {
+	case 1, 2, 3:
+	default:
+		return errors.New("invalid version")
+	}
+
+	c.version = version
+	return nil
+}
+
+type containerLORA struct {
+	version uint32
+}
+
+func (c *containerLORA) Name() string {
+	return "ggla"
+}
+
+func (c *containerLORA) Decode(r io.Reader) error {
+	var version uint32
+	binary.Read(r, binary.LittleEndian, &version)
+
+	switch version {
+	case 1:
+	default:
+		return errors.New("invalid version")
+	}
+
+	c.version = version
+	return nil
+}
+
+const (
+	// / Magic constant for `ggml` files (unversioned).
+	FILE_MAGIC_GGML = 0x67676d6c
+	// / Magic constant for `ggml` files (versioned, ggmf).
+	FILE_MAGIC_GGMF = 0x67676d66
+	// / Magic constant for `ggml` files (versioned, ggjt).
+	FILE_MAGIC_GGJT = 0x67676a74
+	// / Magic constant for `ggla` files (LoRA adapter).
+	FILE_MAGIC_GGLA = 0x67676C61
+)
+
+func DecodeGGML(r io.ReadSeeker, hint ModelFamily) (*GGML, error) {
+	var ggml GGML
+	binary.Read(r, binary.LittleEndian, &ggml.magic)
+
+	switch ggml.magic {
+	case FILE_MAGIC_GGML:
+		ggml.container = &containerGGML{}
+	case FILE_MAGIC_GGMF:
+		ggml.container = &containerGGMF{}
+	case FILE_MAGIC_GGJT:
+		ggml.container = &containerGGJT{}
+	case FILE_MAGIC_GGLA:
+		ggml.container = &containerLORA{}
+	default:
+		return nil, errors.New("invalid file magic")
+	}
+
+	if err := ggml.Decode(r); err != nil {
+		return nil, err
+	}
+
+	// different model types may have different layouts for hyperparameters
+	switch hint {
+	case ModelFamilyLlama:
+		var llama llamaModel
+		binary.Read(r, binary.LittleEndian, &llama.hyperparameters)
+		ggml.model = &llama
+		// TODO: sanity check hyperparameters
+	default:
+		return nil, fmt.Errorf("unsupported model type: %s", hint)
+	}
+
+	// final model type
+	return &ggml, nil
+}

llm/ggml.h

@@ -1,5 +1,5 @@
 /**
- * llama.cpp - git 5bf2a2771886ee86137e01dbc7492f78fb392066
+ * llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
  *
  * MIT License
  *
@@ -209,6 +209,15 @@
 #    define GGML_API
 #endif
 
+// TODO: support for clang
+#ifdef __GNUC__
+#    define GGML_DEPRECATED(func, hint) func __attribute__((deprecated(hint)))
+#elif defined(_MSC_VER)
+#    define GGML_DEPRECATED(func, hint) __declspec(deprecated(hint)) func
+#else
+#    define GGML_DEPRECATED(func, hint) func
+#endif
+
 #include <stdint.h>
 #include <stddef.h>
 #include <stdbool.h>
@@ -225,10 +234,17 @@
 #define GGML_MAX_CONTEXTS      64
 #define GGML_MAX_SRC           6
 #define GGML_MAX_NAME          48
+#define GGML_MAX_OP_PARAMS     32
 #define GGML_DEFAULT_N_THREADS 4
 
+
+#define GGML_EXIT_SUCCESS 0
+#define GGML_EXIT_ABORTED 1
+
 #define GGML_UNUSED(x) (void)(x)
 
+#define GGML_PAD(x, n) (((x) + (n) - 1) & ~((n) - 1))
+
 #define GGML_ASSERT(x) \
     do { \
         if (!(x)) { \
@@ -350,16 +366,6 @@ extern "C" {
         GGML_OP_ARGMAX,
         GGML_OP_REPEAT,
         GGML_OP_REPEAT_BACK,
-        GGML_OP_ABS,
-        GGML_OP_SGN,
-        GGML_OP_NEG,
-        GGML_OP_STEP,
-        GGML_OP_TANH,
-        GGML_OP_ELU,
-        GGML_OP_RELU,
-        GGML_OP_GELU,
-        GGML_OP_GELU_QUICK,
-        GGML_OP_SILU,
         GGML_OP_SILU_BACK,
         GGML_OP_NORM, // normalize
         GGML_OP_RMS_NORM,
@@ -389,6 +395,8 @@ extern "C" {
         GGML_OP_CLAMP,
         GGML_OP_CONV_1D,
         GGML_OP_CONV_2D,
+        GGML_OP_POOL_1D,
+        GGML_OP_POOL_2D,
 
         GGML_OP_FLASH_ATTN,
         GGML_OP_FLASH_FF,
@@ -396,9 +404,15 @@ extern "C" {
         GGML_OP_WIN_PART,
         GGML_OP_WIN_UNPART,
 
+        GGML_OP_UNARY,
+
         GGML_OP_MAP_UNARY,
         GGML_OP_MAP_BINARY,
 
+        GGML_OP_MAP_CUSTOM1_F32,
+        GGML_OP_MAP_CUSTOM2_F32,
+        GGML_OP_MAP_CUSTOM3_F32,
+
         GGML_OP_MAP_CUSTOM1,
         GGML_OP_MAP_CUSTOM2,
         GGML_OP_MAP_CUSTOM3,
@@ -409,6 +423,24 @@ extern "C" {
         GGML_OP_COUNT,
     };
 
+    enum ggml_unary_op {
+        GGML_UNARY_OP_ABS,
+        GGML_UNARY_OP_SGN,
+        GGML_UNARY_OP_NEG,
+        GGML_UNARY_OP_STEP,
+        GGML_UNARY_OP_TANH,
+        GGML_UNARY_OP_ELU,
+        GGML_UNARY_OP_RELU,
+        GGML_UNARY_OP_GELU,
+        GGML_UNARY_OP_GELU_QUICK,
+        GGML_UNARY_OP_SILU,
+    };
+
+    enum ggml_object_type {
+        GGML_OBJECT_TENSOR,
+        GGML_OBJECT_GRAPH,
+        GGML_OBJECT_WORK_BUFFER
+    };
 
     // ggml object
     struct ggml_object {
@@ -417,7 +449,9 @@ extern "C" {
 
         struct ggml_object * next;
 
-        char padding[8];
+        enum ggml_object_type type;
+
+        char padding[4];
     };
 
     static const size_t GGML_OBJECT_SIZE = sizeof(struct ggml_object);
@@ -437,6 +471,9 @@ extern "C" {
         // compute data
         enum ggml_op op;
 
+        // op params - allocated as int32_t for alignment
+        int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
+
         bool is_param;
 
         struct ggml_tensor * grad;
@@ -453,7 +490,7 @@ extern "C" {
 
         void * extra; // extra things e.g. for ggml-cuda.cu
 
-        char padding[8];
+        char padding[4];
     };
 
     static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
@@ -468,8 +505,17 @@ extern "C" {
 
         // the `n_tasks` of nodes, 1:1 mapping to cgraph nodes
         int n_tasks[GGML_MAX_NODES];
+
+        // abort ggml_graph_compute when true
+        bool (*abort_callback)(void * data);
+        void * abort_callback_data;
     };
 
+    // next prime after GGML_MAX_NODES
+    // #define GGML_GRAPH_HASHTABLE_SIZE 4099
+    // next prime after GGML_MAX_NODES * 2 (nodes + leafs)
+    #define GGML_GRAPH_HASHTABLE_SIZE 8273
+
     // computation graph
     struct ggml_cgraph {
         int n_nodes;
@@ -479,12 +525,16 @@ extern "C" {
         struct ggml_tensor * grads[GGML_MAX_NODES];
         struct ggml_tensor * leafs[GGML_MAX_NODES];
 
+        void * visited_hash_table[GGML_GRAPH_HASHTABLE_SIZE];
+
         // performance
         int     perf_runs;
         int64_t perf_cycles;
         int64_t perf_time_us;
     };
 
+    static const size_t GGML_GRAPH_SIZE = sizeof(struct ggml_cgraph);
+
     // scratch buffer
     struct ggml_scratch {
         size_t offs;
@@ -546,6 +596,7 @@ extern "C" {
 
     GGML_API const char * ggml_type_name(enum ggml_type type);
     GGML_API const char * ggml_op_name  (enum ggml_op   op);
+    GGML_API const char * ggml_op_symbol(enum ggml_op   op);
 
     GGML_API size_t  ggml_element_size(const struct ggml_tensor * tensor);
 
@@ -558,6 +609,8 @@ extern "C" {
     GGML_API bool ggml_is_contiguous(const struct ggml_tensor * tensor);
     GGML_API bool ggml_is_permuted  (const struct ggml_tensor * tensor);
 
+    GGML_API bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
+
     // use this to compute the memory overhead of a tensor
     GGML_API size_t ggml_tensor_overhead(void);
 
@@ -569,6 +622,7 @@ extern "C" {
     GGML_API size_t  ggml_used_mem(const struct ggml_context * ctx);
 
     GGML_API size_t  ggml_set_scratch (struct ggml_context * ctx, struct ggml_scratch scratch);
+    GGML_API bool    ggml_get_no_alloc(struct ggml_context * ctx);
     GGML_API void    ggml_set_no_alloc(struct ggml_context * ctx, bool no_alloc);
 
     GGML_API void *  ggml_get_mem_buffer     (const struct ggml_context * ctx);
@@ -628,9 +682,11 @@ extern "C" {
     GGML_API void *  ggml_get_data    (const struct ggml_tensor * tensor);
     GGML_API float * ggml_get_data_f32(const struct ggml_tensor * tensor);
 
-    GGML_API const char *         ggml_get_name(const struct ggml_tensor * tensor);
-    GGML_API struct ggml_tensor * ggml_set_name(struct ggml_tensor * tensor, const char * name);
-    GGML_API struct ggml_tensor * ggml_format_name(struct ggml_tensor * tensor, const char * fmt, ...);
+    GGML_API enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor);
+
+    GGML_API const char *         ggml_get_name   (const struct ggml_tensor * tensor);
+    GGML_API struct ggml_tensor * ggml_set_name   (      struct ggml_tensor * tensor, const char * name);
+    GGML_API struct ggml_tensor * ggml_format_name(      struct ggml_tensor * tensor, const char * fmt, ...);
 
     //
     // operations on tensors with backpropagation
@@ -640,6 +696,11 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_dup_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_add(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -864,14 +925,17 @@ extern "C" {
 
     GGML_API struct ggml_tensor * ggml_rms_norm(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a);
+            struct ggml_tensor  * a,
+            float                 eps);
 
     GGML_API struct ggml_tensor * ggml_rms_norm_inplace(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a);
+            struct ggml_tensor  * a,
+            float                 eps);
 
     // a - x
     // b - dy
+    // TODO: update with configurable eps
     GGML_API struct ggml_tensor * ggml_rms_norm_back(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -963,11 +1027,22 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    // a -> b, in-place, return view(b)
+    GGML_API struct ggml_tensor * ggml_cpy_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     // make contiguous
     GGML_API struct ggml_tensor * ggml_cont(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // make contiguous, in-place
+    GGML_API struct ggml_tensor * ggml_cont_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     // return view(a), b specifies the new shape
     // TODO: when we start computing gradient, make a copy instead of view
     GGML_API struct ggml_tensor * ggml_reshape(
@@ -1136,6 +1211,28 @@ extern "C" {
             int                   mode,
             int                   n_ctx);
 
+    // custom RoPE
+    GGML_API struct ggml_tensor * ggml_rope_custom(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   n_past,
+            int                   n_dims,
+            int                   mode,
+            int                   n_ctx,
+            float                 freq_base,
+            float                 freq_scale);
+
+    // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_rope_custom_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   n_past,
+            int                   n_dims,
+            int                   mode,
+            int                   n_ctx,
+            float                 freq_base,
+            float                 freq_scale);
+
     // rotary position embedding backward, i.e compute dx from dy
     // a - dy
     GGML_API struct ggml_tensor * ggml_rope_back(
@@ -1143,7 +1240,8 @@ extern "C" {
             struct ggml_tensor  * a,
             int                   n_past,
             int                   n_dims,
-            int                   mode);
+            int                   mode,
+            int                   n_ctx);
 
     // alibi position embedding
     // in-place, returns view(a)
@@ -1183,13 +1281,38 @@ extern "C" {
 
     // conv_1d with padding = half
     // alias for ggml_conv_1d(a, b, s, a->ne[0]/2, d)
-    GGML_API struct ggml_tensor* ggml_conv_1d_ph(
+    GGML_API struct ggml_tensor * ggml_conv_1d_ph(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
             int                   s,
             int                   d);
 
+    enum ggml_op_pool {
+        GGML_OP_POOL_MAX,
+        GGML_OP_POOL_AVG,
+        GGML_OP_POOL_COUNT,
+    };
+
+    GGML_API struct ggml_tensor * ggml_pool_1d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            enum ggml_op_pool     op,
+            int                   k0, // kernel size
+            int                   s0, // stride
+            int                   p0); // padding
+
+    GGML_API struct ggml_tensor * ggml_pool_2d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            enum ggml_op_pool     op,
+            int                   k0,
+            int                   k1,
+            int                   s0,
+            int                   s1,
+            int                   p0,
+            int                   p1);
+
     GGML_API struct ggml_tensor * ggml_flash_attn(
             struct ggml_context * ctx,
             struct ggml_tensor  * q,
@@ -1233,6 +1356,16 @@ extern "C" {
             int                   h0,
             int                   w);
 
+    GGML_API struct ggml_tensor * ggml_unary(
+            struct ggml_context * ctx,
+             struct ggml_tensor * a,
+             enum ggml_unary_op op);
+
+    GGML_API struct ggml_tensor * ggml_unary_inplace(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        enum ggml_unary_op op);
+
     // custom operators
 
     typedef void (*ggml_unary_op_f32_t) (const int, float *, const float *);
@@ -1242,63 +1375,129 @@ extern "C" {
     typedef void (*ggml_custom2_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
     typedef void (*ggml_custom3_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
 
-    GGML_API struct ggml_tensor * ggml_map_unary_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_f32(
             struct ggml_context        * ctx,
             struct ggml_tensor         * a,
-                   ggml_unary_op_f32_t   fun);
+                   ggml_unary_op_f32_t   fun),
+        "use ggml_map_custom1 instead");
 
-    GGML_API struct ggml_tensor * ggml_map_unary_inplace_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_inplace_f32(
             struct ggml_context        * ctx,
             struct ggml_tensor         * a,
-                   ggml_unary_op_f32_t   fun);
+                   ggml_unary_op_f32_t   fun),
+        "use ggml_map_custom1_inplace instead");
 
-    GGML_API struct ggml_tensor * ggml_map_binary_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_f32(
             struct ggml_context         * ctx,
             struct ggml_tensor          * a,
             struct ggml_tensor          * b,
-                   ggml_binary_op_f32_t   fun);
+                   ggml_binary_op_f32_t   fun),
+        "use ggml_map_custom2 instead");
 
-    GGML_API struct ggml_tensor * ggml_map_binary_inplace_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_inplace_f32(
             struct ggml_context         * ctx,
             struct ggml_tensor          * a,
             struct ggml_tensor          * b,
-                   ggml_binary_op_f32_t   fun);
+                   ggml_binary_op_f32_t   fun),
+        "use ggml_map_custom2_inplace instead");
 
-    GGML_API struct ggml_tensor * ggml_map_custom1_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_f32(
             struct ggml_context          * ctx,
             struct ggml_tensor           * a,
-                   ggml_custom1_op_f32_t   fun);
+                   ggml_custom1_op_f32_t   fun),
+        "use ggml_map_custom1 instead");
 
-    GGML_API struct ggml_tensor * ggml_map_custom1_inplace_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_inplace_f32(
             struct ggml_context          * ctx,
             struct ggml_tensor           * a,
-                   ggml_custom1_op_f32_t   fun);
+                   ggml_custom1_op_f32_t   fun),
+        "use ggml_map_custom1_inplace instead");
 
-    GGML_API struct ggml_tensor * ggml_map_custom2_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_f32(
             struct ggml_context          * ctx,
             struct ggml_tensor           * a,
             struct ggml_tensor           * b,
-                   ggml_custom2_op_f32_t   fun);
+                   ggml_custom2_op_f32_t   fun),
+        "use ggml_map_custom2 instead");
 
-    GGML_API struct ggml_tensor * ggml_map_custom2_inplace_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_inplace_f32(
             struct ggml_context          * ctx,
             struct ggml_tensor           * a,
             struct ggml_tensor           * b,
-                   ggml_custom2_op_f32_t   fun);
+                   ggml_custom2_op_f32_t   fun),
+        "use ggml_map_custom2_inplace instead");
 
-    GGML_API struct ggml_tensor * ggml_map_custom3_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_f32(
             struct ggml_context          * ctx,
             struct ggml_tensor           * a,
             struct ggml_tensor           * b,
             struct ggml_tensor           * c,
-                   ggml_custom3_op_f32_t   fun);
+                   ggml_custom3_op_f32_t   fun),
+        "use ggml_map_custom3 instead");
 
-    GGML_API struct ggml_tensor * ggml_map_custom3_inplace_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_inplace_f32(
             struct ggml_context          * ctx,
             struct ggml_tensor           * a,
             struct ggml_tensor           * b,
             struct ggml_tensor           * c,
-                   ggml_custom3_op_f32_t   fun);
+                   ggml_custom3_op_f32_t   fun),
+        "use ggml_map_custom3_inplace instead");
+
+    // custom operators v2
+
+    typedef void (*ggml_custom1_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, int ith, int nth, void * userdata);
+    typedef void (*ggml_custom2_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, int ith, int nth, void * userdata);
+    typedef void (*ggml_custom3_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, const struct ggml_tensor * c, int ith, int nth, void * userdata);
+
+    #define GGML_N_TASKS_MAX -1
+
+    GGML_API struct ggml_tensor * ggml_map_custom1(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            ggml_custom1_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
+
+    GGML_API struct ggml_tensor * ggml_map_custom1_inplace(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            ggml_custom1_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
+
+    GGML_API struct ggml_tensor * ggml_map_custom2(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            struct ggml_tensor    * b,
+            ggml_custom2_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
+
+    GGML_API struct ggml_tensor * ggml_map_custom2_inplace(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            struct ggml_tensor    * b,
+            ggml_custom2_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
+
+    GGML_API struct ggml_tensor * ggml_map_custom3(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            struct ggml_tensor    * b,
+            struct ggml_tensor    * c,
+            ggml_custom3_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
+
+    GGML_API struct ggml_tensor * ggml_map_custom3_inplace(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            struct ggml_tensor    * b,
+            struct ggml_tensor    * c,
+            ggml_custom3_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
 
     // loss function
 
@@ -1321,15 +1520,21 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * tensor);
 
+
     GGML_API void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
 
     GGML_API struct ggml_cgraph ggml_build_forward (struct ggml_tensor * tensor);
     GGML_API struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cgraph * gf, bool keep);
 
+    // graph allocation in a context
+    GGML_API struct ggml_cgraph * ggml_new_graph        (struct ggml_context * ctx);
+    GGML_API struct ggml_cgraph * ggml_build_forward_ctx(struct ggml_context * ctx, struct ggml_tensor * tensor);
+    GGML_API size_t ggml_graph_overhead(void);
+
     // ggml_graph_plan() has to be called before ggml_graph_compute()
     // when plan.work_size > 0, caller must allocate memory for plan.work_data
     GGML_API struct ggml_cplan ggml_graph_plan   (struct ggml_cgraph * cgraph, int n_threads /*= GGML_DEFAULT_N_THREADS*/);
-    GGML_API              void ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
+    GGML_API               int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
     GGML_API              void ggml_graph_reset  (struct ggml_cgraph * cgraph);
 
     // same as ggml_graph_compute() but the work data is allocated as a part of the context

llm/k_quants.c

@@ -1,5 +1,5 @@
 /**
- * llama.cpp - git 5bf2a2771886ee86137e01dbc7492f78fb392066
+ * llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
  *
  * MIT License
  *
@@ -65,6 +65,8 @@
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 
+#define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
+
 //
 // 2-6 bit quantization in super-blocks
 //
@@ -1379,7 +1381,7 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
         const __m256i all_scales = _mm256_cvtepi8_epi16(scales8);
         const __m128i l_scales = _mm256_extracti128_si256(all_scales, 0);
         const __m128i h_scales = _mm256_extracti128_si256(all_scales, 1);
-        const __m256i scales[2] = {_mm256_set_m128i(l_scales, l_scales), _mm256_set_m128i(h_scales, h_scales)};
+        const __m256i scales[2] = {MM256_SET_M128I(l_scales, l_scales), MM256_SET_M128I(h_scales, h_scales)};
 
         __m256i sumi = _mm256_setzero_si256();
 
@@ -1447,7 +1449,7 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
         const __m128i summs_1 = _mm_madd_epi16(mins_1, _mm_loadu_si128((const __m128i*)&y[i].bsums[8]));
 
         // sumf += -dmin * summs in 32bits*8
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dmin), _mm256_cvtepi32_ps(_mm256_set_m128i(summs_1, summs_0))), acc);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dmin), _mm256_cvtepi32_ps(MM256_SET_M128I(summs_1, summs_0))), acc);
 
         const __m128i scales_0 = _mm_cvtepi8_epi16(scales16);
         const __m128i scales_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(scales16, scales16));
@@ -1519,7 +1521,7 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
         }
 
         // sumf += dall * isum - dmin * summs in 32bits
-        __m256i sumi = _mm256_set_m128i(sumi_1, sumi_0);
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
         acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dall), _mm256_cvtepi32_ps(sumi)), acc);
     }
 
@@ -1670,8 +1672,8 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
         summs += dmin * smin;
 
         const __m128i q2bits = _mm_loadu_si128((const __m128i*)q2);
-        const __m256i q2_0 = _mm256_and_si256(_mm256_set_m128i(_mm_srli_epi16(q2bits, 2), q2bits), m3);
-        const __m256i q2_1 = _mm256_and_si256(_mm256_set_m128i(_mm_srli_epi16(q2bits, 6), _mm_srli_epi16(q2bits, 4)), m3);
+        const __m256i q2_0 = _mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(q2bits, 2), q2bits), m3);
+        const __m256i q2_1 = _mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(q2bits, 6), _mm_srli_epi16(q2bits, 4)), m3);
 
         const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
         const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
@@ -1692,6 +1694,62 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc) + summs;
 
+#elif defined __AVX__
+
+    const __m128i m3 = _mm_set1_epi8(3);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    uint32_t ud, um;
+    const uint8_t * restrict db = (const uint8_t *)&ud;
+    const uint8_t * restrict mb = (const uint8_t *)&um;
+
+    float summs = 0;
+
+    // TODO: optimize this
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
+        const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin);
+
+        const uint8_t * restrict q2 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const uint32_t * restrict sc = (const uint32_t *)x[i].scales;
+        ud = (sc[0] >> 0) & 0x0f0f0f0f;
+        um = (sc[0] >> 4) & 0x0f0f0f0f;
+
+        int32_t smin = mb[0] * y[i].bsums[0] + mb[1] * y[i].bsums[1] + mb[2] * y[i].bsums[2] + mb[3] * y[i].bsums[3];
+        summs += dmin * smin;
+
+        const __m128i q2bits = _mm_loadu_si128((const __m128i*)q2);
+        const __m128i q2_0 = _mm_and_si128(q2bits, m3);
+        const __m128i q2_1 = _mm_and_si128(_mm_srli_epi16(q2bits, 2), m3);
+        const __m128i q2_2 = _mm_and_si128(_mm_srli_epi16(q2bits, 4), m3);
+        const __m128i q2_3 = _mm_and_si128(_mm_srli_epi16(q2bits, 6), m3);
+
+        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
+        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
+
+        const __m128i p0 = _mm_maddubs_epi16(q2_0, _mm256_extractf128_si256(q8_0, 0));
+        const __m128i p1 = _mm_maddubs_epi16(q2_1, _mm256_extractf128_si256(q8_0, 1));
+        const __m128i p2 = _mm_maddubs_epi16(q2_2, _mm256_extractf128_si256(q8_1, 0));
+        const __m128i p3 = _mm_maddubs_epi16(q2_3, _mm256_extractf128_si256(q8_1, 1));
+
+        const __m256i p_0 = MM256_SET_M128I(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p0, p0)), _mm_cvtepi16_epi32(p0));
+        const __m256i p_1 = MM256_SET_M128I(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p1, p1)), _mm_cvtepi16_epi32(p1));
+        const __m256i p_2 = MM256_SET_M128I(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p2, p2)), _mm_cvtepi16_epi32(p2));
+        const __m256i p_3 = MM256_SET_M128I(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p3, p3)), _mm_cvtepi16_epi32(p3));
+
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[0]), _mm256_cvtepi32_ps(p_0)), acc);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[1]), _mm256_cvtepi32_ps(p_1)), acc);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[2]), _mm256_cvtepi32_ps(p_2)), acc);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[3]), _mm256_cvtepi32_ps(p_3)), acc);
+    }
+
+    *s = hsum_float_8(acc) + summs;
+
 #else
 
     float sumf = 0;
@@ -1887,7 +1945,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
         const __m256i all_scales = _mm256_cvtepi8_epi16(scales128);
         const __m128i l_scales = _mm256_extracti128_si256(all_scales, 0);
         const __m128i h_scales = _mm256_extracti128_si256(all_scales, 1);
-        const __m256i scales[2] = {_mm256_set_m128i(l_scales, l_scales), _mm256_set_m128i(h_scales, h_scales)};
+        const __m256i scales[2] = {MM256_SET_M128I(l_scales, l_scales), MM256_SET_M128I(h_scales, h_scales)};
 
         // high bit
         const __m256i hbits = _mm256_loadu_si256((const __m256i*)x[i].hmask);
@@ -2098,7 +2156,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
         }
 
         // multiply with block scale and accumulate
-        __m256i sumi = _mm256_set_m128i(sumi_1, sumi_0);
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
         acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi)), acc);
 
     }
@@ -2273,13 +2331,13 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
         aux16[0] = a & 0x0f0f;
         aux16[1] = (a >> 4) & 0x0f0f;
 
-        const __m256i scale_0 = _mm256_set_m128i(_mm_set1_epi16(aux8[2] - 8), _mm_set1_epi16(aux8[0] - 8));
-        const __m256i scale_1 = _mm256_set_m128i(_mm_set1_epi16(aux8[3] - 8), _mm_set1_epi16(aux8[1] - 8));
+        const __m256i scale_0 = MM256_SET_M128I(_mm_set1_epi16(aux8[2] - 8), _mm_set1_epi16(aux8[0] - 8));
+        const __m256i scale_1 = MM256_SET_M128I(_mm_set1_epi16(aux8[3] - 8), _mm_set1_epi16(aux8[1] - 8));
 
         memcpy(&aux64, x[i].hmask, 8);
 
         const __m128i haux = _mm_set_epi64x(aux64 >> 1, aux64 >> 0);
-        __m256i q3h_0 = _mm256_set_m128i(_mm_srli_epi16(haux, 2), haux);
+        __m256i q3h_0 = MM256_SET_M128I(_mm_srli_epi16(haux, 2), haux);
         __m256i q3h_1 = _mm256_srli_epi16(q3h_0, 4);
         q3h_0 = _mm256_slli_epi16(_mm256_andnot_si256(q3h_0, m1), 2);
         q3h_1 = _mm256_slli_epi16(_mm256_andnot_si256(q3h_1, m1), 2);
@@ -2288,7 +2346,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
         const __m128i q3bits = _mm_loadu_si128((const __m128i*)q3);
 
         // prepare low and high bits
-        const __m256i q3aux  = _mm256_set_m128i(_mm_srli_epi16(q3bits, 2), q3bits);
+        const __m256i q3aux  = MM256_SET_M128I(_mm_srli_epi16(q3bits, 2), q3bits);
         const __m256i q3l_0 = _mm256_and_si256(q3aux, m3);
         const __m256i q3l_1 = _mm256_and_si256(_mm256_srli_epi16(q3aux, 4), m3);
 
@@ -2321,6 +2379,93 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc);
 
+#elif defined __AVX__
+
+    const __m128i m3 = _mm_set1_epi8(3);
+    const __m128i m1 = _mm_set1_epi8(1);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    uint64_t aux64;
+
+    uint16_t aux16[2];
+    const int8_t * aux8 = (const int8_t *)aux16;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
+
+        const uint8_t * restrict q3 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const uint16_t a = *(const uint16_t *)x[i].scales;
+        aux16[0] = a & 0x0f0f;
+        aux16[1] = (a >> 4) & 0x0f0f;
+
+        const __m128i scale_0 = _mm_set1_epi16(aux8[0] - 8);
+        const __m128i scale_1 = _mm_set1_epi16(aux8[2] - 8);
+        const __m128i scale_2 = _mm_set1_epi16(aux8[1] - 8);
+        const __m128i scale_3 = _mm_set1_epi16(aux8[3] - 8);
+
+        memcpy(&aux64, x[i].hmask, 8);
+
+        __m128i q3h_0 = _mm_set_epi64x(aux64 >> 1, aux64 >> 0);
+        __m128i q3h_1 = _mm_srli_epi16(q3h_0, 2);
+        __m128i q3h_2 = _mm_srli_epi16(q3h_0, 4);
+        __m128i q3h_3 = _mm_srli_epi16(q3h_0, 6);
+        q3h_0 = _mm_slli_epi16(_mm_andnot_si128(q3h_0, m1), 2);
+        q3h_1 = _mm_slli_epi16(_mm_andnot_si128(q3h_1, m1), 2);
+        q3h_2 = _mm_slli_epi16(_mm_andnot_si128(q3h_2, m1), 2);
+        q3h_3 = _mm_slli_epi16(_mm_andnot_si128(q3h_3, m1), 2);
+
+        // load low 2 bits
+        const __m128i q3bits = _mm_loadu_si128((const __m128i*)q3);
+
+        // prepare low and high bits
+        const __m128i q3l_0 = _mm_and_si128(q3bits, m3);
+        const __m128i q3l_1 = _mm_and_si128(_mm_srli_epi16(q3bits, 2), m3);
+        const __m128i q3l_2 = _mm_and_si128(_mm_srli_epi16(q3bits, 4), m3);
+        const __m128i q3l_3 = _mm_and_si128(_mm_srli_epi16(q3bits, 6), m3);
+
+        // load Q8 quants
+        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
+        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
+
+        // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use _mm_maddubs_epi16,
+        // and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set,
+        // and 2 if the high bit was set)
+        const __m128i q8s_0 = _mm_maddubs_epi16(q3h_0, _mm256_extractf128_si256(q8_0, 0));
+        const __m128i q8s_1 = _mm_maddubs_epi16(q3h_1, _mm256_extractf128_si256(q8_0, 1));
+        const __m128i q8s_2 = _mm_maddubs_epi16(q3h_2, _mm256_extractf128_si256(q8_1, 0));
+        const __m128i q8s_3 = _mm_maddubs_epi16(q3h_3, _mm256_extractf128_si256(q8_1, 1));
+
+        __m128i p16_0 = _mm_maddubs_epi16(q3l_0, _mm256_extractf128_si256(q8_0, 0));
+        __m128i p16_1 = _mm_maddubs_epi16(q3l_1, _mm256_extractf128_si256(q8_0, 1));
+        __m128i p16_2 = _mm_maddubs_epi16(q3l_2, _mm256_extractf128_si256(q8_1, 0));
+        __m128i p16_3 = _mm_maddubs_epi16(q3l_3, _mm256_extractf128_si256(q8_1, 1));
+
+        p16_0 = _mm_sub_epi16(p16_0, q8s_0);
+        p16_1 = _mm_sub_epi16(p16_1, q8s_1);
+        p16_2 = _mm_sub_epi16(p16_2, q8s_2);
+        p16_3 = _mm_sub_epi16(p16_3, q8s_3);
+
+        // multiply with scales
+        p16_0 = _mm_madd_epi16(scale_0, p16_0);
+        p16_1 = _mm_madd_epi16(scale_1, p16_1);
+        p16_2 = _mm_madd_epi16(scale_2, p16_2);
+        p16_3 = _mm_madd_epi16(scale_3, p16_3);
+
+        p16_0 = _mm_add_epi32(p16_0, p16_2);
+        p16_1 = _mm_add_epi32(p16_1, p16_3);
+        __m256i p16 = MM256_SET_M128I(p16_1, p16_0);
+
+        // multiply with block scale and accumulate
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(p16)), acc);
+
+    }
+
+    *s = hsum_float_8(acc);
+
 #else
 
     int8_t  aux8[QK_K];
@@ -2503,7 +2648,7 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
         acc_m = _mm_fmadd_ps(_mm_set1_ps(dmin), _mm_cvtepi32_ps(prod), acc_m);
 
         const __m128i sc128  = _mm256_extracti128_si256(mins_and_scales, 0);
-        const __m256i scales = _mm256_set_m128i(sc128, sc128);
+        const __m256i scales = MM256_SET_M128I(sc128, sc128);
 
         __m256i sumi = _mm256_setzero_si256();
 
@@ -2610,7 +2755,7 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
         }
 
         __m256 vd = _mm256_set1_ps(d);
-        __m256i sumi = _mm256_set_m128i(sumi_1, sumi_0);
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
         acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(sumi)), acc);
 
     }
@@ -2807,6 +2952,60 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc) - summs;
 
+#elif defined __AVX__
+
+    const __m128i m4 = _mm_set1_epi8(0xF);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    float summs = 0;
+
+    uint16_t aux16[2];
+    const uint8_t * scales = (const uint8_t *)aux16;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = ggml_fp16_to_fp32(x[i].d[0]) * y[i].d;
+        const float m = ggml_fp16_to_fp32(x[i].d[1]) * y[i].d;
+        const __m256 vd = _mm256_set1_ps(d);
+
+        const uint16_t * a = (const uint16_t *)x[i].scales;
+        aux16[0] = a[0] & 0x0f0f;
+        aux16[1] = (a[0] >> 4) & 0x0f0f;
+
+        summs += m * (scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3]));
+
+        const uint8_t * restrict q4 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const __m256i q4bits = _mm256_loadu_si256((const __m256i*)q4);
+        const __m128i q4bits_0 = _mm256_extractf128_si256(q4bits, 0);
+        const __m128i q4bits_1 = _mm256_extractf128_si256(q4bits, 1);
+        const __m128i q4_0 = _mm_and_si128(q4bits_0, m4);
+        const __m128i q4_1 = _mm_and_si128(q4bits_1, m4);
+        const __m128i q4_2 = _mm_and_si128(_mm_srli_epi16(q4bits_0, 4), m4);
+        const __m128i q4_3 = _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4);
+
+        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
+        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
+
+        const __m128i p16_0 = _mm_maddubs_epi16(q4_0, _mm256_extractf128_si256(q8_0, 0));
+        const __m128i p16_1 = _mm_maddubs_epi16(q4_1, _mm256_extractf128_si256(q8_0, 1));
+        const __m128i p16_2 = _mm_maddubs_epi16(q4_2, _mm256_extractf128_si256(q8_1, 0));
+        const __m128i p16_3 = _mm_maddubs_epi16(q4_3, _mm256_extractf128_si256(q8_1, 1));
+
+        const __m128i p32_0 = _mm_madd_epi16(_mm_set1_epi16(scales[0]), p16_0);
+        const __m128i p32_1 = _mm_madd_epi16(_mm_set1_epi16(scales[0]), p16_1);
+        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(MM256_SET_M128I(p32_1, p32_0))), acc);
+
+        const __m128i p32_2 = _mm_madd_epi16(_mm_set1_epi16(scales[1]), p16_2);
+        const __m128i p32_3 = _mm_madd_epi16(_mm_set1_epi16(scales[1]), p16_3);
+        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(MM256_SET_M128I(p32_3, p32_2))), acc);
+
+    }
+
+    *s = hsum_float_8(acc) - summs;
+
 #else
 
     uint8_t aux8[QK_K];
@@ -2989,7 +3188,7 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
         summs += dmin * _mm_extract_epi32(hsum, 0);
 
         const __m128i sc128  = _mm256_extracti128_si256(mins_and_scales, 0);
-        const __m256i scales = _mm256_set_m128i(sc128, sc128);
+        const __m256i scales = MM256_SET_M128I(sc128, sc128);
 
         const __m256i hbits = _mm256_loadu_si256((const __m256i*)x[i].qh);
         __m256i hmask = mone;
@@ -3128,7 +3327,7 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
         }
 
         __m256 vd = _mm256_set1_ps(d);
-        __m256i sumi = _mm256_set_m128i(sumi_1, sumi_0);
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
         acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(sumi)), acc);
 
     }
@@ -3291,13 +3490,13 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
 
         const __m256i q5bits = _mm256_loadu_si256((const __m256i*)q5);
 
-        const __m256i scale_l = _mm256_set_m128i(_mm_set1_epi16(x[i].scales[1]), _mm_set1_epi16(x[i].scales[0]));
-        const __m256i scale_h = _mm256_set_m128i(_mm_set1_epi16(x[i].scales[3]), _mm_set1_epi16(x[i].scales[2]));
+        const __m256i scale_l = MM256_SET_M128I(_mm_set1_epi16(x[i].scales[1]), _mm_set1_epi16(x[i].scales[0]));
+        const __m256i scale_h = MM256_SET_M128I(_mm_set1_epi16(x[i].scales[3]), _mm_set1_epi16(x[i].scales[2]));
 
         int64_t aux64;
         memcpy(&aux64, x[i].qh, 8);
         const __m128i haux128 = _mm_set_epi64x(aux64 >> 1, aux64);
-        const __m256i haux256 = _mm256_set_m128i(_mm_srli_epi16(haux128, 2), haux128);
+        const __m256i haux256 = MM256_SET_M128I(_mm_srli_epi16(haux128, 2), haux128);
 
         const __m256i q5h_0 = _mm256_slli_epi16(_mm256_andnot_si256(haux256, mone), 4);
         const __m256i q5h_1 = _mm256_slli_epi16(_mm256_andnot_si256(_mm256_srli_epi16(haux256, 4), mone), 4);
@@ -3321,10 +3520,66 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc);
 
+#elif defined __AVX__
+
+    const __m128i m4 = _mm_set1_epi8(0xF);
+    const __m128i mone  = _mm_set1_epi8(1);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+
+        const uint8_t * restrict q5 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
+
+        const __m256i q5bits = _mm256_loadu_si256((const __m256i*)q5);
+
+        const __m128i scale_0 = _mm_set1_epi16(x[i].scales[0]);
+        const __m128i scale_1 = _mm_set1_epi16(x[i].scales[1]);
+        const __m128i scale_2 = _mm_set1_epi16(x[i].scales[2]);
+        const __m128i scale_3 = _mm_set1_epi16(x[i].scales[3]);
+
+        int64_t aux64;
+        memcpy(&aux64, x[i].qh, 8);
+        const __m128i haux128_0 = _mm_set_epi64x(aux64 >> 1, aux64);
+        const __m128i haux128_1 = _mm_srli_epi16(haux128_0, 2);
+
+        const __m128i q5h_0 = _mm_slli_epi16(_mm_andnot_si128(haux128_0, mone), 4);
+        const __m128i q5h_1 = _mm_slli_epi16(_mm_andnot_si128(haux128_1, mone), 4);
+        const __m128i q5h_2 = _mm_slli_epi16(_mm_andnot_si128(_mm_srli_epi16(haux128_0, 4), mone), 4);
+        const __m128i q5h_3 = _mm_slli_epi16(_mm_andnot_si128(_mm_srli_epi16(haux128_1, 4), mone), 4);
+
+        const __m128i q5l_0 = _mm_and_si128(_mm256_extractf128_si256(q5bits, 0), m4);
+        const __m128i q5l_1 = _mm_and_si128(_mm256_extractf128_si256(q5bits, 1), m4);
+        const __m128i q5l_2 = _mm_and_si128(_mm_srli_epi16(_mm256_extractf128_si256(q5bits, 0), 4), m4);
+        const __m128i q5l_3 = _mm_and_si128(_mm_srli_epi16(_mm256_extractf128_si256(q5bits, 1), 4), m4);
+
+        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
+        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
+
+        const __m128i p16_0 = _mm_madd_epi16(scale_0, _mm_maddubs_epi16(q5l_0, _mm256_extractf128_si256(q8_0, 0)));
+        const __m128i p16_1 = _mm_madd_epi16(scale_1, _mm_maddubs_epi16(q5l_1, _mm256_extractf128_si256(q8_0, 1)));
+        const __m128i p16_2 = _mm_madd_epi16(scale_2, _mm_maddubs_epi16(q5l_2, _mm256_extractf128_si256(q8_1, 0)));
+        const __m128i p16_3 = _mm_madd_epi16(scale_3, _mm_maddubs_epi16(q5l_3, _mm256_extractf128_si256(q8_1, 1)));
+        const __m128i s16_0 = _mm_madd_epi16(scale_0, _mm_maddubs_epi16(q5h_0, _mm256_extractf128_si256(q8_0, 0)));
+        const __m128i s16_1 = _mm_madd_epi16(scale_1, _mm_maddubs_epi16(q5h_1, _mm256_extractf128_si256(q8_0, 1)));
+        const __m128i s16_2 = _mm_madd_epi16(scale_2, _mm_maddubs_epi16(q5h_2, _mm256_extractf128_si256(q8_1, 0)));
+        const __m128i s16_3 = _mm_madd_epi16(scale_3, _mm_maddubs_epi16(q5h_3, _mm256_extractf128_si256(q8_1, 1)));
+
+        const __m128i dot_0 = _mm_sub_epi32(_mm_add_epi32(p16_0, p16_2), _mm_add_epi32(s16_0, s16_2));
+        const __m128i dot_1 = _mm_sub_epi32(_mm_add_epi32(p16_1, p16_3), _mm_add_epi32(s16_1, s16_3));
+
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(dot_1, dot_0))), acc);
+
+    }
+
+    *s = hsum_float_8(acc);
+
 #else
 
-
-    uint8_t aux8[QK_K];
+    int8_t aux8[QK_K];
     int16_t aux16[16];
     float   sums [8];
     memset(sums, 0, 8*sizeof(float));
@@ -3334,7 +3589,7 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
         const uint8_t * restrict q4 = x[i].qs;
         const uint8_t * restrict hm = x[i].qh;
         const  int8_t * restrict q8 = y[i].qs;
-        uint8_t * restrict a = aux8;
+        int8_t * restrict a = aux8;
         for (int l = 0; l < 32; ++l) {
             a[l+ 0] = q4[l] & 0xF;
             a[l+32] = q4[l]  >> 4;
@@ -3698,7 +3953,7 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
 
         }
 
-        __m256i sumi = _mm256_set_m128i(sumi_1, sumi_0);
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
         acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi)), acc);
     }
 
@@ -3856,8 +4111,8 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
         const __m256i q4bits1 = _mm256_loadu_si256((const __m256i*)q4);
         const __m128i q4bitsH = _mm_loadu_si128((const __m128i*)qh);
 
-        const __m256i q4h_0 = _mm256_slli_epi16(_mm256_and_si256(_mm256_set_m128i(_mm_srli_epi16(q4bitsH, 2), q4bitsH), m2), 4);
-        const __m256i q4h_1 = _mm256_slli_epi16(_mm256_and_si256(_mm256_set_m128i(_mm_srli_epi16(q4bitsH, 6), _mm_srli_epi16(q4bitsH, 4)), m2), 4);
+        const __m256i q4h_0 = _mm256_slli_epi16(_mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(q4bitsH, 2), q4bitsH), m2), 4);
+        const __m256i q4h_1 = _mm256_slli_epi16(_mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(q4bitsH, 6), _mm_srli_epi16(q4bitsH, 4)), m2), 4);
 
         const __m256i q4_0 = _mm256_or_si256(_mm256_and_si256(q4bits1, m4), q4h_0);
         const __m256i q4_1 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits1, 4), m4), q4h_1);
@@ -3884,6 +4139,77 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc);
 
+#elif defined __AVX__
+
+    const __m128i m4 = _mm_set1_epi8(0xF);
+    const __m128i m2 = _mm_set1_epi8(3);
+    const __m128i m32s = _mm_set1_epi8(32);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
+
+        const uint8_t * restrict q4 = x[i].ql;
+        const uint8_t * restrict qh = x[i].qh;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const __m64 scales_1 = _mm_set1_pi8(x[i].scales[0]);
+        const __m64 scales_2 = _mm_set1_pi8(x[i].scales[1]);
+        const __m64 scales_3 = _mm_set1_pi8(x[i].scales[2]);
+        const __m64 scales_4 = _mm_set1_pi8(x[i].scales[3]);
+
+        __m128i sumi_0 = _mm_setzero_si128();
+        __m128i sumi_1 = _mm_setzero_si128();
+
+        const __m128i scale_0 = _mm_set_epi64(scales_2, scales_1);
+        const __m128i scale_1 = _mm_set_epi64(scales_4, scales_3);
+
+        const __m256i q4bits1 = _mm256_loadu_si256((const __m256i*)q4);
+        const __m128i q4bitsH = _mm_loadu_si128((const __m128i*)qh);
+
+        const __m128i q4h_0 = _mm_slli_epi16(_mm_and_si128(q4bitsH, m2), 4);
+        const __m128i q4h_1 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH, 2), m2), 4);
+        const __m128i q4h_2 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH, 4), m2), 4);
+        const __m128i q4h_3 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH, 6), m2), 4);
+
+        const __m128i q4_0 = _mm_or_si128(_mm_and_si128(_mm256_extractf128_si256(q4bits1, 0), m4), q4h_0);
+        const __m128i q4_1 = _mm_or_si128(_mm_and_si128(_mm256_extractf128_si256(q4bits1, 1), m4), q4h_1);
+        const __m128i q4_2 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(_mm256_extractf128_si256(q4bits1, 0), 4), m4), q4h_2);
+        const __m128i q4_3 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(_mm256_extractf128_si256(q4bits1, 1), 4), m4), q4h_3);
+
+        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
+        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
+
+        __m128i q8s_0 = _mm_maddubs_epi16(m32s, _mm256_extractf128_si256(q8_0, 0));
+        __m128i q8s_1 = _mm_maddubs_epi16(m32s, _mm256_extractf128_si256(q8_0, 1));
+        __m128i q8s_2 = _mm_maddubs_epi16(m32s, _mm256_extractf128_si256(q8_1, 0));
+        __m128i q8s_3 = _mm_maddubs_epi16(m32s, _mm256_extractf128_si256(q8_1, 1));
+
+        __m128i p16_0 = _mm_maddubs_epi16(q4_0, _mm256_extractf128_si256(q8_0, 0));
+        __m128i p16_1 = _mm_maddubs_epi16(q4_1, _mm256_extractf128_si256(q8_0, 1));
+        __m128i p16_2 = _mm_maddubs_epi16(q4_2, _mm256_extractf128_si256(q8_1, 0));
+        __m128i p16_3 = _mm_maddubs_epi16(q4_3, _mm256_extractf128_si256(q8_1, 1));
+
+        p16_0 = _mm_sub_epi16(p16_0, q8s_0);
+        p16_1 = _mm_sub_epi16(p16_1, q8s_1);
+        p16_2 = _mm_sub_epi16(p16_2, q8s_2);
+        p16_3 = _mm_sub_epi16(p16_3, q8s_3);
+
+        p16_0 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_0), p16_0);
+        p16_1 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_0, scale_0)), p16_1);
+        p16_2 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_1), p16_2);
+        p16_3 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_1, scale_1)), p16_3);
+
+        sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
+        sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
+
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(MM256_SET_M128I(sumi_1, sumi_0))), acc);
+    }
+
+    *s = hsum_float_8(acc);
+
 #else
 
     int8_t  aux8[QK_K];

llm/k_quants.h

@@ -1,5 +1,5 @@
 /**
- * llama.cpp - git 5bf2a2771886ee86137e01dbc7492f78fb392066
+ * llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
  *
  * MIT License
  *
@@ -41,6 +41,14 @@
 #define K_SCALE_SIZE 12
 #endif
 
+#ifndef static_assert
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
+#define static_assert(cond, msg) _Static_assert(cond, msg)
+#else
+#define static_assert(cond, msg) struct global_scope_noop_trick
+#endif
+#endif
+
 //
 // Super-block quantization structures
 //

llm/llama-util.h

@@ -1,5 +1,5 @@
 /**
- * llama.cpp - git 5bf2a2771886ee86137e01dbc7492f78fb392066
+ * llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
  *
  * MIT License
  *
@@ -175,6 +175,46 @@ struct llama_file {
     }
 };
 
+// llama_context_data
+struct llama_data_context {
+    virtual void write(const void * src, size_t size) = 0;
+    virtual size_t get_size_written() = 0;
+    virtual ~llama_data_context() = default;
+};
+
+struct llama_data_buffer_context : llama_data_context {
+    uint8_t* ptr;
+    size_t size_written = 0;
+
+    llama_data_buffer_context(uint8_t * p) : ptr(p) {}
+
+    void write(const void * src, size_t size) override {
+        memcpy(ptr, src, size);
+        ptr += size;
+        size_written += size;
+    }
+
+    size_t get_size_written() override {
+        return size_written;
+    }
+};
+
+struct llama_data_file_context : llama_data_context {
+    llama_file* file;
+    size_t size_written = 0;
+
+    llama_data_file_context(llama_file * f) : file(f) {}
+
+    void write(const void * src, size_t size) override {
+        file->write_raw(src, size);
+        size_written += size;
+    }
+
+    size_t get_size_written() override {
+        return size_written;
+    }
+};
+
 #if defined(_WIN32)
 static std::string llama_format_win_err(DWORD err) {
     LPSTR buf;
@@ -201,13 +241,13 @@ struct llama_mmap {
     llama_mmap(struct llama_file * file, size_t prefetch = (size_t) -1 /* -1 = max value */, bool numa = false) {
         size = file->size;
         int fd = fileno(file->fp);
-        int flags = MAP_PRIVATE;
+        int flags = MAP_SHARED;
         // prefetch/readahead impairs performance on NUMA systems
         if (numa) { prefetch = 0; }
 #ifdef __linux__
-        if (prefetch) { flags |= MAP_POPULATE; }
+        if (prefetch >= file->size) { flags |= MAP_POPULATE; }
 #endif
-        addr = mmap(NULL, file->size, PROT_READ | PROT_WRITE, flags, fd, 0);
+        addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0);
         if (addr == MAP_FAILED) {
             throw std::runtime_error(format("mmap failed: %s", strerror(errno)));
         }
@@ -249,7 +289,7 @@ struct llama_mmap {
             throw std::runtime_error(format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str()));
         }
 
-        addr = MapViewOfFile(hMapping, FILE_MAP_COPY, 0, 0, 0);
+        addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0);
         error = GetLastError();
         CloseHandle(hMapping);
 
@@ -257,20 +297,29 @@ struct llama_mmap {
             throw std::runtime_error(format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str()));
         }
 
-        #if _WIN32_WINNT >= _WIN32_WINNT_WIN8
         if (prefetch) {
-            // Advise the kernel to preload the mapped memory
-            WIN32_MEMORY_RANGE_ENTRY range;
-            range.VirtualAddress = addr;
-            range.NumberOfBytes = (SIZE_T)size;
-            if (!PrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
-                fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n",
-                        llama_format_win_err(GetLastError()).c_str());
+            // The PrefetchVirtualMemory API is only present on Windows 8 and above, so we
+            // will dynamically load it using GetProcAddress.
+            BOOL (WINAPI *pPrefetchVirtualMemory) (HANDLE, ULONG_PTR, PWIN32_MEMORY_RANGE_ENTRY, ULONG);
+            HMODULE hKernel32;
+
+            // This call is guaranteed to succeed.
+            hKernel32 = GetModuleHandleW(L"kernel32.dll");
+
+            // This call may fail if on a pre-Win8 system.
+            pPrefetchVirtualMemory = reinterpret_cast<decltype(pPrefetchVirtualMemory)> (GetProcAddress(hKernel32, "PrefetchVirtualMemory"));
+
+            if (pPrefetchVirtualMemory) {
+                // Advise the kernel to preload the mapped memory.
+                WIN32_MEMORY_RANGE_ENTRY range;
+                range.VirtualAddress = addr;
+                range.NumberOfBytes = (SIZE_T)size;
+                if (!pPrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
+                    fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n",
+                            llama_format_win_err(GetLastError()).c_str());
+                }
             }
         }
-        #else
-        #pragma message("warning: You are building for pre-Windows 8; prefetch not supported")
-        #endif // _WIN32_WINNT >= _WIN32_WINNT_WIN8
     }
 
     ~llama_mmap() {

llm/llama.go

@@ -0,0 +1,569 @@
+package llm
+
+/*
+#cgo CFLAGS: -Ofast -std=c11 -fPIC
+#cgo CPPFLAGS: -Ofast -Wall -Wextra -Wno-unused-function -Wno-unused-variable -DNDEBUG -DGGML_USE_K_QUANTS
+#cgo CXXFLAGS: -std=c++11 -fPIC
+#cgo darwin CPPFLAGS:  -DGGML_USE_ACCELERATE
+#cgo darwin,arm64 CPPFLAGS: -DGGML_USE_METAL -DGGML_METAL_NDEBUG
+#cgo darwin LDFLAGS: -framework Accelerate -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders
+#include <stdlib.h>
+#include "llama.h"
+
+struct llama_sample_options
+{
+	float repeat_penalty;
+	float frequency_penalty;
+	float presence_penalty;
+	float temperature;
+	int32_t top_k;
+	float top_p;
+	float tfs_z;
+	float typical_p;
+	int mirostat;
+	float mirostat_tau;
+	float mirostat_eta;
+	bool penalize_newline;
+};
+
+llama_token llama_sample(
+		struct llama_context *ctx,
+		struct llama_token_data *candidates,
+		size_t n_candidates,
+		const llama_token *last_tokens,
+		size_t n_last_tokens,
+		struct llama_sample_options *opts)
+{
+	llama_token_data_array candidates_p = {
+		candidates,
+		n_candidates,
+		false,
+	};
+
+	struct llama_token_data newline = candidates_p.data[llama_token_nl()];
+
+	llama_sample_repetition_penalty(
+		ctx, &candidates_p,
+		last_tokens, n_last_tokens,
+		opts->repeat_penalty);
+
+	llama_sample_frequency_and_presence_penalties(
+		ctx, &candidates_p,
+		last_tokens, n_last_tokens,
+		opts->frequency_penalty, opts->presence_penalty);
+
+	if (!opts->penalize_newline) {
+		candidates_p.data[llama_token_nl()] = newline;
+	}
+
+	if (opts->temperature <= 0) {
+		return llama_sample_token_greedy(ctx, &candidates_p);
+	}
+
+	if (opts->mirostat == 1) {
+		int mirostat_m = 100;
+		float mirostat_mu = 2.0f * opts->mirostat_tau;
+		llama_sample_temperature(ctx, &candidates_p, opts->temperature);
+		return llama_sample_token_mirostat(
+			ctx, &candidates_p,
+			opts->mirostat_tau, opts->mirostat_eta,
+			mirostat_m, &mirostat_mu);
+	} else if (opts->mirostat == 2) {
+		float mirostat_mu = 2.0f * opts->mirostat_tau;
+		llama_sample_temperature(ctx, &candidates_p, opts->temperature);
+		return llama_sample_token_mirostat_v2(
+			ctx, &candidates_p,
+			opts->mirostat_tau, opts->mirostat_eta,
+			&mirostat_mu);
+	} else {
+		llama_sample_top_k(ctx, &candidates_p, opts->top_k, 1);
+		llama_sample_tail_free(ctx, &candidates_p, opts->tfs_z, 1);
+		llama_sample_typical(ctx, &candidates_p, opts->typical_p, 1);
+		llama_sample_top_p(ctx, &candidates_p, opts->top_p, 1);
+		llama_sample_temperature(ctx, &candidates_p, opts->temperature);
+		return llama_sample_token(ctx, &candidates_p);
+	}
+}
+*/
+import "C"
+
+import (
+	"bytes"
+	"embed"
+	"errors"
+	"fmt"
+	"io"
+	"log"
+	"os"
+	"strings"
+	"sync"
+	"unicode/utf8"
+	"unsafe"
+
+	"github.com/jmorganca/ollama/api"
+)
+
+//go:embed ggml-metal.metal
+var fs embed.FS
+
+const ModelFamilyLlama ModelFamily = "llama"
+
+type llamaModel struct {
+	hyperparameters llamaHyperparameters
+}
+
+func (llm *llamaModel) ModelFamily() ModelFamily {
+	return ModelFamilyLlama
+}
+
+func (llm *llamaModel) ModelType() ModelType {
+	return ModelType30B
+}
+
+func (llm *llamaModel) FileType() FileType {
+	return llm.hyperparameters.FileType
+}
+
+type llamaHyperparameters struct {
+	// NumVocab is the size of the model's vocabulary.
+	NumVocab uint32
+
+	// NumEmbd is the size of the model's embedding layer.
+	NumEmbd uint32
+	NumMult uint32
+	NumHead uint32
+
+	// NumLayer is the number of layers in the model.
+	NumLayer uint32
+	NumRot   uint32
+
+	// FileType describes the quantization level of the model, e.g. Q4_0, Q5_K, etc.
+	FileType llamaFileType
+}
+
+type llamaFileType uint32
+
+const (
+	llamaFileTypeF32 llamaFileType = iota
+	llamaFileTypeF16
+	llamaFileTypeQ4_0
+	llamaFileTypeQ4_1
+	llamaFileTypeQ4_1_F16
+	llamaFileTypeQ8_0 llamaFileType = iota + 2
+	llamaFileTypeQ5_0
+	llamaFileTypeQ5_1
+	llamaFileTypeQ2_K
+	llamaFileTypeQ3_K_S
+	llamaFileTypeQ3_K_M
+	llamaFileTypeQ3_K_L
+	llamaFileTypeQ4_K_S
+	llamaFileTypeQ4_K_M
+	llamaFileTypeQ5_K_S
+	llamaFileTypeQ5_K_M
+	llamaFileTypeQ6_K
+)
+
+func (ft llamaFileType) String() string {
+	switch ft {
+	case llamaFileTypeF32:
+		return "F32"
+	case llamaFileTypeF16:
+		return "F16"
+	case llamaFileTypeQ4_0:
+		return "Q4_0"
+	case llamaFileTypeQ4_1:
+		return "Q4_1"
+	case llamaFileTypeQ4_1_F16:
+		return "Q4_1_F16"
+	case llamaFileTypeQ8_0:
+		return "Q8_0"
+	case llamaFileTypeQ5_0:
+		return "Q5_0"
+	case llamaFileTypeQ5_1:
+		return "Q5_1"
+	case llamaFileTypeQ2_K:
+		return "Q2_K"
+	case llamaFileTypeQ3_K_S:
+		return "Q3_K_S"
+	case llamaFileTypeQ3_K_M:
+		return "Q3_K_M"
+	case llamaFileTypeQ3_K_L:
+		return "Q3_K_L"
+	case llamaFileTypeQ4_K_S:
+		return "Q4_K_S"
+	case llamaFileTypeQ4_K_M:
+		return "Q4_K_M"
+	case llamaFileTypeQ5_K_S:
+		return "Q5_K_S"
+	case llamaFileTypeQ5_K_M:
+		return "Q5_K_M"
+	case llamaFileTypeQ6_K:
+		return "Q6_K"
+	default:
+		return "Unknown"
+	}
+}
+
+type llama struct {
+	params *C.struct_llama_context_params
+	model  *C.struct_llama_model
+	ctx    *C.struct_llama_context
+
+	last   []C.llama_token
+	embd   []C.llama_token
+	cursor int
+
+	mu sync.Mutex
+	gc bool
+
+	api.Options
+}
+
+func newLlama(model string, adapters []string, opts api.Options) (*llama, error) {
+	if _, err := os.Stat(model); err != nil {
+		return nil, err
+	}
+
+	llm := llama{Options: opts}
+
+	C.llama_backend_init(C.bool(llm.UseNUMA))
+
+	params := C.llama_context_default_params()
+	params.seed = C.uint(llm.Seed)
+	params.n_ctx = C.int(llm.NumCtx)
+	params.n_batch = C.int(llm.NumBatch)
+	params.n_gqa = C.int(llm.NumGQA)
+	params.n_gpu_layers = C.int(llm.NumGPU)
+	params.main_gpu = C.int(llm.MainGPU)
+	params.low_vram = C.bool(llm.LowVRAM)
+	params.f16_kv = C.bool(llm.F16KV)
+	params.logits_all = C.bool(llm.LogitsAll)
+	params.vocab_only = C.bool(llm.VocabOnly)
+	params.use_mmap = C.bool(llm.UseMMap)
+	params.use_mlock = C.bool(llm.UseMLock)
+	params.embedding = C.bool(llm.EmbeddingOnly)
+	params.rope_freq_base = C.float(llm.RopeFrequencyBase)
+	params.rope_freq_scale = C.float(llm.RopeFrequencyScale)
+
+	if len(adapters) > 0 && llm.UseMMap {
+		log.Printf("must disable mmap to use lora adapters")
+		params.use_mmap = C.bool(false)
+	}
+
+	llm.params = &params
+
+	cModel := C.CString(model)
+	defer C.free(unsafe.Pointer(cModel))
+
+	llm.model = C.llama_load_model_from_file(cModel, params)
+	if llm.model == nil {
+		return nil, errors.New("failed to load model")
+	}
+
+	llm.ctx = C.llama_new_context_with_model(llm.model, params)
+	if llm.ctx == nil {
+		return nil, errors.New("failed to create context")
+	}
+
+	for _, adapter := range adapters {
+		cAdapter := C.CString(adapter)
+		defer C.free(unsafe.Pointer(cAdapter))
+
+		if retval := C.llama_model_apply_lora_from_file(llm.model, cAdapter, nil, C.int(llm.NumThread)); retval != 0 {
+			return nil, fmt.Errorf("failed to load adapter %s", adapter)
+		}
+	}
+
+	// warm up the model
+	bos := []C.llama_token{C.llama_token_bos()}
+	C.llama_eval(llm.ctx, unsafe.SliceData(bos), C.int(len(bos)), 0, C.int(opts.NumThread))
+	C.llama_reset_timings(llm.ctx)
+
+	return &llm, nil
+}
+
+func (llm *llama) Close() {
+	llm.gc = true
+
+	llm.mu.Lock()
+	defer llm.mu.Unlock()
+
+	defer C.llama_free_model(llm.model)
+	defer C.llama_free(llm.ctx)
+
+	C.llama_print_timings(llm.ctx)
+}
+
+func (llm *llama) SetOptions(opts api.Options) {
+	llm.Options = opts
+}
+
+var errNeedMoreData = errors.New("need more data")
+
+func (llm *llama) Predict(ctx []int, prompt string, fn func(api.GenerateResponse)) error {
+	C.llama_reset_timings(llm.ctx)
+
+	llm.marshalPrompt(ctx, prompt)
+
+	C.llama_set_rng_seed(llm.ctx, C.uint(llm.Seed))
+
+	var b bytes.Buffer
+	for {
+		token, err := llm.next()
+		if llm.gc {
+			return nil
+		} else if errors.Is(err, io.EOF) {
+			break
+		} else if err != nil {
+			return err
+		}
+
+		b.WriteString(llm.Decode(int(token)))
+
+		if err := llm.checkStopConditions(b); err != nil {
+			if errors.Is(err, io.EOF) {
+				break
+			} else if errors.Is(err, errNeedMoreData) {
+				continue
+			}
+
+			return err
+		}
+
+		if utf8.Valid(b.Bytes()) || b.Len() >= utf8.UTFMax {
+			fn(api.GenerateResponse{Response: b.String()})
+			b.Reset()
+		}
+	}
+
+	embd := make([]int, len(llm.embd))
+	for i := range llm.embd {
+		embd[i] = int(llm.embd[i])
+	}
+
+	timings := C.llama_get_timings(llm.ctx)
+	fn(api.GenerateResponse{
+		Done:               true,
+		Context:            embd,
+		SampleCount:        int(timings.n_sample),
+		SampleDuration:     parseDurationMs(float64(timings.t_sample_ms)),
+		PromptEvalCount:    int(timings.n_p_eval),
+		PromptEvalDuration: parseDurationMs(float64(timings.t_p_eval_ms)),
+		EvalCount:          int(timings.n_eval),
+		EvalDuration:       parseDurationMs(float64(timings.t_eval_ms)),
+	})
+
+	return nil
+}
+
+func (llm *llama) checkStopConditions(b bytes.Buffer) error {
+	for _, stopCondition := range llm.Stop {
+		if stopCondition == strings.TrimSpace(b.String()) {
+			return io.EOF
+		} else if strings.HasPrefix(stopCondition, strings.TrimSpace(b.String())) {
+			return errNeedMoreData
+		}
+	}
+
+	return nil
+}
+
+func (llm *llama) marshalPrompt(ctx []int, prompt string) []C.llama_token {
+	tokens := append(ctx, llm.Encode(prompt)...)
+	if llm.NumKeep < 0 {
+		llm.NumKeep = len(tokens)
+	}
+
+	cTokens := make([]C.llama_token, len(tokens))
+	for i := range tokens {
+		cTokens[i] = C.llama_token(tokens[i])
+	}
+
+	// min(llm.NumCtx - 4, llm.NumKeep)
+	if llm.NumCtx-4 < llm.NumKeep {
+		llm.NumKeep = llm.NumCtx - 4
+	}
+
+	if len(tokens) >= llm.NumCtx {
+		// truncate input
+		numLeft := (llm.NumCtx - llm.NumKeep) / 2
+		truncated := cTokens[:llm.NumKeep]
+		erasedBlocks := (len(cTokens) - llm.NumKeep - numLeft - 1) / numLeft
+		truncated = append(truncated, cTokens[llm.NumKeep+erasedBlocks*numLeft:]...)
+		copy(llm.last, cTokens[len(cTokens)-llm.NumCtx:])
+
+		cTokens = truncated
+		log.Printf("input truncated: num_ctx=%d num_keep=%d num_left=%d num_tokens=%d", llm.NumCtx, llm.NumKeep, numLeft, len(truncated))
+	} else {
+		llm.last = make([]C.llama_token, llm.NumCtx-len(cTokens))
+		llm.last = append(llm.last, cTokens...)
+	}
+
+	var i int
+	for i = 0; i < len(llm.embd) && i < len(cTokens) && llm.embd[i] == cTokens[i]; i++ {
+		// noop
+	}
+
+	llm.embd = cTokens
+	if i == len(cTokens) {
+		// evaluate at least one token to generate logits
+		i--
+	}
+
+	llm.cursor = i
+
+	log.Printf("prompt: num_past=%d cached=%v eval=%v", i, len(llm.embd[:i]), len(llm.embd[i:]))
+	return cTokens
+}
+
+func (llm *llama) Encode(prompt string) []int {
+	cPrompt := C.CString(prompt)
+	defer C.free(unsafe.Pointer(cPrompt))
+
+	cTokens := make([]C.llama_token, len(prompt)+1)
+	if n := C.llama_tokenize(llm.ctx, cPrompt, unsafe.SliceData(cTokens), C.int(len(cTokens)), true); n > 0 {
+		tokens := make([]int, n)
+		for i := range cTokens[:n] {
+			tokens[i] = int(cTokens[i])
+		}
+
+		return tokens
+	}
+
+	return nil
+}
+
+func (llm *llama) Decode(tokens ...int) string {
+	var sb strings.Builder
+	for _, token := range tokens {
+		sb.WriteString(C.GoString(C.llama_token_to_str(llm.ctx, C.llama_token(token))))
+	}
+
+	return sb.String()
+}
+
+func (llm *llama) next() (C.llama_token, error) {
+	llm.mu.Lock()
+	defer llm.mu.Unlock()
+
+	if len(llm.embd) >= llm.NumCtx {
+		numLeft := (llm.NumCtx - llm.NumKeep) / 2
+		truncated := llm.embd[:llm.NumKeep]
+		truncated = append(truncated, llm.embd[len(llm.embd)-numLeft:]...)
+
+		llm.embd = truncated
+		llm.cursor = llm.NumKeep
+		log.Printf("input truncated: num_ctx=%d num_keep=%d num_left=%d num_tokens=%d cursor=%d", llm.NumCtx, llm.NumKeep, numLeft, len(truncated), llm.cursor)
+	}
+
+	for {
+		if llm.gc {
+			return 0, io.EOF
+		}
+
+		if llm.cursor >= len(llm.embd) {
+			break
+		}
+
+		numEval := len(llm.embd) - llm.cursor
+		if numEval > llm.NumBatch {
+			numEval = llm.NumBatch
+		}
+
+		if retval := C.llama_eval(llm.ctx, unsafe.SliceData(llm.embd[llm.cursor:]), C.int(numEval), C.int(llm.cursor), C.int(llm.NumThread)); retval != 0 {
+			return 0, fmt.Errorf("llama_eval: %d", retval)
+		}
+
+		llm.cursor += numEval
+	}
+
+	var sampleOpts C.struct_llama_sample_options
+	sampleOpts.repeat_penalty = C.float(llm.RepeatPenalty)
+	sampleOpts.frequency_penalty = C.float(llm.FrequencyPenalty)
+	sampleOpts.presence_penalty = C.float(llm.PresencePenalty)
+	sampleOpts.temperature = C.float(llm.Temperature)
+	sampleOpts.top_k = C.int(llm.TopK)
+	sampleOpts.top_p = C.float(llm.TopP)
+	sampleOpts.tfs_z = C.float(llm.TFSZ)
+	sampleOpts.typical_p = C.float(llm.TypicalP)
+	sampleOpts.mirostat = C.int(llm.Mirostat)
+	sampleOpts.mirostat_tau = C.float(llm.MirostatTau)
+	sampleOpts.mirostat_eta = C.float(llm.MirostatEta)
+	sampleOpts.penalize_newline = C.bool(llm.PenalizeNewline)
+
+	numVocab := C.llama_n_vocab(llm.ctx)
+	logits := unsafe.Slice(C.llama_get_logits(llm.ctx), numVocab)
+
+	// TODO: logit bias
+
+	candidates := make([]C.llama_token_data, numVocab)
+	for i := range logits {
+		candidates[i] = C.llama_token_data{
+			id:    C.int(i),
+			logit: logits[i],
+			p:     0,
+		}
+	}
+
+	repeatLastN := llm.RepeatLastN
+	if len(llm.last) < repeatLastN {
+		repeatLastN = len(llm.last)
+	}
+
+	if llm.NumCtx < repeatLastN {
+		repeatLastN = llm.NumCtx
+	}
+
+	lastN := llm.last[len(llm.last)-repeatLastN:]
+
+	token := C.llama_sample(
+		llm.ctx,
+		unsafe.SliceData(candidates), C.size_t(len(candidates)),
+		unsafe.SliceData(lastN), C.size_t(len(lastN)),
+		&sampleOpts,
+	)
+
+	llm.last = append(llm.last, token)
+	llm.embd = append(llm.embd, token)
+
+	if token == C.llama_token_eos() {
+		return 0, io.EOF
+	}
+
+	return token, nil
+}
+
+func (llm *llama) Embedding(input string) ([]float64, error) {
+	if !llm.EmbeddingOnly {
+		return nil, errors.New("llama: embedding not enabled")
+	}
+
+	tokens := llm.Encode(input)
+	if tokens == nil {
+		return nil, errors.New("llama: tokenize embedding")
+	}
+
+	cTokens := make([]C.llama_token, len(tokens))
+	for i := range tokens {
+		cTokens[i] = C.llama_token(tokens[i])
+	}
+
+	retval := C.llama_eval(llm.ctx, unsafe.SliceData(cTokens), C.int(len(tokens)), 0, C.int(llm.NumThread))
+	if retval != 0 {
+		return nil, errors.New("llama: eval")
+	}
+
+	C.llama_print_timings(llm.ctx)
+
+	n := C.llama_n_embd(llm.ctx)
+	if n <= 0 {
+		return nil, errors.New("llama: no embeddings generated")
+	}
+	cEmbeddings := unsafe.Slice(C.llama_get_embeddings(llm.ctx), n)
+
+	embeddings := make([]float64, len(cEmbeddings))
+	for i, v := range cEmbeddings {
+		embeddings[i] = float64(v)
+	}
+	return embeddings, nil
+}

llm/llama.h

@@ -1,5 +1,5 @@
 /**
- * llama.cpp - git 5bf2a2771886ee86137e01dbc7492f78fb392066
+ * llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
  *
  * MIT License
  *
@@ -79,6 +79,10 @@
 #define LLAMA_SUPPORTS_GPU_OFFLOAD
 #endif
 
+#ifndef LLAMA_DEFAULT_RMS_EPS
+#define LLAMA_DEFAULT_RMS_EPS 5e-6f
+#endif
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -108,13 +112,34 @@ extern "C" {
 
     typedef void (*llama_progress_callback)(float progress, void *ctx);
 
-   struct llama_context_params {
-        uint32_t seed;                         // RNG seed, -1 for random
-        int32_t  n_ctx;                        // text context
-        int32_t  n_batch;                      // prompt processing batch size
-        int32_t  n_gpu_layers;                 // number of layers to store in VRAM
-        int32_t  main_gpu;                     // the GPU that is used for scratch and small tensors
-        float tensor_split[LLAMA_MAX_DEVICES]; // how to split layers across multiple GPUs
+    enum llama_log_level {
+        LLAMA_LOG_LEVEL_ERROR = 2,
+        LLAMA_LOG_LEVEL_WARN  = 3,
+        LLAMA_LOG_LEVEL_INFO  = 4
+    };
+
+    // Signature for logging events
+    // Note that text includes the new line character at the end for most events.
+    // If your logging mechanism cannot handle that, check if the last character is '\n' and strip it
+    // if it exists.
+    // It might not exist for progress report where '.' is output repeatedly.
+    typedef void (*llama_log_callback)(enum llama_log_level level, const char * text, void * user_data);
+
+    struct llama_context_params {
+        uint32_t seed;         // RNG seed, -1 for random
+        int32_t  n_ctx;        // text context
+        int32_t  n_batch;      // prompt processing batch size
+        int32_t  n_gqa;        // grouped-query attention (TEMP - will be moved to model hparams)
+        float    rms_norm_eps; // rms norm epsilon (TEMP - will be moved to model hparams)
+        int32_t  n_gpu_layers; // number of layers to store in VRAM
+        int32_t  main_gpu;     // the GPU that is used for scratch and small tensors
+
+        const float * tensor_split; // how to split layers across multiple GPUs (size: LLAMA_MAX_DEVICES)
+
+        // ref: https://github.com/ggerganov/llama.cpp/pull/2054
+        float    rope_freq_base;  // RoPE base frequency
+        float    rope_freq_scale; // RoPE frequency scaling factor
+
         // called with a progress value between 0 and 1, pass NULL to disable
         llama_progress_callback progress_callback;
         // context pointer passed to the progress callback
@@ -122,6 +147,7 @@ extern "C" {
 
         // Keep the booleans together to avoid misalignment during copy-by-value.
         bool low_vram;   // if true, reduce VRAM usage at the cost of performance
+        bool mul_mat_q;  // if true, use experimental mul_mat_q kernels
         bool f16_kv;     // use fp16 for KV cache
         bool logits_all; // the llama_eval() call computes all logits, not just the last one
         bool vocab_only; // only load the vocabulary, no weights
@@ -160,6 +186,40 @@ extern "C" {
         bool quantize_output_tensor; // quantize output.weight
     } llama_model_quantize_params;
 
+    // grammar types
+    struct llama_grammar;
+
+    // grammar element type
+    enum llama_gretype {
+        // end of rule definition
+        LLAMA_GRETYPE_END            = 0,
+
+        // start of alternate definition for rule
+        LLAMA_GRETYPE_ALT            = 1,
+
+        // non-terminal element: reference to rule
+        LLAMA_GRETYPE_RULE_REF       = 2,
+
+        // terminal element: character (code point)
+        LLAMA_GRETYPE_CHAR           = 3,
+
+        // inverse char(s) ([^a], [^a-b] [^abc])
+        LLAMA_GRETYPE_CHAR_NOT       = 4,
+
+        // modifies a preceding LLAMA_GRETYPE_CHAR or LLAMA_GRETYPE_CHAR_ALT to
+        // be an inclusive range ([a-z])
+        LLAMA_GRETYPE_CHAR_RNG_UPPER = 5,
+
+        // modifies a preceding LLAMA_GRETYPE_CHAR or
+        // LLAMA_GRETYPE_CHAR_RNG_UPPER to add an alternate char to match ([ab], [a-zA])
+        LLAMA_GRETYPE_CHAR_ALT       = 6,
+    };
+
+    typedef struct llama_grammar_element {
+        enum llama_gretype type;
+        uint32_t           value; // Unicode code point or rule ID
+    } llama_grammar_element;
+
     // performance timing information
     struct llama_timings {
         double t_start_ms;
@@ -174,6 +234,12 @@ extern "C" {
         int32_t n_eval;
     };
 
+    // Set callback for all future logging events.
+    // If this is not called, or NULL is supplied, everything is output on stderr.
+    LLAMA_API void llama_log_set(llama_log_callback log_callback, void * user_data);
+
+    LLAMA_API int llama_max_devices();
+
     LLAMA_API struct llama_context_params llama_context_default_params();
     LLAMA_API struct llama_model_quantize_params llama_model_quantize_default_params();
 
@@ -296,10 +362,21 @@ extern "C" {
                              int   n_max_tokens,
                             bool   add_bos);
 
+    LLAMA_API int llama_tokenize_with_model(
+        const struct llama_model * model,
+                      const char * text,
+                     llama_token * tokens,
+                             int   n_max_tokens,
+                            bool   add_bos);
+
     LLAMA_API int llama_n_vocab(const struct llama_context * ctx);
     LLAMA_API int llama_n_ctx  (const struct llama_context * ctx);
     LLAMA_API int llama_n_embd (const struct llama_context * ctx);
 
+    LLAMA_API int llama_n_vocab_from_model(const struct llama_model * model);
+    LLAMA_API int llama_n_ctx_from_model  (const struct llama_model * model);
+    LLAMA_API int llama_n_embd_from_model (const struct llama_model * model);
+
     // Get the vocabulary as output parameters.
     // Returns number of results.
     LLAMA_API int llama_get_vocab(
@@ -308,6 +385,12 @@ extern "C" {
                                  float * scores,
                                    int   capacity);
 
+    LLAMA_API int llama_get_vocab_from_model(
+              const struct llama_model * model,
+                          const char * * strings,
+                                 float * scores,
+                                   int   capacity);
+
     // Token logits obtained from the last call to llama_eval()
     // The logits for the last token are stored in the last row
     // Can be mutated in order to change the probabilities of the next token
@@ -320,13 +403,28 @@ extern "C" {
     LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
 
     // Token Id -> String. Uses the vocabulary in the provided context
-    LLAMA_API const char * llama_token_to_str(const struct llama_context * ctx, llama_token token);
+    LLAMA_API const char * llama_token_to_str(
+            const struct llama_context * ctx,
+                           llama_token   token);
+
+    LLAMA_API const char * llama_token_to_str_with_model(
+              const struct llama_model * model,
+                           llama_token   token);
 
     // Special tokens
     LLAMA_API llama_token llama_token_bos();  // beginning-of-sentence
     LLAMA_API llama_token llama_token_eos();  // end-of-sentence
     LLAMA_API llama_token llama_token_nl();   // next-line
 
+    // Grammar
+    //
+    LLAMA_API struct llama_grammar * llama_grammar_init(
+            const llama_grammar_element ** rules,
+                                 size_t    n_rules,
+                                 size_t    start_rule_index);
+
+    LLAMA_API void llama_grammar_free(struct llama_grammar * grammar);
+
     // Sampling functions
 
     /// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix.
@@ -339,13 +437,11 @@ extern "C" {
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, the logits must be directly extracted from the original generation context without being sorted.
     /// @params guidance_ctx A separate context from the same model. Other than a negative prompt at the beginning, it should have all generated and user input tokens copied from the main context.
     /// @params scale Guidance strength. 1.0f means no guidance. Higher values mean stronger guidance.
-    /// @params smooth_factor Smooth factor between guidance logits and original logits. 1.0f means only use guidance logits. 0.0f means only original logits.
     LLAMA_API void llama_sample_classifier_free_guidance(
               struct llama_context * ctx,
             llama_token_data_array * candidates,
               struct llama_context * guidance_ctx,
-                             float   scale,
-                             float   smooth_factor);
+                             float   scale);
 
     /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
     LLAMA_API void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * candidates);
@@ -363,6 +459,9 @@ extern "C" {
     LLAMA_API void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
     LLAMA_API void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array * candidates, float temp);
 
+    /// @details Apply constraints from grammar
+    LLAMA_API void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * candidates, const struct llama_grammar * grammar);
+
     /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
     /// @param tau  The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
@@ -384,6 +483,9 @@ extern "C" {
     /// @details Randomly selects a token from the candidates based on their probabilities.
     LLAMA_API llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_array * candidates);
 
+    /// @details Accepts the sampled token into the grammar
+    LLAMA_API void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar * grammar, llama_token token);
+
     // Performance information
     LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx);
     LLAMA_API void llama_print_timings(struct llama_context * ctx);

llm/llama_darwin.go

@@ -0,0 +1,81 @@
+package llm
+
+import (
+	"bytes"
+	"crypto/sha256"
+	"errors"
+	"io"
+	"log"
+	"os"
+	"path/filepath"
+)
+
+func init() {
+	if err := initBackend(); err != nil {
+		log.Printf("WARNING: GPU could not be initialized correctly: %v", err)
+		log.Printf("WARNING: falling back to CPU")
+	}
+}
+
+func initBackend() error {
+	exec, err := os.Executable()
+	if err != nil {
+		return err
+	}
+
+	exec, err = filepath.EvalSymlinks(exec)
+	if err != nil {
+		return err
+	}
+
+	metal := filepath.Join(filepath.Dir(exec), "ggml-metal.metal")
+	fi, err := os.Stat(metal)
+	if err != nil && !errors.Is(err, os.ErrNotExist) {
+		return err
+	}
+
+	if fi != nil {
+		actual, err := os.Open(metal)
+		if err != nil {
+			return err
+		}
+		defer actual.Close()
+
+		actualSum := sha256.New()
+		if _, err := io.Copy(actualSum, actual); err != nil {
+			return err
+		}
+
+		expect, err := fs.Open("ggml-metal.metal")
+		if err != nil {
+			return err
+		}
+
+		expectSum := sha256.New()
+		if _, err := io.Copy(expectSum, expect); err != nil {
+			return err
+		}
+
+		if bytes.Equal(actualSum.Sum(nil), expectSum.Sum(nil)) {
+			return nil
+		}
+	}
+
+	dst, err := os.Create(filepath.Join(filepath.Dir(exec), "ggml-metal.metal"))
+	if err != nil {
+		return err
+	}
+	defer dst.Close()
+
+	src, err := fs.Open("ggml-metal.metal")
+	if err != nil {
+		return err
+	}
+	defer src.Close()
+
+	if _, err := io.Copy(dst, src); err != nil {
+		return err
+	}
+
+	return nil
+}

llm/llm.go

@@ -0,0 +1,74 @@
+package llm
+
+import (
+	"fmt"
+	"log"
+	"os"
+
+	"github.com/pbnjay/memory"
+
+	"github.com/jmorganca/ollama/api"
+)
+
+type LLM interface {
+	Predict([]int, string, func(api.GenerateResponse)) error
+	Embedding(string) ([]float64, error)
+	Encode(string) []int
+	Decode(...int) string
+	SetOptions(api.Options)
+	Close()
+}
+
+func New(model string, adapters []string, opts api.Options) (LLM, error) {
+	if _, err := os.Stat(model); err != nil {
+		return nil, err
+	}
+
+	f, err := os.Open(model)
+	if err != nil {
+		return nil, err
+	}
+	defer f.Close()
+
+	ggml, err := DecodeGGML(f, ModelFamilyLlama)
+	if err != nil {
+		return nil, err
+	}
+
+	switch ggml.FileType().String() {
+	case "F32", "F16", "Q5_0", "Q5_1", "Q8_0":
+		if opts.NumGPU != 0 {
+			// F32, F16, Q5_0, Q5_1, and Q8_0 do not support Metal API and will
+			// cause the runner to segmentation fault so disable GPU
+			log.Printf("WARNING: GPU disabled for F32, F16, Q5_0, Q5_1, and Q8_0")
+			opts.NumGPU = 0
+		}
+	}
+
+	totalResidentMemory := memory.TotalMemory()
+	switch ggml.ModelType() {
+	case ModelType3B, ModelType7B:
+		if totalResidentMemory < 8*1024*1024 {
+			return nil, fmt.Errorf("model requires at least 8GB of memory")
+		}
+	case ModelType13B:
+		if totalResidentMemory < 16*1024*1024 {
+			return nil, fmt.Errorf("model requires at least 16GB of memory")
+		}
+	case ModelType30B:
+		if totalResidentMemory < 32*1024*1024 {
+			return nil, fmt.Errorf("model requires at least 32GB of memory")
+		}
+	case ModelType65B:
+		if totalResidentMemory < 64*1024*1024 {
+			return nil, fmt.Errorf("model requires at least 64GB of memory")
+		}
+	}
+
+	switch ggml.ModelFamily() {
+	case ModelFamilyLlama:
+		return newLlama(model, adapters, opts)
+	default:
+		return nil, fmt.Errorf("unknown ggml type: %s", ggml.ModelFamily())
+	}
+}

llm/update-llama-cpp.sh

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

llm/utils.go

@@ -0,0 +1,15 @@
+package llm
+
+import (
+	"fmt"
+	"time"
+)
+
+func parseDurationMs(ms float64) time.Duration {
+	dur, err := time.ParseDuration(fmt.Sprintf("%fms", ms))
+	if err != nil {
+		panic(err)
+	}
+
+	return dur
+}

main.go

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

models.json

@@ -1,38 +0,0 @@
-[
-  {
-    "name": "orca",
-    "display_name": "Orca Mini",
-    "parameters": "3B",
-    "url": "https://huggingface.co/TheBloke/orca_mini_3B-GGML/resolve/main/orca-mini-3b.ggmlv3.q4_1.bin",
-    "short_description": "Follow instructions. Great small model that runs fast even without GPU support.",
-    "description": "An OpenLLaMa-3B model trained on explain tuned datasets, created using Instructions and Input from WizardLM, Alpaca & Dolly-V2 datasets and applying Orca Research Paper dataset construction approaches.",
-    "published_by": "TheBloke",
-    "original_author": "psmathur",
-    "original_url": "https://huggingface.co/psmathur/orca_mini_3b",
-    "license": "CC-BY-SA-4.0"
-  },
-  {
-    "name": "nous-hermes",
-    "display_name": "Nous Hermes",
-    "parameters": "13B",
-    "url": "https://huggingface.co/TheBloke/Nous-Hermes-13B-GGML/resolve/main/nous-hermes-13b.ggmlv3.q2_K.bin",
-    "short_description": "Currently one of the best 13B general model.",
-    "description": "It is suitable for a wide range of language tasks, from generating creative text to understanding and following complex instructions. This model was fine-tuned by Nous Research, with Teknium and Karan4D leading the fine tuning process and dataset curation, Redmond AI sponsoring the compute, and several other contributors. The result is an enhanced Llama 13b model that rivals GPT-3.5-turbo in performance across a variety of tasks. \n \n This model stands out for its long responses, low hallucination rate, and absence of OpenAI censorship mechanisms. The fine-tuning process was performed with a 2000 sequence length on an 8x a100 80GB DGX machine for over 50 hours.",
-    "published_by": "TheBloke",
-    "original_author": "NousResearch",
-    "original_url": "https://huggingface.co/NousResearch/Nous-Hermes-13b",
-    "license": "GPL"
-  },
-  {
-    "name": "vicuna",
-    "display_name": "Vicuna",
-    "parameters": "7B",
-    "url": "https://huggingface.co/TheBloke/vicuna-7B-v1.3-GGML/resolve/main/vicuna-7b-v1.3.ggmlv3.q4_0.bin",
-    "short_description": "Vicuna is a chat assistant trained by fine-tuning LLaMA on user-shared conversations collected from ShareGPT.",
-    "description": "The primary use of Vicuna is research on large language models and chatbots. The primary intended users of the model are researchers and hobbyists in natural language processing, machine learning, and artificial intelligence.",
-    "published_by": "TheBloke",
-    "original_author": "LMSYS",
-    "original_url": "https://huggingface.co/lmsys/vicuna-7b-v1.3",
-    "license:": "Non-commercial"
-  }
-]

parser/parser.go

@@ -2,76 +2,115 @@ package parser
 
 import (
 	"bufio"
+	"bytes"
+	"errors"
 	"fmt"
 	"io"
-	"strings"
+	"log"
 )
 
 type Command struct {
 	Name string
-	Arg  string
+	Args string
+}
+
+func (c *Command) Reset() {
+	c.Name = ""
+	c.Args = ""
 }
 
 func Parse(reader io.Reader) ([]Command, error) {
 	var commands []Command
-	var foundModel bool
+	var command, modelCommand Command
 
 	scanner := bufio.NewScanner(reader)
-	multiline := false
-	var multilineCommand *Command
+	scanner.Buffer(make([]byte, 0, bufio.MaxScanTokenSize), bufio.MaxScanTokenSize)
+	scanner.Split(scanModelfile)
 	for scanner.Scan() {
-		line := scanner.Text()
-		if multiline {
-			// If we're in a multiline string and the line is """, end the multiline string.
-			if strings.TrimSpace(line) == `"""` {
-				multiline = false
-				commands = append(commands, *multilineCommand)
-			} else {
-				// Otherwise, append the line to the multiline string.
-				multilineCommand.Arg += "\n" + line
-			}
-			continue
-		}
-		fields := strings.Fields(line)
-		if len(fields) == 0 {
+		line := scanner.Bytes()
+
+		fields := bytes.SplitN(line, []byte(" "), 2)
+		if len(fields) == 0 || len(fields[0]) == 0 {
 			continue
 		}
 
-		command := Command{}
-		switch strings.ToUpper(fields[0]) {
+		switch string(bytes.ToUpper(fields[0])) {
 		case "FROM":
 			command.Name = "model"
-			command.Arg = fields[1]
-			if command.Arg == "" {
-				return nil, fmt.Errorf("no model specified in FROM line")
-			}
-			foundModel = true
-		case "PROMPT":
-			command.Name = "prompt"
-			if fields[1] == `"""` {
-				multiline = true
-				multilineCommand = &command
-				multilineCommand.Arg = ""
-			} else {
-				command.Arg = strings.Join(fields[1:], " ")
-			}
+			command.Args = string(fields[1])
+			// copy command for validation
+			modelCommand = command
+		case "LICENSE", "TEMPLATE", "SYSTEM", "PROMPT", "EMBED", "ADAPTER":
+			command.Name = string(bytes.ToLower(fields[0]))
+			command.Args = string(fields[1])
 		case "PARAMETER":
-			command.Name = fields[1]
-			command.Arg = strings.Join(fields[2:], " ")
+			fields = bytes.SplitN(fields[1], []byte(" "), 2)
+			if len(fields) < 2 {
+				return nil, fmt.Errorf("missing value for %s", fields)
+			}
+
+			command.Name = string(fields[0])
+			command.Args = string(fields[1])
 		default:
+			if !bytes.HasPrefix(fields[0], []byte("#")) {
+				// log a warning for unknown commands
+				log.Printf("WARNING: Unknown command: %s", fields[0])
+			}
 			continue
 		}
-		if !multiline {
-			commands = append(commands, command)
-		}
+
+		commands = append(commands, command)
+		command.Reset()
 	}
 
-	if !foundModel {
-		return nil, fmt.Errorf("no FROM line for the model was specified")
+	if modelCommand.Args == "" {
+		return nil, errors.New("no FROM line for the model was specified")
 	}
 
-	if multiline {
-		return nil, fmt.Errorf("unclosed multiline string")
-	}
 	return commands, scanner.Err()
 }
+
+func scanModelfile(data []byte, atEOF bool) (advance int, token []byte, err error) {
+	advance, token, err = scan([]byte(`"""`), []byte(`"""`), data, atEOF)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	if advance > 0 && token != nil {
+		return advance, token, nil
+	}
+
+	advance, token, err = scan([]byte(`"`), []byte(`"`), data, atEOF)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	if advance > 0 && token != nil {
+		return advance, token, nil
+	}
+
+	return bufio.ScanLines(data, atEOF)
+}
+
+func scan(openBytes, closeBytes, data []byte, atEOF bool) (advance int, token []byte, err error) {
+	newline := bytes.IndexByte(data, '\n')
+
+	if start := bytes.Index(data, openBytes); start >= 0 && start < newline {
+		end := bytes.Index(data[start+len(openBytes):], closeBytes)
+		if end < 0 {
+			if atEOF {
+				return 0, nil, fmt.Errorf("unterminated %s: expecting %s", openBytes, closeBytes)
+			} else {
+				return 0, nil, nil
+			}
+		}
+
+		n := start + len(openBytes) + end + len(closeBytes)
+
+		newData := data[:start]
+		newData = append(newData, data[start+len(openBytes):n-len(closeBytes)]...)
+		return n, newData, nil
+	}
+
+	return 0, nil, nil
+}