introduce build.go for controlling distribution builds

This commit aims to provide the Ollama maintainers with maximum control
of the distribution build process by creating a cross-platform shim.

Currently, we have no flexibility, or control of the process (pre and
post) or even the quality of the build.

By introducing a shim, and propagating it out to Homebrew, et al., we
can soon after ensure that the build process is consistent, and
reliable.

This also happens to remove the requirement for go generate and the
build tag hacks, but it does still support go generate in the flow, at
least until we can remove it after the major distribution use the new
build process.

About the script:

Beyond giving the Ollama maintainers drastically more control over the
build process, the script also provides a few other benefits:

- It is cross-platform, and can be run on any platform that supports Go
  (a hard requirement for building Ollama anyway).

- It can can check for correct versions of cmake, and other dependencies
  before starting the build process, and provide helpful error messages
  to the user if they are not met.

- It can be used to build the distribution for any platform,
  architecture, or build type (debug, release, etc.) with a single
  command. Currently, it is two commands.

- It can skip parts of the build process if they are already done, such
  as build the C dependencies. Of course there is a -f flag to force
  rebuild.

- So much more!

.github/workflows/test.yaml

@@ -73,12 +73,12 @@ jobs:
           $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
           $env:PATH="$gopath;$gccpath;$env:PATH"
           echo $env:PATH
-          go generate -x ./...
+          $env:GOARCH=""; $env:OLLAMA_BUILD_TARGET_ARCH="${{ matrix.arch }}"; go generate -x ./...
         if: ${{ startsWith(matrix.os, 'windows-') }}
-        name: 'Windows Go Generate'
-      - run: go generate -x ./...
+        name: 'Windows Generate'
+      - run: GOARCH= OLLAMA_BUILD_TARGET_ARCH=${{ matrix.arch }} go generate -x ./...
         if: ${{ ! startsWith(matrix.os, 'windows-') }}
-        name: 'Unix Go Generate'
+        name: 'Unix Generate'
       - uses: actions/upload-artifact@v4
         with:
           name: ${{ matrix.os }}-${{ matrix.arch }}-libraries
@@ -184,7 +184,7 @@ jobs:
           $env:OLLAMA_SKIP_CPU_GENERATE="1"
           $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
           go generate -x ./...
-        name: go generate
+        name: go generate -x ./...
         env:
           OLLAMA_SKIP_CPU_GENERATE: '1'
       # TODO - do we need any artifacts?
@@ -217,7 +217,7 @@ jobs:
       - name: 'Verify CUDA'
         run: nvcc -V
       - run: go get ./...
-      - name: go generate
+      - name: go generate -x ./...
         run: |
           $gopath=(get-command go).source | split-path -parent
           $cudabin=(get-command nvcc).source | split-path
@@ -312,7 +312,10 @@ jobs:
           touch llm/build/darwin/$ARCH/stub/bin/ollama_llama_server
         if: ${{ startsWith(matrix.os, 'macos-') }}
         shell: bash
-      - run: go generate ./...
+      - run: $env:GOARCH=""; $env:OLLAMA_BUILD_TARGET_ARCH="${{ matrix.arch }}"; go generate -x ./...
+        if: ${{ startsWith(matrix.os, 'windows-') }}
+      - run: GOARCH= OLLAMA_BUILD_TARGET_ARCH=${{ matrix.arch }} go generate -x ./...
+        if: ${{ ! startsWith(matrix.os, 'windows-') }}
       - run: go build
       - run: go test -v ./...
       - uses: actions/upload-artifact@v4

README.md

@@ -1,12 +1,12 @@
 <div align="center">
- <img alt="ollama" height="200px" src="https://github.com/ollama/ollama/assets/3325447/0d0b44e2-8f4a-4e99-9b52-a5c1c741c8f7">
+  <img alt="ollama" height="200px" src="https://github.com/ollama/ollama/assets/3325447/0d0b44e2-8f4a-4e99-9b52-a5c1c741c8f7">
 </div>
 
 # Ollama
 
 [![Discord](https://dcbadge.vercel.app/api/server/ollama?style=flat&compact=true)](https://discord.gg/ollama)
 
-Get up and running with large language models.
+Get up and running with large language models locally.
 
 ### macOS
 
@@ -51,17 +51,15 @@ Here are some example models that can be downloaded:
 | ------------------ | ---------- | ----- | ------------------------------ |
 | Llama 3            | 8B         | 4.7GB | `ollama run llama3`            |
 | Llama 3            | 70B        | 40GB  | `ollama run llama3:70b`        |
-| Phi 3 Mini         | 3.8B       | 2.3GB | `ollama run phi3`              |
-| Phi 3 Medium       | 14B        | 7.9GB | `ollama run phi3:medium`       |
-| Gemma              | 2B         | 1.4GB | `ollama run gemma:2b`          |
-| Gemma              | 7B         | 4.8GB | `ollama run gemma:7b`          |
+| Phi-3              | 3,8B       | 2.3GB | `ollama run phi3`              |
 | Mistral            | 7B         | 4.1GB | `ollama run mistral`           |
-| Moondream 2        | 1.4B       | 829MB | `ollama run moondream`         |
 | Neural Chat        | 7B         | 4.1GB | `ollama run neural-chat`       |
 | Starling           | 7B         | 4.1GB | `ollama run starling-lm`       |
 | Code Llama         | 7B         | 3.8GB | `ollama run codellama`         |
 | Llama 2 Uncensored | 7B         | 3.8GB | `ollama run llama2-uncensored` |
 | LLaVA              | 7B         | 4.5GB | `ollama run llava`             |
+| Gemma              | 2B         | 1.4GB | `ollama run gemma:2b`          |
+| Gemma              | 7B         | 4.8GB | `ollama run gemma:7b`          |
 | Solar              | 10.7B      | 6.1GB | `ollama run solar`             |
 
 > Note: You should have at least 8 GB of RAM available to run the 7B models, 16 GB to run the 13B models, and 32 GB to run the 33B models.
@@ -175,19 +173,13 @@ I'm a basic program that prints the famous "Hello, world!" message to the consol
 The image features a yellow smiley face, which is likely the central focus of the picture.
 ```
 
-### Pass the prompt as an argument
+### Pass in prompt as arguments
 
 ```
 $ ollama run llama3 "Summarize this file: $(cat README.md)"
  Ollama is a lightweight, extensible framework for building and running language models on the local machine. It provides a simple API for creating, running, and managing models, as well as a library of pre-built models that can be easily used in a variety of applications.
 ```
 
-### Show model information
-
-```
-ollama show llama3
-```
-
 ### List models on your computer
 
 ```
@@ -200,7 +192,19 @@ ollama list
 
 ## Building
 
-See the [developer guide](https://github.com/ollama/ollama/blob/main/docs/development.md)
+Install `cmake` and `go`:
+
+```
+brew install cmake go
+```
+
+Then build the binary:
+
+```
+go run build.go
+```
+
+More detailed instructions can be found in the [developer guide](https://github.com/ollama/ollama/blob/main/docs/development.md)
 
 ### Running local builds
 
@@ -248,7 +252,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 
 - [Open WebUI](https://github.com/open-webui/open-webui)
 - [Enchanted (macOS native)](https://github.com/AugustDev/enchanted)
-- [Hollama](https://github.com/fmaclen/hollama)
 - [Lollms-Webui](https://github.com/ParisNeo/lollms-webui)
 - [LibreChat](https://github.com/danny-avila/LibreChat)
 - [Bionic GPT](https://github.com/bionic-gpt/bionic-gpt)
@@ -275,23 +278,17 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [OllamaGUI](https://github.com/enoch1118/ollamaGUI)
 - [OpenAOE](https://github.com/InternLM/OpenAOE)
 - [Odin Runes](https://github.com/leonid20000/OdinRunes)
-- [LLM-X](https://github.com/mrdjohnson/llm-x) (Progressive Web App)
+- [LLM-X: Progressive Web App](https://github.com/mrdjohnson/llm-x)
 - [AnythingLLM (Docker + MacOs/Windows/Linux native app)](https://github.com/Mintplex-Labs/anything-llm)
 - [Ollama Basic Chat: Uses HyperDiv Reactive UI](https://github.com/rapidarchitect/ollama_basic_chat)
 - [Ollama-chats RPG](https://github.com/drazdra/ollama-chats)
-- [QA-Pilot](https://github.com/reid41/QA-Pilot) (Chat with Code Repository)
-- [ChatOllama](https://github.com/sugarforever/chat-ollama) (Open Source Chatbot based on Ollama with Knowledge Bases)
-- [CRAG Ollama Chat](https://github.com/Nagi-ovo/CRAG-Ollama-Chat) (Simple Web Search with Corrective RAG)
-- [RAGFlow](https://github.com/infiniflow/ragflow) (Open-source Retrieval-Augmented Generation engine based on deep document understanding)
-- [StreamDeploy](https://github.com/StreamDeploy-DevRel/streamdeploy-llm-app-scaffold) (LLM Application Scaffold)
-- [chat](https://github.com/swuecho/chat) (chat web app for teams)
+- [QA-Pilot: Chat with Code Repository](https://github.com/reid41/QA-Pilot)
+- [ChatOllama: Open Source Chatbot based on Ollama with Knowledge Bases](https://github.com/sugarforever/chat-ollama)
+- [CRAG Ollama Chat: Simple Web Search with Corrective RAG](https://github.com/Nagi-ovo/CRAG-Ollama-Chat)
+- [RAGFlow: Open-source Retrieval-Augmented Generation engine based on deep document understanding](https://github.com/infiniflow/ragflow)
+- [chat: chat web app for teams](https://github.com/swuecho/chat)
 - [Lobe Chat](https://github.com/lobehub/lobe-chat) with [Integrating Doc](https://lobehub.com/docs/self-hosting/examples/ollama)
-- [Ollama RAG Chatbot](https://github.com/datvodinh/rag-chatbot.git) (Local Chat with multiple PDFs using Ollama and RAG)
-- [BrainSoup](https://www.nurgo-software.com/products/brainsoup) (Flexible native client with RAG & multi-agent automation)
-- [macai](https://github.com/Renset/macai) (macOS client for Ollama, ChatGPT, and other compatible API back-ends)
-- [Olpaka](https://github.com/Otacon/olpaka) (User-friendly Flutter Web App for Ollama)
-- [OllamaSpring](https://github.com/CrazyNeil/OllamaSpring) (Ollama Client for macOS)
-- [LLocal.in](https://github.com/kartikm7/llocal) (Easy to use Electron Desktop Client for Ollama)
+- [Ollama RAG Chatbot: Local Chat with multiples PDFs using Ollama and RAG.](https://github.com/datvodinh/rag-chatbot.git)
 
 ### Terminal
 
@@ -314,7 +311,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [ShellOracle](https://github.com/djcopley/ShellOracle)
 - [tlm](https://github.com/yusufcanb/tlm)
 - [podman-ollama](https://github.com/ericcurtin/podman-ollama)
-- [gollama](https://github.com/sammcj/gollama)
 
 ### Database
 
@@ -325,20 +321,17 @@ See the [API documentation](./docs/api.md) for all endpoints.
 
 - [Pacman](https://archlinux.org/packages/extra/x86_64/ollama/)
 - [Helm Chart](https://artifacthub.io/packages/helm/ollama-helm/ollama)
-- [Guix channel](https://codeberg.org/tusharhero/ollama-guix)
 
 ### Libraries
 
 - [LangChain](https://python.langchain.com/docs/integrations/llms/ollama) and [LangChain.js](https://js.langchain.com/docs/modules/model_io/models/llms/integrations/ollama) with [example](https://js.langchain.com/docs/use_cases/question_answering/local_retrieval_qa)
 - [LangChainGo](https://github.com/tmc/langchaingo/) with [example](https://github.com/tmc/langchaingo/tree/main/examples/ollama-completion-example)
 - [LangChain4j](https://github.com/langchain4j/langchain4j) with [example](https://github.com/langchain4j/langchain4j-examples/tree/main/ollama-examples/src/main/java)
-- [LangChainRust](https://github.com/Abraxas-365/langchain-rust) with [example](https://github.com/Abraxas-365/langchain-rust/blob/main/examples/llm_ollama.rs)
 - [LlamaIndex](https://gpt-index.readthedocs.io/en/stable/examples/llm/ollama.html)
 - [LiteLLM](https://github.com/BerriAI/litellm)
 - [OllamaSharp for .NET](https://github.com/awaescher/OllamaSharp)
 - [Ollama for Ruby](https://github.com/gbaptista/ollama-ai)
 - [Ollama-rs for Rust](https://github.com/pepperoni21/ollama-rs)
-- [Ollama-hpp for C++](https://github.com/jmont-dev/ollama-hpp)
 - [Ollama4j for Java](https://github.com/amithkoujalgi/ollama4j)
 - [ModelFusion Typescript Library](https://modelfusion.dev/integration/model-provider/ollama)
 - [OllamaKit for Swift](https://github.com/kevinhermawan/OllamaKit)
@@ -349,13 +342,9 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Haystack](https://github.com/deepset-ai/haystack-integrations/blob/main/integrations/ollama.md)
 - [Elixir LangChain](https://github.com/brainlid/langchain)
 - [Ollama for R - rollama](https://github.com/JBGruber/rollama)
-- [Ollama for R - ollama-r](https://github.com/hauselin/ollama-r)
 - [Ollama-ex for Elixir](https://github.com/lebrunel/ollama-ex)
 - [Ollama Connector for SAP ABAP](https://github.com/b-tocs/abap_btocs_ollama)
 - [Testcontainers](https://testcontainers.com/modules/ollama/)
-- [Portkey](https://portkey.ai/docs/welcome/integration-guides/ollama)
-- [PromptingTools.jl](https://github.com/svilupp/PromptingTools.jl) with an [example](https://svilupp.github.io/PromptingTools.jl/dev/examples/working_with_ollama)
-- [LlamaScript](https://github.com/Project-Llama/llamascript)
 
 ### Mobile
 
@@ -375,23 +364,18 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Ollama Telegram Bot](https://github.com/ruecat/ollama-telegram)
 - [Hass Ollama Conversation](https://github.com/ej52/hass-ollama-conversation)
 - [Rivet plugin](https://github.com/abrenneke/rivet-plugin-ollama)
+- [Llama Coder](https://github.com/ex3ndr/llama-coder) (Copilot alternative using Ollama)
 - [Obsidian BMO Chatbot plugin](https://github.com/longy2k/obsidian-bmo-chatbot)
 - [Cliobot](https://github.com/herval/cliobot) (Telegram bot with Ollama support)
 - [Copilot for Obsidian plugin](https://github.com/logancyang/obsidian-copilot)
 - [Obsidian Local GPT plugin](https://github.com/pfrankov/obsidian-local-gpt)
 - [Open Interpreter](https://docs.openinterpreter.com/language-model-setup/local-models/ollama)
-- [Llama Coder](https://github.com/ex3ndr/llama-coder) (Copilot alternative using Ollama)
-- [Ollama Copilot](https://github.com/bernardo-bruning/ollama-copilot) (Proxy that allows you to use ollama as a copilot like Github copilot)
 - [twinny](https://github.com/rjmacarthy/twinny) (Copilot and Copilot chat alternative using Ollama)
 - [Wingman-AI](https://github.com/RussellCanfield/wingman-ai) (Copilot code and chat alternative using Ollama and HuggingFace)
 - [Page Assist](https://github.com/n4ze3m/page-assist) (Chrome Extension)
 - [AI Telegram Bot](https://github.com/tusharhero/aitelegrambot) (Telegram bot using Ollama in backend)
 - [AI ST Completion](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (Sublime Text 4 AI assistant plugin with Ollama support)
 - [Discord-Ollama Chat Bot](https://github.com/kevinthedang/discord-ollama) (Generalized TypeScript Discord Bot w/ Tuning Documentation)
-- [Discord AI chat/moderation bot](https://github.com/rapmd73/Companion) Chat/moderation bot written in python. Uses Ollama to create personalities.
-- [Headless Ollama](https://github.com/nischalj10/headless-ollama) (Scripts to automatically install ollama client & models on any OS for apps that depends on ollama server)
-
-### Supported backends
-
-- [llama.cpp](https://github.com/ggerganov/llama.cpp) project founded by Georgi Gerganov.
 
+### Supported backends 
+- [llama.cpp](https://github.com/ggerganov/llama.cpp) project founded by Georgi Gerganov. 

build.go

@@ -0,0 +1,199 @@
+//go:build ignore
+
+package main
+
+import (
+	"cmp"
+	"errors"
+	"flag"
+	"log"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"runtime"
+)
+
+// Flags
+var (
+	flagRegenerateDestroy = flag.Bool("d", false, "force regenerate the dependencies (destructive)")
+	flagRegenerateGently  = flag.Bool("g", false, "regenerate the dependencies (non-destructive)")
+	flagSkipBuild         = flag.Bool("s", false, "generate dependencies only (e.g. skip 'go build .')")
+
+	// Flags to set GOARCH explicitly for cross-platform builds,
+	// e.g., in CI to target a different platform than the build matrix
+	// default. These allows us to run generate without a separate build
+	// step for building the script binary for the host ARCH and then
+	// runing the generate script for the target ARCH. Instead, we can
+	// just run `go run build.go -target=$GOARCH` to generate the
+	// deps.
+	flagGOARCH = flag.String("target", "", "sets GOARCH to use when generating dependencies and building")
+)
+
+func buildEnv() []string {
+	return append(os.Environ(), "GOARCH="+cmp.Or(
+		*flagGOARCH,
+		os.Getenv("OLLAMA_BUILD_TARGET_ARCH"),
+		runtime.GOARCH,
+	))
+}
+
+func main() {
+	log.SetFlags(0)
+	flag.Usage = func() {
+		log.Printf("Usage: go run build.go [flags]")
+		log.Println()
+		log.Println("Flags:")
+		flag.PrintDefaults()
+		log.Println()
+		log.Println("This script builds the Ollama server binary and generates the llama.cpp")
+		log.Println("bindings for the current platform. It assumes that the current working")
+		log.Println("directory is the root directory of the Ollama project.")
+		log.Println()
+		log.Println("If the -d flag is provided, the script will force regeneration of the")
+		log.Println("dependencies; removing the 'llm/build' directory before starting.")
+		log.Println()
+		log.Println("If the -g flag is provided, the script will regenerate the dependencies")
+		log.Println("without removing the 'llm/build' directory.")
+		log.Println()
+		log.Println("If the -s flag is provided, the script will skip building the Ollama binary")
+		log.Println()
+		log.Println("If the -target flag is provided, the script will set GOARCH to the value")
+		log.Println("of the flag. This is useful for cross-platform builds.")
+		log.Println()
+		log.Println("The script will check for the required dependencies (cmake, gcc) and")
+		log.Println("print their version.")
+		log.Println()
+		log.Println("The script will also check if it is being run from the root directory of")
+		log.Println("the Ollama project.")
+		log.Println()
+		os.Exit(1)
+	}
+	flag.Parse()
+
+	log.Printf("=== Building Ollama ===")
+	defer func() {
+		log.Printf("=== Done building Ollama ===")
+		if !*flagSkipBuild {
+			log.Println()
+			log.Println("To run the Ollama server, use:")
+			log.Println()
+			log.Println("    ./ollama serve")
+			log.Println()
+		}
+	}()
+
+	if flag.NArg() > 0 {
+		flag.Usage()
+	}
+
+	if !inRootDir() {
+		log.Fatalf("Please run this script from the root directory of the Ollama project.")
+	}
+
+	if err := checkDependencies(); err != nil {
+		log.Fatalf("Failed dependency check: %v", err)
+	}
+	if err := buildLlammaCPP(); err != nil {
+		log.Fatalf("Failed to build llama.cpp: %v", err)
+	}
+	if err := goBuildOllama(); err != nil {
+		log.Fatalf("Failed to build ollama Go binary: %v", err)
+	}
+}
+
+// checkDependencies does a quick check to see if the required dependencies are
+// installed on the system and functioning enough to print their version.
+//
+// TODO(bmizerany): Check the actual version of the dependencies? Seems a
+// little daunting given diff versions might print diff things. This should
+// be good enough for now.
+func checkDependencies() error {
+	var err error
+	check := func(name string, args ...string) {
+		log.Printf("=== Checking for %s ===", name)
+		defer log.Printf("=== Done checking for %s ===\n\n", name)
+		cmd := exec.Command(name, args...)
+		cmd.Stdout = os.Stdout
+		cmd.Stderr = os.Stderr
+		err = errors.Join(err, cmd.Run())
+	}
+
+	check("cmake", "--version")
+	check("gcc", "--version")
+	return err
+}
+
+func goBuildOllama() error {
+	log.Println("=== Building Ollama binary ===")
+	defer log.Printf("=== Done building Ollama binary ===\n\n")
+	if *flagSkipBuild {
+		log.Println("Skipping 'go build -o ollama .'")
+		return nil
+	}
+	cmd := exec.Command("go", "build", "-o", "ollama", ".")
+	cmd.Stdout = os.Stdout
+	cmd.Stderr = os.Stderr
+	cmd.Env = buildEnv()
+	return cmd.Run()
+}
+
+// buildLlammaCPP generates the llama.cpp bindings for the current platform.
+//
+// It assumes that the current working directory is the root directory of the
+// Ollama project.
+func buildLlammaCPP() error {
+	log.Println("=== Generating dependencies ===")
+	defer log.Printf("=== Done generating dependencies ===\n\n")
+	if *flagRegenerateDestroy {
+		if err := os.RemoveAll(filepath.Join("llm", "build")); err != nil {
+			return err
+		}
+	}
+	if isDirectory(filepath.Join("llm", "build")) && !*flagRegenerateGently {
+		log.Println("llm/build already exists; skipping.  Use -d or -g to re-generate.")
+		return nil
+	}
+
+	scriptDir, err := filepath.Abs(filepath.Join("llm", "generate"))
+	if err != nil {
+		return err
+	}
+
+	var cmd *exec.Cmd
+	switch runtime.GOOS {
+	case "windows":
+		script := filepath.Join(scriptDir, "gen_windows.ps1")
+		cmd = exec.Command("powershell", "-ExecutionPolicy", "Bypass", "-File", script)
+	case "linux":
+		script := filepath.Join(scriptDir, "gen_linux.sh")
+		cmd = exec.Command("bash", script)
+	case "darwin":
+		script := filepath.Join(scriptDir, "gen_darwin.sh")
+		cmd = exec.Command("bash", script)
+	default:
+		log.Fatalf("Unsupported OS: %s", runtime.GOOS)
+	}
+	cmd.Dir = filepath.Join("llm", "generate")
+	cmd.Stdout = os.Stdout
+	cmd.Stderr = os.Stderr
+	cmd.Env = buildEnv()
+
+	log.Printf("Running GOOS=%s GOARCH=%s %s", runtime.GOOS, runtime.GOARCH, cmd.Args)
+
+	return cmd.Run()
+}
+
+func isDirectory(path string) bool {
+	info, err := os.Stat(path)
+	if err != nil {
+		return false
+	}
+	return info.IsDir()
+}
+
+// inRootDir returns true if the current working directory is the root
+// directory of the Ollama project. It looks for a file named "go.mod".
+func inRootDir() bool {
+	_, err := os.Stat("go.mod")
+	return err == nil
+}

cmd/cmd.go

@@ -162,9 +162,6 @@ func tempZipFiles(path string) (string, error) {
 	}
 	defer tempfile.Close()
 
-	zipfile := zip.NewWriter(tempfile)
-	defer zipfile.Close()
-
 	detectContentType := func(path string) (string, error) {
 		f, err := os.Open(path)
 		if err != nil {
@@ -233,6 +230,9 @@ func tempZipFiles(path string) (string, error) {
 		files = append(files, tks...)
 	}
 
+	zipfile := zip.NewWriter(tempfile)
+	defer zipfile.Close()
+
 	for _, file := range files {
 		f, err := os.Open(file)
 		if err != nil {
@@ -624,13 +624,13 @@ func ShowHandler(cmd *cobra.Command, args []string) error {
 		return errors.New("only one of '--license', '--modelfile', '--parameters', '--system', or '--template' can be specified")
 	}
 
-	if flagsSet == 1 {
-		req := api.ShowRequest{Name: args[0]}
-		resp, err := client.Show(cmd.Context(), &req)
-		if err != nil {
-			return err
-		}
+	req := api.ShowRequest{Name: args[0]}
+	resp, err := client.Show(cmd.Context(), &req)
+	if err != nil {
+		return err
+	}
 
+	if flagsSet == 1 {
 		switch showType {
 		case "license":
 			fmt.Println(resp.License)
@@ -647,12 +647,12 @@ func ShowHandler(cmd *cobra.Command, args []string) error {
 		return nil
 	}
 
-	req := api.ShowRequest{Name: args[0]}
-	resp, err := client.Show(cmd.Context(), &req)
-	if err != nil {
-		return err
-	}
+	showInfo(resp)
 
+	return nil
+}
+
+func showInfo(resp *api.ShowResponse) {
 	arch := resp.ModelInfo["general.architecture"].(string)
 
 	modelData := [][]string{
@@ -672,11 +672,17 @@ func ShowHandler(cmd *cobra.Command, args []string) error {
 		projectorData := [][]string{
 			{"arch", "clip"},
 			{"parameters", format.HumanNumber(uint64(resp.ProjectorInfo["general.parameter_count"].(float64)))},
-			{"projector type", resp.ProjectorInfo["clip.projector_type"].(string)},
-			{"embedding length", fmt.Sprintf("%v", resp.ProjectorInfo["clip.vision.embedding_length"].(float64))},
-			{"projection dimensionality", fmt.Sprintf("%v", resp.ProjectorInfo["clip.vision.projection_dim"].(float64))},
 		}
 
+		if projectorType, ok := resp.ProjectorInfo["clip.projector_type"]; ok {
+			projectorData = append(projectorData, []string{"projector type", projectorType.(string)})
+		}
+
+		projectorData = append(projectorData,
+			[]string{"embedding length", fmt.Sprintf("%v", resp.ProjectorInfo["clip.vision.embedding_length"].(float64))},
+			[]string{"projection dimensionality", fmt.Sprintf("%v", resp.ProjectorInfo["clip.vision.projection_dim"].(float64))},
+		)
+
 		mainTableData = append(mainTableData,
 			[]string{"Projector"},
 			[]string{renderSubTable(projectorData, false)},
@@ -705,8 +711,6 @@ func ShowHandler(cmd *cobra.Command, args []string) error {
 	}
 
 	table.Render()
-
-	return nil
 }
 
 func renderSubTable(data [][]string, file bool) string {

cmd/interactive.go

@@ -404,15 +404,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 
 				switch args[1] {
 				case "info":
-					fmt.Println("Model details:")
-					if len(resp.Details.Families) > 0 {
-						fmt.Printf("Family              %s\n", strings.Join(resp.Details.Families, ", "))
-					} else if resp.Details.Family != "" {
-						fmt.Printf("Family              %s\n", resp.Details.Family)
-					}
-					fmt.Printf("Parameter Size      %s\n", resp.Details.ParameterSize)
-					fmt.Printf("Quantization Level  %s\n", resp.Details.QuantizationLevel)
-					fmt.Println("")
+					showInfo(resp)
 				case "license":
 					if resp.License == "" {
 						fmt.Println("No license was specified for this model.")

docs/api.md

@@ -26,7 +26,7 @@ All durations are returned in nanoseconds.
 
 ### Streaming responses
 
-Certain endpoints stream responses as JSON objects and can optional return non-streamed responses.
+Certain endpoints stream responses as JSON objects. Streaming can be disabled by providing `{"stream": false}` for these endpoints.
 
 ## Generate a completion
 

docs/development.md

@@ -25,13 +25,7 @@ export OLLAMA_DEBUG=1
 Get the required libraries and build the native LLM code:
 
 ```bash
-go generate ./...
-```
-
-Then build ollama:
-
-```bash
-go build .
+go run build.go
 ```
 
 Now you can run `ollama`:
@@ -40,6 +34,16 @@ Now you can run `ollama`:
 ./ollama
 ```
 
+### Rebuilding the native code
+
+If at any point you need to rebuild the native code, you can run the
+build.go script again using the `-f` flag to force a rebuild, and,
+optionally, the `-d` flag to skip building the Go binary:
+
+```bash
+go run build.go -d -s
+```
+
 ### Linux
 
 #### Linux CUDA (NVIDIA)
@@ -55,16 +59,10 @@ specifying an environment variable `CUDA_LIB_DIR` to the location of the shared
 libraries, and `CUDACXX` to the location of the nvcc compiler. You can customize
 a set of target CUDA architectures by setting `CMAKE_CUDA_ARCHITECTURES` (e.g. "50;60;70")
 
-Then generate dependencies:
-
-```
-go generate ./...
-```
-
 Then build the binary:
 
 ```
-go build .
+go run build.go
 ```
 
 #### Linux ROCm (AMD)
@@ -80,21 +78,17 @@ install (typically `/opt/rocm`), and `CLBlast_DIR` to the location of the
 CLBlast install (typically `/usr/lib/cmake/CLBlast`). You can also customize
 the AMD GPU targets by setting AMDGPU_TARGETS (e.g. `AMDGPU_TARGETS="gfx1101;gfx1102"`)
 
-```
-go generate ./...
-```
-
 Then build the binary:
 
 ```
-go build .
+go run build.go
 ```
 
 ROCm requires elevated privileges to access the GPU at runtime. On most distros you can add your user account to the `render` group, or run as root.
 
 #### Advanced CPU Settings
 
-By default, running `go generate ./...` will compile a few different variations
+By default, running `go run build.go` will compile a few different variations
 of the LLM library based on common CPU families and vector math capabilities,
 including a lowest-common-denominator which should run on almost any 64 bit CPU
 somewhat slowly. At runtime, Ollama will auto-detect the optimal variation to
@@ -104,8 +98,7 @@ like to use. For example, to compile an optimized binary for an Intel i9-9880H,
 you might use:
 
 ```
-OLLAMA_CUSTOM_CPU_DEFS="-DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_F16C=on -DLLAMA_FMA=on" go generate ./...
-go build .
+OLLAMA_CUSTOM_CPU_DEFS="-DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_F16C=on -DLLAMA_FMA=on" go run build.go
 ```
 
 #### Containerized Linux Build
@@ -129,8 +122,7 @@ Then, build the `ollama` binary:
 
 ```powershell
 $env:CGO_ENABLED="1"
-go generate ./...
-go build .
+go run build.go
 ```
 
 #### Windows CUDA (NVIDIA)

docs/gpu.md

@@ -18,7 +18,7 @@ Check your compute compatibility to see if your card is supported:
 |                    | Quadro              | `RTX 8000` `RTX 6000` `RTX 5000` `RTX 4000`                                                                 |
 | 7.0                | NVIDIA              | `TITAN V` `V100` `Quadro GV100`                                                                             |
 | 6.1                | NVIDIA TITAN        | `TITAN Xp` `TITAN X`                                                                                        |
-|                    | GeForce GTX         | `GTX 1080 Ti` `GTX 1080` `GTX 1070 Ti` `GTX 1070` `GTX 1060` `GTX 1050`                                     |
+|                    | GeForce GTX         | `GTX 1080 Ti` `GTX 1080` `GTX 1070 Ti` `GTX 1070` `GTX 1060` `GTX 1050 Ti` `GTX 1050`                       |
 |                    | Quadro              | `P6000` `P5200` `P4200` `P3200` `P5000` `P4000` `P3000` `P2200` `P2000` `P1000` `P620` `P600` `P500` `P520` |
 |                    | Tesla               | `P40` `P4`                                                                                                  |
 | 6.0                | NVIDIA              | `Tesla P100` `Quadro GP100`                                                                                 |

docs/openai.md

@@ -104,7 +104,6 @@ curl http://localhost:11434/v1/chat/completions \
 
 #### Notes
 
-- `finish_reason` will always be `stop`
 - `usage.prompt_tokens` will be 0 for completions where prompt evaluation is cached
 
 ## Models

llm/ext_server/server.cpp

@@ -1650,26 +1650,41 @@ struct llama_server_context
                     }
                     slot.params.n_keep = std::min(slot.n_ctx - 4, slot.params.n_keep);
 
+                    char buf[256];
+                    llama_model_meta_val_str(model, "general.architecture", buf, 256);
+                    bool gemma2 = strcmp(buf, "gemma2") == 0;
+
+                    int32_t truncate_at = slot.n_ctx;
+
+                    // truncate at 2/3 of the context length for gemma2 models
+                    // as they do not support context shifts (from the sliding window implementation).
+                    // this way, prompts that almost fit the context length can still generate a full
+                    // response without a sudden stop from hitting the context limit
+                    if (gemma2) {
+                        truncate_at = 2 * slot.n_ctx / 3;
+                    }
+
                     // if input prompt is too big, truncate it, if group attention self-extend is disabled
-                    if (slot.ga_n == 1 && slot.n_prompt_tokens >= slot.n_ctx)
+                    if (slot.ga_n == 1 && slot.n_prompt_tokens >= truncate_at)
                     {
                         const int n_left = slot.n_ctx - slot.params.n_keep;
-                        const int n_block_size = n_left / 2;
-                        const int erased_blocks = (slot.n_prompt_tokens - slot.params.n_keep - n_block_size) / n_block_size;
+                        const int n_shift = n_left / 2;
+                        const int n_erase = slot.n_prompt_tokens - slot.params.n_keep - n_shift;
 
                         std::vector<llama_token> new_tokens(
                             prompt_tokens.begin(),
                             prompt_tokens.begin() + slot.params.n_keep);
                         new_tokens.insert(
                             new_tokens.end(),
-                            prompt_tokens.begin() + slot.params.n_keep + erased_blocks * n_block_size,
+                            prompt_tokens.begin() + slot.params.n_keep + n_erase,
                             prompt_tokens.end());
 
-                        LOG_VERBOSE("input truncated", {
-                            {"n_ctx",      slot.n_ctx},
-                            {"n_keep",     slot.params.n_keep},
-                            {"n_left",     n_left},
-                            {"new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend())},
+                        LOG_INFO("input truncated", {
+                            {"n_ctx",        slot.n_ctx},
+                            {"n_keep",       slot.params.n_keep},
+                            {"n_left",       n_left},
+                            {"n_shift",      n_shift},
+                            {"n_erase",      n_erase},
                         });
                         slot.truncated = true;
                         prompt_tokens = new_tokens;
@@ -1678,6 +1693,19 @@ struct llama_server_context
                         GGML_ASSERT(slot.n_prompt_tokens < slot.n_ctx);
                     }
 
+                    // Models with sliding window attention do not work with context shifts, so
+                    // limit their prediction to the context length
+                    if (gemma2) {
+                        int32_t limit = slot.n_ctx - slot.n_prompt_tokens;
+                        slot.n_predict = limit;
+                        slot.params.n_predict = limit;
+                        LOG_INFO("model does not support sliding window, limiting generation", {
+                            {"n_ctx", slot.n_ctx},
+                            {"n_prompt_tokens", slot.n_prompt_tokens},
+                            {"n_predict", slot.n_predict}
+                        });
+                    }
+
                     if (!slot.params.cache_prompt)
                     {
                         llama_sampling_reset(slot.ctx_sampling);

llm/generate/gen_darwin.sh

@@ -1,6 +1,6 @@
 #!/bin/bash
-# This script is intended to run inside the go generate
-# working directory must be ./llm/generate/
+# This script is intended to run inside the `go run build.go` script, which
+# sets the working directory to the correct location: ./llm/generate/.
 
 # TODO - add hardening to detect missing tools (cmake, etc.)
 
@@ -92,10 +92,10 @@ case "${GOARCH}" in
     ;;
 *)
     echo "GOARCH must be set"
-    echo "this script is meant to be run from within go generate"
+    echo "this script is meant to be run from within 'go run build.go'"
     exit 1
     ;;
 esac
 
 cleanup
-echo "go generate completed.  LLM runners: $(cd ${BUILD_DIR}/..; echo *)"
+echo "code generation completed.  LLM runners: $(cd ${BUILD_DIR}/..; echo *)"

llm/generate/gen_linux.sh

@@ -1,6 +1,6 @@
 #!/bin/bash
-# This script is intended to run inside the go generate
-# working directory must be llm/generate/
+# This script is intended to run with the `go run build.go` script, which
+# sets the working directory to the correct location: ./llm/generate/.
 
 # First we build one or more CPU based LLM libraries
 #
@@ -281,4 +281,4 @@ if [ -z "${OLLAMA_SKIP_ROCM_GENERATE}" -a -d "${ROCM_PATH}" ]; then
 fi
 
 cleanup
-echo "go generate completed.  LLM runners: $(cd ${BUILD_DIR}/..; echo *)"
+echo "code generation completed.  LLM runners: $(cd ${BUILD_DIR}/..; echo *)"

llm/generate/gen_windows.ps1

@@ -26,26 +26,15 @@ function amdGPUs {
     $GPU_LIST -join ';'
 }
 
-
 function init_vars {
-    if (!$script:SRC_DIR) {
-        $script:SRC_DIR = $(resolve-path "..\..\")
-    }
-    if (!$script:llamacppDir) {
-        $script:llamacppDir = "../llama.cpp"
-    }
-    if (!$script:cmakeTargets) {
-        $script:cmakeTargets = @("ollama_llama_server")
-    }
+    $script:SRC_DIR = $(resolve-path "..\..\")
+    $script:llamacppDir = "../llama.cpp"
     $script:cmakeDefs = @(
         "-DBUILD_SHARED_LIBS=on",
-        "-DLLAMA_NATIVE=off",
-        "-DLLAMA_OPENMP=off"
+        "-DLLAMA_NATIVE=off"
         )
-    $script:commonCpuDefs = @("-DCMAKE_POSITION_INDEPENDENT_CODE=on")
-    $script:ARCH = $Env:PROCESSOR_ARCHITECTURE.ToLower()
-    $script:DIST_BASE = "${script:SRC_DIR}\dist\windows-${script:ARCH}\ollama_runners"
-    md "$script:DIST_BASE" -ea 0 > $null
+    $script:cmakeTargets = @("ollama_llama_server")
+    $script:ARCH = "amd64" # arm not yet supported.
     if ($env:CGO_CFLAGS -contains "-g") {
         $script:cmakeDefs += @("-DCMAKE_VERBOSE_MAKEFILE=on", "-DLLAMA_SERVER_VERBOSE=on", "-DCMAKE_BUILD_TYPE=RelWithDebInfo")
         $script:config = "RelWithDebInfo"
@@ -66,6 +55,7 @@ function init_vars {
     } else {
         $script:CUDA_LIB_DIR=$env:CUDA_LIB_DIR
     }
+    $script:GZIP=(get-command -ea 'silentlycontinue' gzip).path
     $script:DUMPBIN=(get-command -ea 'silentlycontinue' dumpbin).path
     if ($null -eq $env:CMAKE_CUDA_ARCHITECTURES) {
         $script:CMAKE_CUDA_ARCHITECTURES="50;52;61;70;75;80"
@@ -123,13 +113,8 @@ function build {
     & cmake --version
     & cmake -S "${script:llamacppDir}" -B $script:buildDir $script:cmakeDefs
     if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
-    if ($cmakeDefs -contains "-G") {
-        $extra=@("-j8")
-    } else {
-        $extra= @("--", "/p:CL_MPcount=8")
-    }
-    write-host "building with: cmake --build $script:buildDir --config $script:config $($script:cmakeTargets | ForEach-Object { `"--target`", $_ }) $extra"
-    & cmake --build $script:buildDir --config $script:config ($script:cmakeTargets | ForEach-Object { "--target", $_ }) $extra
+    write-host "building with: cmake --build $script:buildDir --config $script:config $($script:cmakeTargets | ForEach-Object { `"--target`", $_ })"
+    & cmake --build $script:buildDir --config $script:config ($script:cmakeTargets | ForEach-Object { "--target", $_ })
     if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
     # Rearrange output to be consistent between different generators
     if ($null -ne ${script:config} -And (test-path -path "${script:buildDir}/bin/${script:config}" ) ) {
@@ -149,18 +134,21 @@ function sign {
     }
 }
 
-function install {
-    write-host "Installing binaries to dist dir ${script:distDir}"
-    mkdir ${script:distDir} -ErrorAction SilentlyContinue
+function compress {
+    if ($script:GZIP -eq $null) {
+        write-host "gzip not installed, not compressing files"
+        return
+    }
+    write-host "Compressing binaries..."
     $binaries = dir "${script:buildDir}/bin/*.exe"
     foreach ($file in $binaries) {
-        copy-item -Path $file -Destination ${script:distDir} -Force
+        & "$script:GZIP" --best -f $file
     }
 
-    write-host "Installing dlls to dist dir ${script:distDir}"
+    write-host "Compressing dlls..."
     $dlls = dir "${script:buildDir}/bin/*.dll"
     foreach ($file in $dlls) {
-        copy-item -Path $file -Destination ${script:distDir} -Force
+        & "$script:GZIP" --best -f $file
     }
 }
 
@@ -181,252 +169,132 @@ function cleanup {
     }
 }
 
+init_vars
+git_module_setup
+apply_patches
 
 # -DLLAMA_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
 # -DLLAMA_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
 # -DLLAMA_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
 
+$script:commonCpuDefs = @("-DCMAKE_POSITION_INDEPENDENT_CODE=on")
 
-function build_static() {
-    if ((-not "${env:OLLAMA_SKIP_STATIC_GENERATE}") -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "static"))) {
-        # GCC build for direct linking into the Go binary
-        init_vars
-        # cmake will silently fallback to msvc compilers if mingw isn't in the path, so detect and fail fast
-        # as we need this to be compiled by gcc for golang to be able to link with itx
-        write-host "Checking for MinGW..."
-        # error action ensures we exit on failure
-        get-command gcc
-        get-command mingw32-make
-        $oldTargets = $script:cmakeTargets
-        $script:cmakeTargets = @("llama", "ggml")
-        $script:cmakeDefs = @(
-            "-G", "MinGW Makefiles"
-            "-DCMAKE_C_COMPILER=gcc.exe",
-            "-DCMAKE_CXX_COMPILER=g++.exe",
-            "-DBUILD_SHARED_LIBS=off",
-            "-DLLAMA_NATIVE=off",
-            "-DLLAMA_AVX=off",
-            "-DLLAMA_AVX2=off",
-            "-DLLAMA_AVX512=off",
-            "-DLLAMA_F16C=off",
-            "-DLLAMA_FMA=off",
-            "-DLLAMA_OPENMP=off")
-        $script:buildDir="../build/windows/${script:ARCH}_static"
-        write-host "Building static library"
-        build
-        $script:cmakeTargets = $oldTargets
-    } else {
-        write-host "Skipping CPU generation step as requested"
-    }
+if ($null -eq ${env:OLLAMA_SKIP_CPU_GENERATE}) {
+
+# GCC build for direct linking into the Go binary
+init_vars
+# cmake will silently fallback to msvc compilers if mingw isn't in the path, so detect and fail fast
+# as we need this to be compiled by gcc for golang to be able to link with itx
+write-host "Checking for MinGW..."
+# error action ensures we exit on failure
+get-command gcc
+get-command mingw32-make
+$script:cmakeTargets = @("llama", "ggml")
+$script:cmakeDefs = @(
+    "-G", "MinGW Makefiles"
+    "-DCMAKE_C_COMPILER=gcc.exe",
+    "-DCMAKE_CXX_COMPILER=g++.exe",
+    "-DBUILD_SHARED_LIBS=off",
+    "-DLLAMA_NATIVE=off",
+    "-DLLAMA_AVX=off",
+    "-DLLAMA_AVX2=off",
+    "-DLLAMA_AVX512=off",
+    "-DLLAMA_F16C=off",
+    "-DLLAMA_FMA=off")
+$script:buildDir="../build/windows/${script:ARCH}_static"
+write-host "Building static library"
+build
+
+# remaining llama.cpp builds use MSVC 
+    init_vars
+    $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=off", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
+    $script:buildDir="../build/windows/${script:ARCH}/cpu"
+    write-host "Building LCD CPU"
+    build
+    sign
+    compress
+
+    init_vars
+    $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
+    $script:buildDir="../build/windows/${script:ARCH}/cpu_avx"
+    write-host "Building AVX CPU"
+    build
+    sign
+    compress
+
+    init_vars
+    $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=on", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=on", "-DLLAMA_F16C=on") + $script:cmakeDefs
+    $script:buildDir="../build/windows/${script:ARCH}/cpu_avx2"
+    write-host "Building AVX2 CPU"
+    build
+    sign
+    compress
+} else {
+    write-host "Skipping CPU generation step as requested"
 }
 
-function build_cpu($gen_arch) {
-    if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu"))) {
-        # remaining llama.cpp builds use MSVC 
-        init_vars
-        $script:cmakeDefs = $script:commonCpuDefs + @("-A", $gen_arch, "-DLLAMA_AVX=off", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
-        $script:buildDir="../build/windows/${script:ARCH}/cpu"
-        $script:distDir="$script:DIST_BASE\cpu"
-        write-host "Building LCD CPU"
-        build
-        sign
-        install
-    } else {
-        write-host "Skipping CPU generation step as requested"
-    }
-}
-
-function build_cpu_avx() {
-    if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu_avx"))) {
-        init_vars
-        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
-        $script:buildDir="../build/windows/${script:ARCH}/cpu_avx"
-        $script:distDir="$script:DIST_BASE\cpu_avx"
-        write-host "Building AVX CPU"
-        build
-        sign
-        install
-    } else {
-        write-host "Skipping CPU AVX generation step as requested"
-    }
-}
-
-function build_cpu_avx2() {
-    if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu_avx2"))) {
-        init_vars
-        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=on", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=on", "-DLLAMA_F16C=on") + $script:cmakeDefs
-        $script:buildDir="../build/windows/${script:ARCH}/cpu_avx2"
-        $script:distDir="$script:DIST_BASE\cpu_avx2"
-        write-host "Building AVX2 CPU"
-        build
-        sign
-        install
-    } else {
-        write-host "Skipping CPU AVX2 generation step as requested"
-    }
-}
-
-function build_cuda() {
-    if ((-not "${env:OLLAMA_SKIP_CUDA_GENERATE}") -and ("${script:CUDA_LIB_DIR}")) {
-        # Then build cuda as a dynamically loaded library
-        $nvcc = "$script:CUDA_LIB_DIR\nvcc.exe"
-        $script:CUDA_VERSION=(get-item ($nvcc | split-path | split-path)).Basename
-        if ($null -ne $script:CUDA_VERSION) {
-            $script:CUDA_VARIANT="_"+$script:CUDA_VERSION
-        }
-        init_vars
-        $script:buildDir="../build/windows/${script:ARCH}/cuda$script:CUDA_VARIANT"
-        $script:distDir="$script:DIST_BASE\cuda$script:CUDA_VARIANT"
-        $script:cmakeDefs += @(
-            "-A", "x64",
-            "-DLLAMA_CUDA=ON",
-            "-DLLAMA_AVX=on",
-            "-DLLAMA_AVX2=off",
-            "-DCUDAToolkit_INCLUDE_DIR=$script:CUDA_INCLUDE_DIR",
-            "-DCMAKE_CUDA_FLAGS=-t8",
-            "-DCMAKE_CUDA_ARCHITECTURES=${script:CMAKE_CUDA_ARCHITECTURES}"
-            )
-        if ($null -ne $env:OLLAMA_CUSTOM_CUDA_DEFS) {
-            write-host "OLLAMA_CUSTOM_CUDA_DEFS=`"${env:OLLAMA_CUSTOM_CUDA_DEFS}`""
-            $script:cmakeDefs +=@("${env:OLLAMA_CUSTOM_CUDA_DEFS}")
-            write-host "building custom CUDA GPU"
-        }
-        build
-        sign
-        install
-
-        rm -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
-        md "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\" -ea 0 > $null
-        write-host "copying CUDA dependencies to ${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
-        cp "${script:CUDA_LIB_DIR}\cudart64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
-        cp "${script:CUDA_LIB_DIR}\cublas64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
-        cp "${script:CUDA_LIB_DIR}\cublasLt64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
-    } else {
-        write-host "Skipping CUDA generation step"
-    }
-}
-
-function build_oneapi() {
-  if ((-not "${env:OLLAMA_SKIP_ONEAPI_GENERATE}") -and ("${env:ONEAPI_ROOT}"))  {
-    # Get oneAPI version
-    $script:ONEAPI_VERSION = icpx --version
-    $script:ONEAPI_VERSION = [regex]::Match($script:ONEAPI_VERSION, '(?<=oneAPI DPC\+\+/C\+\+ Compiler )(?<version>\d+\.\d+\.\d+)').Value
-    if ($null -ne $script:ONEAPI_VERSION) {
-      $script:ONEAPI_VARIANT = "_v" + $script:ONEAPI_VERSION
+if ($null -ne $script:CUDA_LIB_DIR) {
+    # Then build cuda as a dynamically loaded library
+    $nvcc = "$script:CUDA_LIB_DIR\nvcc.exe"
+    $script:CUDA_VERSION=(get-item ($nvcc | split-path | split-path)).Basename
+    if ($null -ne $script:CUDA_VERSION) {
+        $script:CUDA_VARIANT="_"+$script:CUDA_VERSION
     }
     init_vars
-    $script:buildDir = "../build/windows/${script:ARCH}/oneapi$script:ONEAPI_VARIANT"
-    $script:distDir ="$script:DIST_BASE\oneapi$script:ONEAPI_VARIANT"
-    $script:cmakeDefs += @(
-      "-G", "MinGW Makefiles",
-      "-DLLAMA_SYCL=ON",
-      "-DCMAKE_C_COMPILER=icx",
-      "-DCMAKE_CXX_COMPILER=icx",
-      "-DCMAKE_BUILD_TYPE=Release"
-    )
+    $script:buildDir="../build/windows/${script:ARCH}/cuda$script:CUDA_VARIANT"
+    $script:cmakeDefs += @("-A", "x64", "-DLLAMA_CUDA=ON", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=off", "-DCUDAToolkit_INCLUDE_DIR=$script:CUDA_INCLUDE_DIR", "-DCMAKE_CUDA_ARCHITECTURES=${script:CMAKE_CUDA_ARCHITECTURES}")
+    if ($null -ne $env:OLLAMA_CUSTOM_CUDA_DEFS) {
+        write-host "OLLAMA_CUSTOM_CUDA_DEFS=`"${env:OLLAMA_CUSTOM_CUDA_DEFS}`""
+        $script:cmakeDefs +=@("${env:OLLAMA_CUSTOM_CUDA_DEFS}")
+        write-host "building custom CUDA GPU"
+    }
+    build
+    sign
+    compress
+}
 
-    Write-Host "Building oneAPI"
+if ($null -ne $env:HIP_PATH) {
+    $script:ROCM_VERSION=(get-item $env:HIP_PATH).Basename
+    if ($null -ne $script:ROCM_VERSION) {
+        $script:ROCM_VARIANT="_v"+$script:ROCM_VERSION
+    }
+
+    init_vars
+    $script:buildDir="../build/windows/${script:ARCH}/rocm$script:ROCM_VARIANT"
+    $script:cmakeDefs += @(
+        "-G", "Ninja", 
+        "-DCMAKE_C_COMPILER=clang.exe",
+        "-DCMAKE_CXX_COMPILER=clang++.exe",
+        "-DLLAMA_HIPBLAS=on",
+        "-DHIP_PLATFORM=amd",
+        "-DLLAMA_AVX=on",
+        "-DLLAMA_AVX2=off",
+        "-DCMAKE_POSITION_INDEPENDENT_CODE=on",
+        "-DAMDGPU_TARGETS=$(amdGPUs)",
+        "-DGPU_TARGETS=$(amdGPUs)"
+        )
+
+    # Make sure the ROCm binary dir is first in the path
+    $env:PATH="$env:HIP_PATH\bin;$env:PATH"
+
+    # We have to clobber the LIB var from the developer shell for clang to work properly
+    $env:LIB=""
+    if ($null -ne $env:OLLAMA_CUSTOM_ROCM_DEFS) {
+        write-host "OLLAMA_CUSTOM_ROCM_DEFS=`"${env:OLLAMA_CUSTOM_ROCM_DEFS}`""
+        $script:cmakeDefs += @("${env:OLLAMA_CUSTOM_ROCM_DEFS}")
+        write-host "building custom ROCM GPU"
+    }
+    write-host "Building ROCm"
     build
     # Ninja doesn't prefix with config name
+    ${script:config}=""
     if ($null -ne $script:DUMPBIN) {
-      & "$script:DUMPBIN" /dependents "${script:buildDir}/bin/ollama_llama_server.exe" | Select-String ".dll"
+        & "$script:DUMPBIN" /dependents "${script:buildDir}/bin/ollama_llama_server.exe" | select-string ".dll"
     }
     sign
-    install
-
-    rm -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    md "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\" -ea 0 > $null
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libirngmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libmmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_level_zero.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_unified_runtime.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_win_proxy_loader.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\svml_dispmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\sycl7.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_core.2.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_sycl_blas.4.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_tbb_thread.2.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-  } else {
-    Write-Host "Skipping oneAPI generation step"
-  }
+    compress
 }
 
-function build_rocm() {
-    if ((-not "${env:OLLAMA_SKIP_ROCM_GENERATE}") -and ("${env:HIP_PATH}")) {
-        $script:ROCM_VERSION=(get-item $env:HIP_PATH).Basename
-        if ($null -ne $script:ROCM_VERSION) {
-            $script:ROCM_VARIANT="_v"+$script:ROCM_VERSION
-        }
 
-        init_vars
-        $script:buildDir="../build/windows/${script:ARCH}/rocm$script:ROCM_VARIANT"
-        $script:distDir="$script:DIST_BASE\rocm$script:ROCM_VARIANT"
-        $script:cmakeDefs += @(
-            "-G", "Ninja", 
-            "-DCMAKE_C_COMPILER=clang.exe",
-            "-DCMAKE_CXX_COMPILER=clang++.exe",
-            "-DLLAMA_HIPBLAS=on",
-            "-DHIP_PLATFORM=amd",
-            "-DLLAMA_AVX=on",
-            "-DLLAMA_AVX2=off",
-            "-DCMAKE_POSITION_INDEPENDENT_CODE=on",
-            "-DAMDGPU_TARGETS=$(amdGPUs)",
-            "-DGPU_TARGETS=$(amdGPUs)"
-            )
-
-        # Make sure the ROCm binary dir is first in the path
-        $env:PATH="$env:HIP_PATH\bin;$env:PATH"
-
-        # We have to clobber the LIB var from the developer shell for clang to work properly
-        $env:LIB=""
-        if ($null -ne $env:OLLAMA_CUSTOM_ROCM_DEFS) {
-            write-host "OLLAMA_CUSTOM_ROCM_DEFS=`"${env:OLLAMA_CUSTOM_ROCM_DEFS}`""
-            $script:cmakeDefs += @("${env:OLLAMA_CUSTOM_ROCM_DEFS}")
-            write-host "building custom ROCM GPU"
-        }
-        write-host "Building ROCm"
-        build
-        # Ninja doesn't prefix with config name
-        ${script:config}=""
-        if ($null -ne $script:DUMPBIN) {
-            & "$script:DUMPBIN" /dependents "${script:buildDir}/bin/ollama_llama_server.exe" | select-string ".dll"
-        }
-        sign
-        install
-
-        # Assumes v5.7, may need adjustments for v6
-        rm -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\"
-        md "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\rocblas\library\" -ea 0 > $null
-        cp "${env:HIP_PATH}\bin\hipblas.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\"
-        cp "${env:HIP_PATH}\bin\rocblas.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\"
-        # amdhip64.dll dependency comes from the driver and must be installed on the host to use AMD GPUs
-        cp "${env:HIP_PATH}\bin\rocblas\library\*" "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\rocblas\library\"
-    } else {
-        write-host "Skipping ROCm generation step"
-    }
-}
-
-init_vars
-if ($($args.count) -eq 0) {
-    git_module_setup
-    apply_patches
-    build_static
-    if ($script:ARCH -eq "arm64") {
-        build_cpu("ARM64")
-    } else { # amd64
-        build_cpu("x64")
-        build_cpu_avx
-        build_cpu_avx2
-        build_cuda
-        build_oneapi
-        build_rocm
-    }
-
-    cleanup
-    write-host "`ngo generate completed.  LLM runners: $(get-childitem -path $script:DIST_BASE)"
-} else {
-    for ( $i = 0; $i -lt $args.count; $i++ ) {
-        write-host "performing $($args[$i])"
-        & $($args[$i])
-    } 
-}
+cleanup
+write-host "`code generation completed.  LLM runners: $(get-childitem -path ${script:SRC_DIR}\llm\build\windows\${script:ARCH})"

llm/generate/generate_darwin.go

@@ -1,3 +0,0 @@
-package generate
-
-//go:generate bash ./gen_darwin.sh

llm/generate/generate_linux.go

@@ -1,3 +0,0 @@
-package generate
-
-//go:generate bash ./gen_linux.sh

llm/generate/generate_windows.go

@@ -1,3 +0,0 @@
-package generate
-
-//go:generate powershell -ExecutionPolicy Bypass -File ./gen_windows.ps1

llm/ggml.go

@@ -366,9 +366,18 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 				4*batch*(1+2*embedding+context*(1+heads))+embedding*(6*context*headsKV/heads+embedding*9/16),
 			)
 		}
-	case "gemma":
-		fullOffload = 4 * batch * (embedding + vocab)
-		partialOffload = 4*batch*(2*embedding+vocab+1) + embedding*vocab*105/128
+	case "gemma", "gemma2":
+		fullOffload = max(
+			4*batch*(embedding+vocab),
+			4*batch*(2+context+context*heads+2*embedding+2*embeddingHeadsK*heads),
+		)
+
+		partialOffload = max(
+			4*embedding*batch+embedding*vocab*105/128+4*vocab*batch,
+			4*batch*(2*embedding+1+2*embeddingHeadsK*heads+context+context*heads)+
+				4*embeddingHeadsK*context*8+
+				embedding*embeddingHeadsK*heads*9/16,
+		)
 	case "command-r":
 		fullOffload = max(
 			4*batch*(embedding+vocab),

llm/patches/07-gemma.diff

@@ -0,0 +1,305 @@
+From 5cadb45f39d001ffbad95b690d6cf0abcb4a6d96 Mon Sep 17 00:00:00 2001
+From: Ollama maintainers <hello@ollama.com>
+Date: Wed, 26 Jun 2024 16:18:09 -0700
+Subject: [PATCH] Architecture support
+
+---
+ llama.cpp | 194 +++++++++++++++++++++++++++++++++++++++++++++++++++++-
+ 1 file changed, 193 insertions(+), 1 deletion(-)
+
+diff --git a/llama.cpp b/llama.cpp
+index 61948751..3b4196f5 100644
+--- a/llama.cpp
++++ b/llama.cpp
+@@ -217,6 +217,7 @@ enum llm_arch {
+     LLM_ARCH_INTERNLM2,
+     LLM_ARCH_MINICPM,
+     LLM_ARCH_GEMMA,
++    LLM_ARCH_GEMMA2,
+     LLM_ARCH_STARCODER2,
+     LLM_ARCH_MAMBA,
+     LLM_ARCH_XVERSE,
+@@ -255,6 +256,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
+     { LLM_ARCH_INTERNLM2,       "internlm2"    },
+     { LLM_ARCH_MINICPM,         "minicpm"      },
+     { LLM_ARCH_GEMMA,           "gemma"        },
++    { LLM_ARCH_GEMMA2,          "gemma2"       },
+     { LLM_ARCH_STARCODER2,      "starcoder2"   },
+     { LLM_ARCH_MAMBA,           "mamba"        },
+     { LLM_ARCH_XVERSE,          "xverse"       },
+@@ -464,10 +466,12 @@ enum llm_tensor {
+     LLM_TENSOR_ATTN_NORM,
+     LLM_TENSOR_ATTN_NORM_2,
+     LLM_TENSOR_ATTN_OUT_NORM,
++    LLM_TENSOR_ATTN_POST_NORM,
+     LLM_TENSOR_ATTN_ROT_EMBD,
+     LLM_TENSOR_FFN_GATE_INP,
+     LLM_TENSOR_FFN_GATE_INP_SHEXP,
+     LLM_TENSOR_FFN_NORM,
++    LLM_TENSOR_FFN_POST_NORM,
+     LLM_TENSOR_FFN_GATE,
+     LLM_TENSOR_FFN_DOWN,
+     LLM_TENSOR_FFN_UP,
+@@ -960,6 +964,24 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
+             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+         },
+     },
++    {
++        LLM_ARCH_GEMMA2,
++        {
++            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
++            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
++            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
++            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
++            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
++            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
++            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
++            { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
++            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
++            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
++            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
++            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
++            { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
++        },
++    },
+     {
+         LLM_ARCH_STARCODER2,
+         {
+@@ -1941,6 +1963,8 @@ enum e_model {
+     MODEL_8x22B,
+     MODEL_16x12B,
+     MODEL_10B_128x3_66B,
++    MODEL_9B,
++    MODEL_27B,
+ };
+ 
+ static const size_t kiB = 1024;
+@@ -2114,6 +2138,7 @@ struct llama_layer {
+     struct ggml_tensor * attn_out_norm_b;
+     struct ggml_tensor * attn_q_a_norm;
+     struct ggml_tensor * attn_kv_a_norm;
++    struct ggml_tensor * attn_post_norm;
+ 
+     // attention
+     struct ggml_tensor * wq;
+@@ -2136,6 +2161,7 @@ struct llama_layer {
+     // normalization
+     struct ggml_tensor * ffn_norm;
+     struct ggml_tensor * ffn_norm_b;
++    struct ggml_tensor * ffn_post_norm;
+     struct ggml_tensor * layer_out_norm;
+     struct ggml_tensor * layer_out_norm_b;
+     struct ggml_tensor * ffn_norm_exps;
+@@ -4529,6 +4555,16 @@ static void llm_load_hparams(
+                 }
+             } break;
+         case LLM_ARCH_GEMMA:
++            {
++                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
++
++                switch (hparams.n_layer) {
++                    case 18: model.type = e_model::MODEL_9B; break;
++                    case 28: model.type = e_model::MODEL_27B; break;
++                    default: model.type = e_model::MODEL_UNKNOWN;
++               }
++            } break;
++        case LLM_ARCH_GEMMA2:
+             {
+                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+ 
+@@ -6305,6 +6341,40 @@ static bool llm_load_tensors(
+                         layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                     }
+                 } break;
++            case LLM_ARCH_GEMMA2:
++                {
++                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
++
++                    // output
++                    model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
++                    model.output      = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading
++
++                    const int64_t n_ff          = hparams.n_ff;
++                    const int64_t n_embd_head_k = hparams.n_embd_head_k;
++                    const int64_t n_embd_k_gqa  = hparams.n_embd_k_gqa();
++                    const int64_t n_embd_v_gqa  = hparams.n_embd_v_gqa();
++
++                    for (uint32_t i = 0; i < n_layer; ++i) {
++                        ggml_context * ctx_layer = ctx_for_layer(i);
++                        ggml_context * ctx_split = ctx_for_layer_split(i);
++
++                        auto & layer = model.layers[i];
++
++                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
++
++                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * hparams.n_head});
++                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa});
++                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa});
++                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * hparams.n_head, n_embd});
++                        layer.attn_post_norm = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd});
++
++                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
++                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
++                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
++                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
++                        layer.ffn_post_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd});
++                    }
++                } break;
+             case LLM_ARCH_STARCODER2:
+                 {
+                     model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+@@ -10614,6 +10684,123 @@ struct llm_build_context {
+         return gf;
+     }
+ 
++    struct ggml_cgraph * build_gemma2() {
++        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
++
++        const int64_t n_embd_head_k = hparams.n_embd_head_k;
++
++        struct ggml_tensor * cur;
++        struct ggml_tensor * inpL;
++
++        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
++
++        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
++        cb(inpL, "inp_scaled", -1);
++
++        // inp_pos - contains the positions
++        struct ggml_tensor * inp_pos = build_inp_pos();
++
++        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
++        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
++
++        for (int il = 0; il < n_layer; ++il) {
++            // norm
++            cur = llm_build_norm(ctx0, inpL, hparams,
++                    model.layers[il].attn_norm, NULL,
++                    LLM_NORM_RMS, cb, il);
++            cb(cur, "attn_norm", il);
++
++            // self-attention
++            {
++                // compute Q and K and RoPE them
++                struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
++                cb(Qcur, "Qcur", il);
++
++                struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
++                cb(Kcur, "Kcur", il);
++
++                struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
++                cb(Vcur, "Vcur", il);
++
++                Qcur = ggml_rope_ext(
++                        ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head,    n_tokens), inp_pos, nullptr,
++                        n_embd_head_k, rope_type, n_ctx_orig, freq_base, freq_scale,
++                        ext_factor, attn_factor, beta_fast, beta_slow);
++                cb(Qcur, "Qcur", il);
++
++                Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k)));
++                cb(Qcur, "Qcur_scaled", il);
++
++                Kcur = ggml_rope_ext(
++                        ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens), inp_pos, nullptr,
++                        n_embd_head_k, rope_type, n_ctx_orig, freq_base, freq_scale,
++                        ext_factor, attn_factor, beta_fast, beta_slow);
++                cb(Kcur, "Kcur", il);
++
++                cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
++                        model.layers[il].wo, NULL,
++                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il);
++            }
++
++            if (il == n_layer - 1) {
++                // skip computing output for unused tokens
++                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
++                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
++                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
++            }
++
++            cur = llm_build_norm(ctx0, cur, hparams,
++                    model.layers[il].attn_post_norm, NULL,
++                    LLM_NORM_RMS, cb, il);
++            cb(cur, "attn_post_norm", il);
++
++            struct ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
++            cb(sa_out, "sa_out", il);
++
++            cur = llm_build_norm(ctx0, sa_out, hparams,
++                    model.layers[il].ffn_norm, NULL,
++                    LLM_NORM_RMS, cb, il);
++            cb(cur, "ffn_norm", il);
++
++            // feed-forward network
++            {
++                cur = llm_build_ffn(ctx0, cur,
++                        model.layers[il].ffn_up, NULL,
++                        model.layers[il].ffn_gate, NULL,
++                        model.layers[il].ffn_down, NULL,
++                        NULL,
++                        LLM_FFN_GELU, LLM_FFN_PAR, cb, il);
++                cb(cur, "ffn_out", il);
++            }
++
++            cur = llm_build_norm(ctx0, cur, hparams,
++                model.layers[il].ffn_post_norm, NULL,
++                LLM_NORM_RMS, cb, -1);
++            cb(cur, "ffn_post_norm", -1);
++
++            cur = ggml_add(ctx0, cur, sa_out);
++            cb(cur, "l_out", il);
++
++            // input for next layer
++            inpL = cur;
++        }
++
++        cur = inpL;
++
++        cur = llm_build_norm(ctx0, cur, hparams,
++                model.output_norm, NULL,
++                LLM_NORM_RMS, cb, -1);
++        cb(cur, "result_norm", -1);
++
++        // lm_head
++        cur = ggml_mul_mat(ctx0, model.output, cur);
++        cb(cur, "result_output", -1);
++
++        ggml_build_forward_expand(gf, cur);
++
++        return gf;
++    }
++
+     struct ggml_cgraph * build_starcoder2() {
+         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+ 
+@@ -11847,6 +12034,10 @@ static struct ggml_cgraph * llama_build_graph(
+             {
+                 result = llm.build_gemma();
+             } break;
++        case LLM_ARCH_GEMMA2:
++            {
++                result = llm.build_gemma2();
++            } break;
+         case LLM_ARCH_STARCODER2:
+             {
+                 result = llm.build_starcoder2();
+@@ -16671,6 +16862,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
+         case LLM_ARCH_PHI2:
+         case LLM_ARCH_PHI3:
+         case LLM_ARCH_GEMMA:
++        case LLM_ARCH_GEMMA2:
+         case LLM_ARCH_STARCODER2:
+         case LLM_ARCH_GPTNEOX:
+             return LLAMA_ROPE_TYPE_NEOX;
+@@ -18551,7 +18743,7 @@ static int32_t llama_chat_apply_template_internal(
+         if (add_ass) {
+             ss << "<s>assistant\n";
+         }
+-    } else if (tmpl == "gemma" || tmpl.find("<start_of_turn>") != std::string::npos) {
++    } else if (tmpl == "gemma" || tmpl == "gemma2" || tmpl.find("<start_of_turn>") != std::string::npos) {
+         // google/gemma-7b-it
+         std::string system_prompt = "";
+         for (auto message : chat) {
+-- 
+2.45.2
+

main.go

@@ -1,5 +1,7 @@
 package main
 
+//go:generate go run build.go -g -s
+
 import (
 	"context"
 

server/model.go

@@ -11,6 +11,7 @@ import (
 	"net/http"
 	"os"
 	"path/filepath"
+	"strings"
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/convert"
@@ -77,62 +78,80 @@ func parseFromModel(ctx context.Context, name model.Name, fn func(api.ProgressRe
 	return layers, nil
 }
 
-func parseFromZipFile(_ context.Context, file *os.File, digest string, fn func(api.ProgressResponse)) (layers []*layerGGML, err error) {
+func extractFromZipFile(p string, file *os.File, fn func(api.ProgressResponse)) error {
 	stat, err := file.Stat()
 	if err != nil {
-		return nil, err
+		return err
 	}
 
 	r, err := zip.NewReader(file, stat.Size())
 	if err != nil {
-		return nil, err
+		return err
 	}
 
-	tempdir, err := os.MkdirTemp(filepath.Dir(file.Name()), "")
-	if err != nil {
-		return nil, err
-	}
-	defer os.RemoveAll(tempdir)
-
 	fn(api.ProgressResponse{Status: "unpacking model metadata"})
 	for _, f := range r.File {
+		n := filepath.Join(p, f.Name)
+		if !strings.HasPrefix(n, p) {
+			slog.Warn("skipped extracting file outside of context", "name", f.Name)
+			continue
+		}
+
+		if err := os.MkdirAll(filepath.Dir(n), 0o750); err != nil {
+			return err
+		}
+
 		// TODO(mxyng): this should not write out all files to disk
-		outfile, err := os.Create(filepath.Join(tempdir, f.Name))
+		outfile, err := os.Create(n)
 		if err != nil {
-			return nil, err
+			return err
 		}
 		defer outfile.Close()
 
 		infile, err := f.Open()
 		if err != nil {
-			return nil, err
+			return err
 		}
 		defer infile.Close()
 
 		if _, err = io.Copy(outfile, infile); err != nil {
-			return nil, err
+			return err
 		}
 
 		if err := outfile.Close(); err != nil {
-			return nil, err
+			return err
 		}
 
 		if err := infile.Close(); err != nil {
-			return nil, err
+			return err
 		}
 	}
 
-	mf, err := convert.GetModelFormat(tempdir)
+	return nil
+}
+
+func parseFromZipFile(_ context.Context, file *os.File, digest string, fn func(api.ProgressResponse)) (layers []*layerGGML, err error) {
+	tempDir, err := os.MkdirTemp(filepath.Dir(file.Name()), "")
+	if err != nil {
+		return nil, err
+	}
+	defer os.RemoveAll(tempDir)
+
+	if err := extractFromZipFile(tempDir, file, fn); err != nil {
+		return nil, err
+	}
+
+	mf, err := convert.GetModelFormat(tempDir)
 	if err != nil {
 		return nil, err
 	}
 
-	params, err := mf.GetParams(tempdir)
+	params, err := mf.GetParams(tempDir)
 	if err != nil {
 		return nil, err
 	}
 
-	mArch, err := mf.GetModelArch("", tempdir, params)
+	mArch, err := mf.GetModelArch("", tempDir, params)
 	if err != nil {
 		return nil, err
 	}
@@ -150,7 +169,7 @@ func parseFromZipFile(_ context.Context, file *os.File, digest string, fn func(a
 
 	// TODO(mxyng): this should write directly into a layer
 	// e.g. NewLayer(arch.Reader(), "application/vnd.ollama.image.model")
-	temp, err := os.CreateTemp(tempdir, "fp16")
+	temp, err := os.CreateTemp(tempDir, "fp16")
 	if err != nil {
 		return nil, err
 	}

server/model_test.go

@@ -0,0 +1,92 @@
+package server
+
+import (
+	"archive/zip"
+	"bytes"
+	"io"
+	"os"
+	"path/filepath"
+	"slices"
+	"testing"
+
+	"github.com/ollama/ollama/api"
+)
+
+func createZipFile(t *testing.T, name string) *os.File {
+	t.Helper()
+
+	f, err := os.CreateTemp(t.TempDir(), "")
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	zf := zip.NewWriter(f)
+	defer zf.Close()
+
+	zh, err := zf.CreateHeader(&zip.FileHeader{Name: name})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	if _, err := io.Copy(zh, bytes.NewReader([]byte(""))); err != nil {
+		t.Fatal(err)
+	}
+
+	return f
+}
+
+func TestExtractFromZipFile(t *testing.T) {
+	cases := []struct {
+		name   string
+		expect []string
+	}{
+		{
+			name:   "good",
+			expect: []string{"good"},
+		},
+		{
+			name: filepath.Join("..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "bad"),
+		},
+	}
+
+	for _, tt := range cases {
+		t.Run(tt.name, func(t *testing.T) {
+			f := createZipFile(t, tt.name)
+			defer f.Close()
+
+			tempDir := t.TempDir()
+			if err := extractFromZipFile(tempDir, f, func(api.ProgressResponse) {}); err != nil {
+				t.Fatal(err)
+			}
+
+			var matches []string
+			if err := filepath.Walk(tempDir, func(p string, fi os.FileInfo, err error) error {
+				if err != nil {
+					return err
+				}
+
+				if !fi.IsDir() {
+					matches = append(matches, p)
+				}
+
+				return nil
+			}); err != nil {
+				t.Fatal(err)
+			}
+
+			var actual []string
+			for _, match := range matches {
+				rel, err := filepath.Rel(tempDir, match)
+				if err != nil {
+					t.Error(err)
+				}
+
+				actual = append(actual, rel)
+			}
+
+			if !slices.Equal(actual, tt.expect) {
+				t.Fatalf("expected %d files, got %d", len(tt.expect), len(matches))
+			}
+		})
+	}
+}