update patch

write-host in powershell writes directly to the console and will not be picked
up by a pipe.  Echo, or write-output will.

.github/workflows/release.yaml

@@ -274,7 +274,134 @@ jobs:
           path: dist/deps/*
 
 
-  # Import the prior generation steps and build the final windows assets
+  # windows arm64 generate, go build, and zip file (no installer)
+  # Output of this build is aggregated into the final x86 build
+  # for a unified windows installer
+  windows-arm64:
+    runs-on: windows-arm64
+    environment: release
+    env:
+      KEY_CONTAINER: ${{ vars.KEY_CONTAINER }}
+    steps:
+      # The current Windows arm64 beta image has effectively zero dev tools installed...
+      - name: Install git and gzip
+        run: |
+          Set-ExecutionPolicy Bypass -Scope Process -Force
+          [System.Net.ServicePointManager]::SecurityProtocol = [System.Net.ServicePointManager]::SecurityProtocol -bor 3072
+          iex ((New-Object System.Net.WebClient).DownloadString('https://community.chocolatey.org/install.ps1'))
+          choco install -y --no-progress git gzip
+          echo "C:\Program Files\Git\cmd" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "C:\ProgramData\chocolatey\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+      - name: Install Visual Studio 2022
+        run: |
+          $components = @(
+            "Microsoft.VisualStudio.Component.CoreEditor",
+            "Microsoft.VisualStudio.Workload.CoreEditor",
+            "Microsoft.VisualStudio.Component.Roslyn.Compiler",
+            "Microsoft.Component.MSBuild",
+            "Microsoft.VisualStudio.Component.TextTemplating",
+            "Microsoft.VisualStudio.Component.Debugger.JustInTime",
+            "Microsoft.VisualStudio.Component.VC.CoreIde",
+            "Microsoft.VisualStudio.Component.VC.Tools.x86.x64",
+            "Microsoft.VisualStudio.Component.Windows11SDK.22621",
+            "Microsoft.VisualStudio.Component.VC.Tools.ARM64EC",
+            "Microsoft.VisualStudio.Component.VC.Tools.ARM64",
+            "Microsoft.VisualStudio.Component.VC.ATL",
+            "Microsoft.VisualStudio.Component.VC.ATL.ARM64",
+            "Microsoft.VisualStudio.Component.Graphics",
+            "Microsoft.VisualStudio.Component.VC.Redist.14.Latest",
+            "Microsoft.VisualStudio.ComponentGroup.NativeDesktop.Core",
+            "Microsoft.VisualStudio.Component.Windows11Sdk.WindowsPerformanceToolkit",
+            "Microsoft.VisualStudio.Component.CppBuildInsights",
+            "Microsoft.VisualStudio.Component.VC.DiagnosticTools",
+            "Microsoft.VisualStudio.ComponentGroup.WebToolsExtensions.CMake",
+            "Microsoft.VisualStudio.Component.VC.CMake.Project",
+            "Microsoft.VisualStudio.Component.VC.ASAN",
+            "Microsoft.VisualStudio.Component.Vcpkg",
+            "Microsoft.VisualStudio.Workload.NativeDesktop"
+          )
+          $config = @{
+                "version" = "1.0"
+                "components"  = $components
+                "extensions"  = @()
+            }
+          $configPath = "${env:RUNNER_TEMP}\vsconfig"
+          $config | ConvertTo-Json | Out-File -FilePath $configPath
+          $bootstrapperFilePath = "${env:RUNNER_TEMP}\vs_community.exe"
+          write-host "Downloading Visual Studio 2022"
+          Invoke-WebRequest -Uri "https://aka.ms/vs/17/release/vs_community.exe" -outfile $bootstrapperFilePath
+          $bootstrapperArgumentList = ('/c', $bootstrapperFilePath, '--config', $configPath, '--quiet', '--wait' )
+          write-host "Installing Visual Studio 2022"
+          $process = Start-Process -FilePath cmd.exe -ArgumentList $bootstrapperArgumentList -Wait -PassThru
+          $exitCode = $process.ExitCode
+          write-host $exitCode
+      # pacman in mingw/msys2 is ~broken on windows arm right now - hangs consistently during attempts to install
+      # so we'll use this alternative GCC binary
+      - name: Install llvm-mingw GCC
+        run: |
+          $gcc_url="https://github.com/mstorsjo/llvm-mingw/releases/download/20240619/llvm-mingw-20240619-ucrt-aarch64.zip"
+          write-host "Downloading llvm-mingw"
+          Invoke-WebRequest -Uri "${gcc_url}" -OutFile "${env:RUNNER_TEMP}\gcc.zip"
+          write-host "Unpacking llvm-mingw"
+          expand-archive -path "${env:RUNNER_TEMP}\gcc.zip" -destinationpath "c:\"
+          mv c:\llvm-mingw-* c:\llvm-mingw
+          echo "c:\llvm-mingw\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+      - name: Verify GCC
+        run: |
+          echo $env:PATH
+          gcc --version
+      - uses: actions/checkout@v4
+      - name: Set Version
+        run: |
+          $ver=${env:GITHUB_REF_NAME}.trim("v")
+          echo VERSION=$ver | Out-File -FilePath ${env:GITHUB_ENV} -Encoding utf8 -Append
+      - uses: 'google-github-actions/auth@v2'
+        with:
+          project_id: 'ollama'
+          credentials_json: '${{ secrets.GOOGLE_SIGNING_CREDENTIALS }}'
+      - run: echo "${{ vars.OLLAMA_CERT }}" | Out-File -FilePath ollama_inc.crt -Encoding utf8
+      - name: install Windows SDK 8.1 to get signtool
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading SDK"
+          Invoke-WebRequest -Uri "https://go.microsoft.com/fwlink/p/?LinkId=323507" -OutFile "${env:RUNNER_TEMP}\sdksetup.exe"
+          Start-Process "${env:RUNNER_TEMP}\sdksetup.exe" -ArgumentList @("/q") -NoNewWindow -Wait
+          write-host "Win SDK 8.1 installed"
+          gci -path 'C:\Program Files (x86)\Windows Kits\' -r -fi 'signtool.exe'
+      - name: install signing plugin
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading plugin"
+          Invoke-WebRequest -Uri "https://github.com/GoogleCloudPlatform/kms-integrations/releases/download/cng-v1.0/kmscng-1.0-windows-amd64.zip" -OutFile "${env:RUNNER_TEMP}\plugin.zip"
+          Expand-Archive -Path "${env:RUNNER_TEMP}\plugin.zip" -DestinationPath ${env:RUNNER_TEMP}\plugin\
+          write-host "Installing plugin"
+          & "${env:RUNNER_TEMP}\plugin\*\kmscng.msi" /quiet
+          write-host "plugin installed"
+      - uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+          cache: true
+      - run: go get ./...
+      - run: |
+          $gopath=(get-command go).source | split-path -parent
+          $gccpath=(get-command gcc).source | split-path -parent
+          & "C:\Program Files\Microsoft Visual Studio\2022\Community\Common7\Tools\Launch-VsDevShell.ps1"
+          cd $env:GITHUB_WORKSPACE
+          $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
+          $env:PATH="$gopath;$gccpath;$env:PATH;C:\Program Files\Microsoft Visual Studio\2022\Community\Common7\IDE\CommonExtensions\Microsoft\CMake\CMake\bin"
+          echo $env:PATH
+          $env:ARCH="arm64"
+          .\scripts\build_windows.ps1 buildOllama buildApp gatherDependencies distZip
+        name: 'Windows Build'
+      - uses: actions/upload-artifact@v4
+        with:
+          name: windows-arm64
+          path: |
+            dist/windows-arm64/**
+            dist/windows-arm64-app.exe
+            dist/ollama-windows-arm64.zip
+
+  # Import the prior generation steps plus the full arm64 build, and build the final windows assets
   build-windows:
     environment: release
     runs-on: windows
@@ -282,6 +409,7 @@ jobs:
       - generate-windows-cuda
       - generate-windows-rocm
       - generate-windows-cpu
+      - windows-arm64
     env:
       KEY_CONTAINER: ${{ vars.KEY_CONTAINER }}
     steps:
@@ -339,6 +467,10 @@ jobs:
       - uses: actions/download-artifact@v4
         with:
           name: generate-windows-rocm
+      - uses: actions/download-artifact@v4
+        with:
+          name: windows-arm64
+          path: dist
       - run: dir build
       - run: |
           $gopath=(get-command go).source | split-path -parent

README.md

@@ -35,10 +35,10 @@ The official [Ollama Docker image](https://hub.docker.com/r/ollama/ollama) `olla
 
 ## Quickstart
 
-To run and chat with [Llama 3.1](https://ollama.com/library/llama3.1):
+To run and chat with [Llama 3.2](https://ollama.com/library/llama3.2):
 
 ```
-ollama run llama3.1
+ollama run llama3.2
 ```
 
 ## Model library
@@ -49,6 +49,8 @@ Here are some example models that can be downloaded:
 
 | Model              | Parameters | Size  | Download                       |
 | ------------------ | ---------- | ----- | ------------------------------ |
+| Llama 3.2          | 3B         | 2.0GB | `ollama run llama3.2`          |
+| Llama 3.2          | 1B         | 1.3GB | `ollama run llama3.2:1b`       |
 | Llama 3.1          | 8B         | 4.7GB | `ollama run llama3.1`          |
 | Llama 3.1          | 70B        | 40GB  | `ollama run llama3.1:70b`      |
 | Llama 3.1          | 405B       | 231GB | `ollama run llama3.1:405b`     |
@@ -99,16 +101,16 @@ See the [guide](docs/import.md) on importing models for more information.
 
 ### Customize a prompt
 
-Models from the Ollama library can be customized with a prompt. For example, to customize the `llama3.1` model:
+Models from the Ollama library can be customized with a prompt. For example, to customize the `llama3.2` model:
 
 ```
-ollama pull llama3.1
+ollama pull llama3.2
 ```
 
 Create a `Modelfile`:
 
 ```
-FROM llama3.1
+FROM llama3.2
 
 # set the temperature to 1 [higher is more creative, lower is more coherent]
 PARAMETER temperature 1
@@ -143,7 +145,7 @@ ollama create mymodel -f ./Modelfile
 ### Pull a model
 
 ```
-ollama pull llama3.1
+ollama pull llama3.2
 ```
 
 > This command can also be used to update a local model. Only the diff will be pulled.
@@ -151,13 +153,13 @@ ollama pull llama3.1
 ### Remove a model
 
 ```
-ollama rm llama3.1
+ollama rm llama3.2
 ```
 
 ### Copy a model
 
 ```
-ollama cp llama3.1 my-model
+ollama cp llama3.2 my-model
 ```
 
 ### Multiline input
@@ -181,14 +183,14 @@ The image features a yellow smiley face, which is likely the central focus of th
 ### Pass the prompt as an argument
 
 ```
-$ ollama run llama3.1 "Summarize this file: $(cat README.md)"
+$ ollama run llama3.2 "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 llama3.1
+ollama show llama3.2
 ```
 
 ### List models on your computer
@@ -197,6 +199,18 @@ ollama show llama3.1
 ollama list
 ```
 
+### List which models are currently loaded
+
+```
+ollama ps
+```
+
+### Stop a model which is currently running
+
+```
+ollama stop llama3.2
+```
+
 ### Start Ollama
 
 `ollama serve` is used when you want to start ollama without running the desktop application.
@@ -216,7 +230,7 @@ Next, start the server:
 Finally, in a separate shell, run a model:
 
 ```
-./ollama run llama3.1
+./ollama run llama3.2
 ```
 
 ## REST API
@@ -227,7 +241,7 @@ Ollama has a REST API for running and managing models.
 
 ```
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3.1",
+  "model": "llama3.2",
   "prompt":"Why is the sky blue?"
 }'
 ```
@@ -236,7 +250,7 @@ curl http://localhost:11434/api/generate -d '{
 
 ```
 curl http://localhost:11434/api/chat -d '{
-  "model": "llama3.1",
+  "model": "llama3.2",
   "messages": [
     { "role": "user", "content": "why is the sky blue?" }
   ]
@@ -313,6 +327,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [ConfiChat](https://github.com/1runeberg/confichat) (Lightweight, standalone, multi-platform, and privacy focused LLM chat interface with optional encryption)
 - [Archyve](https://github.com/nickthecook/archyve) (RAG-enabling document library)
 - [crewAI with Mesop](https://github.com/rapidarchitect/ollama-crew-mesop) (Mesop Web Interface to run crewAI with Ollama)
+- [LLMChat](https://github.com/trendy-design/llmchat) (Privacy focused, 100% local, intuitive all-in-one chat interface)
 
 ### Terminal
 
@@ -365,7 +380,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [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)
+- [LlamaIndex](https://docs.llamaindex.ai/en/stable/examples/llm/ollama/) and [LlamaIndexTS](https://ts.llamaindex.ai/modules/llms/available_llms/ollama)
 - [LiteLLM](https://github.com/BerriAI/litellm)
 - [OllamaFarm for Go](https://github.com/presbrey/ollamafarm)
 - [OllamaSharp for .NET](https://github.com/awaescher/OllamaSharp)

app/ollama.iss

@@ -28,8 +28,8 @@ AppPublisher={#MyAppPublisher}
 AppPublisherURL={#MyAppURL}
 AppSupportURL={#MyAppURL}
 AppUpdatesURL={#MyAppURL}
-ArchitecturesAllowed=x64 arm64
-ArchitecturesInstallIn64BitMode=x64 arm64
+ArchitecturesAllowed=x64compatible arm64
+ArchitecturesInstallIn64BitMode=x64compatible arm64
 DefaultDirName={localappdata}\Programs\{#MyAppName}
 DefaultGroupName={#MyAppName}
 DisableProgramGroupPage=yes
@@ -48,6 +48,7 @@ OutputDir=..\dist\
 SetupLogging=yes
 CloseApplications=yes
 RestartApplications=no
+RestartIfNeededByRun=no
 
 ; https://jrsoftware.org/ishelp/index.php?topic=setup_wizardimagefile
 WizardSmallImageFile=.\assets\setup.bmp
@@ -86,12 +87,21 @@ Name: "english"; MessagesFile: "compiler:Default.isl"
 DialogFontSize=12
 
 [Files]
-Source: ".\app.exe"; DestDir: "{app}"; DestName: "{#MyAppExeName}" ; Flags: ignoreversion 64bit
-Source: "..\ollama.exe"; DestDir: "{app}"; Flags: ignoreversion 64bit
-Source: "..\dist\windows-{#ARCH}\lib\ollama\runners\*"; DestDir: "{app}\lib\ollama\runners"; Flags: ignoreversion 64bit recursesubdirs
+#if DirExists("..\dist\windows-amd64")
+Source: "..\dist\windows-amd64-app.exe"; DestDir: "{app}"; DestName: "{#MyAppExeName}" ;Check: not IsArm64();  Flags: ignoreversion 64bit
+Source: "..\dist\windows-amd64\ollama.exe"; DestDir: "{app}"; Check: not IsArm64(); Flags: ignoreversion 64bit
+Source: "..\dist\windows-amd64\lib\ollama\*"; DestDir: "{app}\lib\ollama\"; Check: not IsArm64(); Flags: ignoreversion 64bit recursesubdirs
+#endif
+
+#if DirExists("..\dist\windows-arm64")
+Source: "..\dist\windows-arm64\vc_redist.arm64.exe"; DestDir: "{tmp}"; Check: IsArm64() and vc_redist_needed(); Flags: deleteafterinstall
+Source: "..\dist\windows-arm64-app.exe"; DestDir: "{app}"; DestName: "{#MyAppExeName}" ;Check: IsArm64();  Flags: ignoreversion 64bit
+Source: "..\dist\windows-arm64\ollama.exe"; DestDir: "{app}"; Check: IsArm64(); Flags: ignoreversion 64bit
+Source: "..\dist\windows-arm64\lib\ollama\*"; DestDir: "{app}\lib\ollama\"; Check: IsArm64(); Flags: ignoreversion 64bit recursesubdirs
+#endif
+
 Source: "..\dist\ollama_welcome.ps1"; DestDir: "{app}"; Flags: ignoreversion
 Source: ".\assets\app.ico"; DestDir: "{app}"; Flags: ignoreversion
-Source: "..\dist\windows-amd64\lib\ollama\*"; DestDir: "{app}\lib\ollama\"; Flags: ignoreversion recursesubdirs
 
 [Icons]
 Name: "{group}\{#MyAppName}"; Filename: "{app}\{#MyAppExeName}"; IconFilename: "{app}\app.ico"
@@ -99,6 +109,9 @@ Name: "{userstartup}\{#MyAppName}"; Filename: "{app}\{#MyAppExeName}"; IconFilen
 Name: "{userprograms}\{#MyAppName}"; Filename: "{app}\{#MyAppExeName}"; IconFilename: "{app}\app.ico"
 
 [Run]
+#if DirExists("..\dist\windows-arm64")
+Filename: "{tmp}\vc_redist.arm64.exe"; Parameters: "/install /passive /norestart"; Check: IsArm64() and vc_redist_needed(); StatusMsg: "Installing VC++ Redistributables..."; Flags: waituntilterminated
+#endif
 Filename: "{cmd}"; Parameters: "/C set PATH={app};%PATH% & ""{app}\{#MyAppExeName}"""; Flags: postinstall nowait runhidden
 
 [UninstallRun]
@@ -129,7 +142,7 @@ SetupAppRunningError=Another Ollama installer is running.%n%nPlease cancel or fi
 
 
 ;FinishedHeadingLabel=Run your first model
-;FinishedLabel=%nRun this command in a PowerShell or cmd terminal.%n%n%n    ollama run llama3.1
+;FinishedLabel=%nRun this command in a PowerShell or cmd terminal.%n%n%n    ollama run llama3.2
 ;ClickFinish=%n
 
 [Registry]
@@ -154,3 +167,39 @@ begin
   { Pos() returns 0 if not found }
   Result := Pos(';' + ExpandConstant(Param) + ';', ';' + OrigPath + ';') = 0;
 end;
+
+{ --- VC Runtime libraries discovery code - Only install vc_redist if it isn't already installed ----- }
+const VCRTL_MIN_V1 = 14;
+const VCRTL_MIN_V2 = 40;
+const VCRTL_MIN_V3 = 33807;
+const VCRTL_MIN_V4 = 0;
+
+ // check if the minimum required vc redist is installed (by looking the registry)
+function vc_redist_needed (): Boolean;
+var
+  sRegKey: string;
+  v1: Cardinal;
+  v2: Cardinal;
+  v3: Cardinal;
+  v4: Cardinal;
+begin
+  sRegKey := 'SOFTWARE\WOW6432Node\Microsoft\VisualStudio\14.0\VC\Runtimes\arm64';
+  if (RegQueryDWordValue (HKEY_LOCAL_MACHINE, sRegKey, 'Major', v1)  and
+      RegQueryDWordValue (HKEY_LOCAL_MACHINE, sRegKey, 'Minor', v2) and
+      RegQueryDWordValue (HKEY_LOCAL_MACHINE, sRegKey, 'Bld', v3) and
+      RegQueryDWordValue (HKEY_LOCAL_MACHINE, sRegKey, 'RBld', v4)) then
+  begin
+    Log ('VC Redist version: ' + IntToStr (v1) +
+        '.' + IntToStr (v2) + '.' + IntToStr (v3) +
+        '.' + IntToStr (v4));
+    { Version info was found. Return true if later or equal to our
+       minimal required version RTL_MIN_Vx }
+    Result := not (
+        (v1 > VCRTL_MIN_V1) or ((v1 = VCRTL_MIN_V1) and
+         ((v2 > VCRTL_MIN_V2) or ((v2 = VCRTL_MIN_V2) and
+          ((v3 > VCRTL_MIN_V3) or ((v3 = VCRTL_MIN_V3) and
+           (v4 >= VCRTL_MIN_V4)))))));
+  end
+  else
+    Result := TRUE;
+end;

app/ollama_welcome.ps1

@@ -4,5 +4,5 @@ write-host "Welcome to Ollama!"
 write-host ""
 write-host "Run your first model:"
 write-host ""
-write-host "`tollama run llama3.1"
+write-host "`tollama run llama3.2"
 write-host ""

cmd/interactive.go

@@ -30,11 +30,6 @@ const (
 	MultilineSystem
 )
 
-const (
-	scannerPrompt    = ">>> "
-	scannerAltPrompt = "... "
-)
-
 func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 	usage := func() {
 		fmt.Fprintln(os.Stderr, "Available Commands:")
@@ -116,8 +111,8 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 	}
 
 	scanner, err := readline.New(readline.Prompt{
-		Prompt:         scannerPrompt,
-		AltPrompt:      scannerAltPrompt,
+		Prompt:         ">>> ",
+		AltPrompt:      "... ",
 		Placeholder:    "Send a message (/? for help)",
 		AltPlaceholder: `Use """ to end multi-line input`,
 	})
@@ -149,11 +144,6 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 			scanner.Prompt.UseAlt = false
 			sb.Reset()
 
-			continue
-		case errors.Is(err, readline.ErrNewLineDetected):
-			sb.WriteString(line)
-			fmt.Fprintln(&sb)
-			scanner.Prompt.Prompt = scannerAltPrompt
 			continue
 		case err != nil:
 			return err
@@ -179,7 +169,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 
 			multiline = MultilineNone
 			scanner.Prompt.UseAlt = false
-		case strings.HasPrefix(line, `"""`) && !scanner.Pasting:
+		case strings.HasPrefix(line, `"""`):
 			line := strings.TrimPrefix(line, `"""`)
 			line, ok := strings.CutSuffix(line, `"""`)
 			sb.WriteString(line)
@@ -443,7 +433,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 			sb.WriteString(line)
 		}
 
-		if sb.Len() > 0 && strings.TrimSpace(sb.String()) != "" && multiline == MultilineNone {
+		if sb.Len() > 0 && multiline == MultilineNone {
 			newMessage := api.Message{Role: "user", Content: sb.String()}
 
 			if opts.MultiModal {
@@ -474,7 +464,6 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 			}
 
 			sb.Reset()
-			scanner.Prompt.Prompt = scannerPrompt
 		}
 	}
 }

docs/api.md

@@ -69,7 +69,7 @@ Enable JSON mode by setting the `format` parameter to `json`. This will structur
 
 ```shell
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3.1",
+  "model": "llama3.2",
   "prompt": "Why is the sky blue?"
 }'
 ```
@@ -80,7 +80,7 @@ A stream of JSON objects is returned:
 
 ```json
 {
-  "model": "llama3.1",
+  "model": "llama3.2",
   "created_at": "2023-08-04T08:52:19.385406455-07:00",
   "response": "The",
   "done": false
@@ -102,7 +102,7 @@ To calculate how fast the response is generated in tokens per second (token/s),
 
 ```json
 {
-  "model": "llama3.1",
+  "model": "llama3.2",
   "created_at": "2023-08-04T19:22:45.499127Z",
   "response": "",
   "done": true,
@@ -124,7 +124,7 @@ A response can be received in one reply when streaming is off.
 
 ```shell
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3.1",
+  "model": "llama3.2",
   "prompt": "Why is the sky blue?",
   "stream": false
 }'
@@ -136,7 +136,7 @@ If `stream` is set to `false`, the response will be a single JSON object:
 
 ```json
 {
-  "model": "llama3.1",
+  "model": "llama3.2",
   "created_at": "2023-08-04T19:22:45.499127Z",
   "response": "The sky is blue because it is the color of the sky.",
   "done": true,
@@ -194,7 +194,7 @@ curl http://localhost:11434/api/generate -d '{
 
 ```shell
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3.1",
+  "model": "llama3.2",
   "prompt": "What color is the sky at different times of the day? Respond using JSON",
   "format": "json",
   "stream": false
@@ -205,7 +205,7 @@ curl http://localhost:11434/api/generate -d '{
 
 ```json
 {
-  "model": "llama3.1",
+  "model": "llama3.2",
   "created_at": "2023-11-09T21:07:55.186497Z",
   "response": "{\n\"morning\": {\n\"color\": \"blue\"\n},\n\"noon\": {\n\"color\": \"blue-gray\"\n},\n\"afternoon\": {\n\"color\": \"warm gray\"\n},\n\"evening\": {\n\"color\": \"orange\"\n}\n}\n",
   "done": true,
@@ -327,7 +327,7 @@ If you want to set custom options for the model at runtime rather than in the Mo
 
 ```shell
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3.1",
+  "model": "llama3.2",
   "prompt": "Why is the sky blue?",
   "stream": false,
   "options": {
@@ -368,7 +368,7 @@ curl http://localhost:11434/api/generate -d '{
 
 ```json
 {
-  "model": "llama3.1",
+  "model": "llama3.2",
   "created_at": "2023-08-04T19:22:45.499127Z",
   "response": "The sky is blue because it is the color of the sky.",
   "done": true,
@@ -390,7 +390,7 @@ If an empty prompt is provided, the model will be loaded into memory.
 
 ```shell
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3.1"
+  "model": "llama3.2"
 }'
 ```
 
@@ -400,13 +400,40 @@ A single JSON object is returned:
 
 ```json
 {
-  "model": "llama3.1",
+  "model": "llama3.2",
   "created_at": "2023-12-18T19:52:07.071755Z",
   "response": "",
   "done": true
 }
 ```
 
+#### Unload a model
+
+If an empty prompt is provided and the `keep_alive` parameter is set to `0`, a model will be unloaded from memory.
+
+##### Request
+
+```shell
+curl http://localhost:11434/api/generate -d '{
+  "model": "llama3.2",
+  "keep_alive": 0
+}'
+```
+
+##### Response
+
+A single JSON object is returned:
+
+```json
+{
+  "model": "llama3.2",
+  "created_at": "2024-09-12T03:54:03.516566Z",
+  "response": "",
+  "done": true,
+  "done_reason": "unload"
+}
+```
+
 ## Generate a chat completion
 
 ```shell
@@ -445,7 +472,7 @@ Send a chat message with a streaming response.
 
 ```shell
 curl http://localhost:11434/api/chat -d '{
-  "model": "llama3.1",
+  "model": "llama3.2",
   "messages": [
     {
       "role": "user",
@@ -461,7 +488,7 @@ A stream of JSON objects is returned:
 
 ```json
 {
-  "model": "llama3.1",
+  "model": "llama3.2",
   "created_at": "2023-08-04T08:52:19.385406455-07:00",
   "message": {
     "role": "assistant",
@@ -476,7 +503,7 @@ Final response:
 
 ```json
 {
-  "model": "llama3.1",
+  "model": "llama3.2",
   "created_at": "2023-08-04T19:22:45.499127Z",
   "done": true,
   "total_duration": 4883583458,
@@ -494,7 +521,7 @@ Final response:
 
 ```shell
 curl http://localhost:11434/api/chat -d '{
-  "model": "llama3.1",
+  "model": "llama3.2",
   "messages": [
     {
       "role": "user",
@@ -509,7 +536,7 @@ curl http://localhost:11434/api/chat -d '{
 
 ```json
 {
-  "model": "llama3.1",
+  "model": "llama3.2",
   "created_at": "2023-12-12T14:13:43.416799Z",
   "message": {
     "role": "assistant",
@@ -533,7 +560,7 @@ Send a chat message with a conversation history. You can use this same approach
 
 ```shell
 curl http://localhost:11434/api/chat -d '{
-  "model": "llama3.1",
+  "model": "llama3.2",
   "messages": [
     {
       "role": "user",
@@ -557,7 +584,7 @@ A stream of JSON objects is returned:
 
 ```json
 {
-  "model": "llama3.1",
+  "model": "llama3.2",
   "created_at": "2023-08-04T08:52:19.385406455-07:00",
   "message": {
     "role": "assistant",
@@ -571,7 +598,7 @@ Final response:
 
 ```json
 {
-  "model": "llama3.1",
+  "model": "llama3.2",
   "created_at": "2023-08-04T19:22:45.499127Z",
   "done": true,
   "total_duration": 8113331500,
@@ -629,7 +656,7 @@ curl http://localhost:11434/api/chat -d '{
 
 ```shell
 curl http://localhost:11434/api/chat -d '{
-  "model": "llama3.1",
+  "model": "llama3.2",
   "messages": [
     {
       "role": "user",
@@ -647,7 +674,7 @@ curl http://localhost:11434/api/chat -d '{
 
 ```json
 {
-  "model": "llama3.1",
+  "model": "llama3.2",
   "created_at": "2023-12-12T14:13:43.416799Z",
   "message": {
     "role": "assistant",
@@ -669,7 +696,7 @@ curl http://localhost:11434/api/chat -d '{
 
 ```
 curl http://localhost:11434/api/chat -d '{
-  "model": "llama3.1",
+  "model": "llama3.2",
   "messages": [
     {
       "role": "user",
@@ -708,7 +735,7 @@ curl http://localhost:11434/api/chat -d '{
 
 ```json
 {
-  "model": "llama3.1",
+  "model": "llama3.2",
   "created_at": "2024-07-22T20:33:28.123648Z",
   "message": {
     "role": "assistant",
@@ -736,6 +763,64 @@ curl http://localhost:11434/api/chat -d '{
 }
 ```
 
+#### Load a model
+
+If the messages array is empty, the model will be loaded into memory.
+
+##### Request
+
+```
+curl http://localhost:11434/api/chat -d '{
+  "model": "llama3.2",
+  "messages": []
+}'
+```
+
+##### Response
+```json
+{
+  "model": "llama3.2",
+  "created_at":"2024-09-12T21:17:29.110811Z",
+  "message": {
+    "role": "assistant",
+    "content": ""
+  },
+  "done_reason": "load",
+  "done": true
+}
+```
+
+#### Unload a model
+
+If the messages array is empty and the `keep_alive` parameter is set to `0`, a model will be unloaded from memory.
+
+##### Request
+
+```
+curl http://localhost:11434/api/chat -d '{
+  "model": "llama3.2",
+  "messages": [],
+  "keep_alive": 0
+}'
+```
+
+##### Response
+
+A single JSON object is returned:
+
+```json
+{
+  "model": "llama3.2",
+  "created_at":"2024-09-12T21:33:17.547535Z",
+  "message": {
+    "role": "assistant",
+    "content": ""
+  },
+  "done_reason": "unload",
+  "done": true
+}
+```
+
 ## Create a Model
 
 ```shell
@@ -904,7 +989,7 @@ Show information about a model including details, modelfile, template, parameter
 
 ```shell
 curl http://localhost:11434/api/show -d '{
-  "name": "llama3.1"
+  "name": "llama3.2"
 }'
 ```
 
@@ -965,7 +1050,7 @@ Copy a model. Creates a model with another name from an existing model.
 
 ```shell
 curl http://localhost:11434/api/copy -d '{
-  "source": "llama3.1",
+  "source": "llama3.2",
   "destination": "llama3-backup"
 }'
 ```
@@ -1020,7 +1105,7 @@ Download a model from the ollama library. Cancelled pulls are resumed from where
 
 ```shell
 curl http://localhost:11434/api/pull -d '{
-  "name": "llama3.1"
+  "name": "llama3.2"
 }'
 ```
 

docs/development.md

@@ -148,3 +148,22 @@ In addition to the common Windows development tools described above, install AMD
 - [Strawberry Perl](https://strawberryperl.com/)
 
 Lastly, add `ninja.exe` included with MSVC to the system path (e.g. `C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\Common7\IDE\CommonExtensions\Microsoft\CMake\Ninja`).
+
+#### Windows arm64
+
+The default `Developer PowerShell for VS 2022` may default to x86 which is not what you want.  To ensure you get an arm64 development environment, start a plain PowerShell terminal and run:
+
+```powershell
+import-module 'C:\\Program Files\\Microsoft Visual Studio\\2022\\Community\\Common7\\Tools\\Microsoft.VisualStudio.DevShell.dll'
+Enter-VsDevShell -Arch arm64 -vsinstallpath 'C:\\Program Files\\Microsoft Visual Studio\\2022\\Community' -skipautomaticlocation
+```
+
+You can confirm with `write-host $env:VSCMD_ARG_TGT_ARCH`
+
+Follow the instructions at https://www.msys2.org/wiki/arm64/ to set up an arm64 msys2 environment.  Ollama requires gcc and mingw32-make to compile, which is not currently available on Windows arm64, but a gcc compatibility adapter is available via `mingw-w64-clang-aarch64-gcc-compat`. At a minimum you will need to install the following:
+
+```
+pacman -S mingw-w64-clang-aarch64-clang mingw-w64-clang-aarch64-gcc-compat mingw-w64-clang-aarch64-make make
+```
+
+You will need to ensure your PATH includes go, cmake, gcc and clang mingw32-make to build ollama from source. (typically `C:\msys64\clangarm64\bin\`)

docs/docker.md

@@ -63,7 +63,7 @@ docker run -d --device /dev/kfd --device /dev/dri -v ollama:/root/.ollama -p 114
 Now you can run a model:
 
 ```
-docker exec -it ollama ollama run llama3.1
+docker exec -it ollama ollama run llama3.2
 ```
 
 ### Try different models

docs/faq.md

@@ -32,7 +32,7 @@ When using the API, specify the `num_ctx` parameter:
 
 ```shell
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3.1",
+  "model": "llama3.2",
   "prompt": "Why is the sky blue?",
   "options": {
     "num_ctx": 4096
@@ -232,14 +232,18 @@ curl http://localhost:11434/api/chat -d '{"model": "mistral"}'
 
 To preload a model using the CLI, use the command:
 ```shell
-ollama run llama3.1 ""
+ollama run llama3.2 ""
 ```
 
 ## How do I keep a model loaded in memory or make it unload immediately?
 
-By default models are kept in memory for 5 minutes before being unloaded. This allows for quicker response times if you are making numerous requests to the LLM. You may, however, want to free up the memory before the 5 minutes have elapsed or keep the model loaded indefinitely. Use the `keep_alive` parameter with either the `/api/generate` and `/api/chat` API endpoints to control how long the model is left in memory.
+By default models are kept in memory for 5 minutes before being unloaded. This allows for quicker response times if you're making numerous requests to the LLM. If you want to immediately unload a model from memory, use the `ollama stop` command:
 
-The `keep_alive` parameter can be set to:
+```shell
+ollama stop llama3.2
+```
+
+If you're using the API, use the `keep_alive` parameter with the `/api/generate` and `/api/chat` endpoints to set the amount of time that a model stays in memory. The `keep_alive` parameter can be set to:
 * a duration string (such as "10m" or "24h")
 * a number in seconds (such as 3600)
 * any negative number which will keep the model loaded in memory (e.g. -1 or "-1m")
@@ -247,17 +251,17 @@ The `keep_alive` parameter can be set to:
 
 For example, to preload a model and leave it in memory use:
 ```shell
-curl http://localhost:11434/api/generate -d '{"model": "llama3.1", "keep_alive": -1}'
+curl http://localhost:11434/api/generate -d '{"model": "llama3.2", "keep_alive": -1}'
 ```
 
 To unload the model and free up memory use:
 ```shell
-curl http://localhost:11434/api/generate -d '{"model": "llama3.1", "keep_alive": 0}'
+curl http://localhost:11434/api/generate -d '{"model": "llama3.2", "keep_alive": 0}'
 ```
 
-Alternatively, you can change the amount of time all models are loaded into memory by setting the `OLLAMA_KEEP_ALIVE` environment variable when starting the Ollama server. The `OLLAMA_KEEP_ALIVE` variable uses the same parameter types as the `keep_alive` parameter types mentioned above. Refer to section explaining [how to configure the Ollama server](#how-do-i-configure-ollama-server) to correctly set the environment variable.
+Alternatively, you can change the amount of time all models are loaded into memory by setting the `OLLAMA_KEEP_ALIVE` environment variable when starting the Ollama server. The `OLLAMA_KEEP_ALIVE` variable uses the same parameter types as the `keep_alive` parameter types mentioned above. Refer to the section explaining [how to configure the Ollama server](#how-do-i-configure-ollama-server) to correctly set the environment variable.
 
-If you wish to override the `OLLAMA_KEEP_ALIVE` setting, use the `keep_alive` API parameter with the `/api/generate` or `/api/chat` API endpoints.
+The `keep_alive` API parameter with the `/api/generate` and `/api/chat` API endpoints will override the `OLLAMA_KEEP_ALIVE` setting.
 
 ## How do I manage the maximum number of requests the Ollama server can queue?
 

docs/modelfile.md

@@ -50,7 +50,7 @@ INSTRUCTION arguments
 An example of a `Modelfile` creating a mario blueprint:
 
 ```modelfile
-FROM llama3.1
+FROM llama3.2
 # 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
@@ -72,10 +72,10 @@ More examples are available in the [examples directory](../examples).
 To view the Modelfile of a given model, use the `ollama show --modelfile` command.
 
   ```bash
-  > ollama show --modelfile llama3.1
+  > ollama show --modelfile llama3.2
   # Modelfile generated by "ollama show"
   # To build a new Modelfile based on this one, replace the FROM line with:
-  # FROM llama3.1:latest
+  # FROM llama3.2:latest
   FROM /Users/pdevine/.ollama/models/blobs/sha256-00e1317cbf74d901080d7100f57580ba8dd8de57203072dc6f668324ba545f29
   TEMPLATE """{{ if .System }}<|start_header_id|>system<|end_header_id|>
 
@@ -103,7 +103,7 @@ FROM <model name>:<tag>
 #### Build from existing model
 
 ```modelfile
-FROM llama3.1
+FROM llama3.2
 ```
 
 A list of available base models:

docs/openai.md

@@ -25,7 +25,7 @@ chat_completion = client.chat.completions.create(
             'content': 'Say this is a test',
         }
     ],
-    model='llama3.1',
+    model='llama3.2',
 )
 
 response = client.chat.completions.create(
@@ -46,13 +46,13 @@ response = client.chat.completions.create(
 )
 
 completion = client.completions.create(
-    model="llama3.1",
+    model="llama3.2",
     prompt="Say this is a test",
 )
 
 list_completion = client.models.list()
 
-model = client.models.retrieve("llama3.1")
+model = client.models.retrieve("llama3.2")
 
 embeddings = client.embeddings.create(
     model="all-minilm",
@@ -74,7 +74,7 @@ const openai = new OpenAI({
 
 const chatCompletion = await openai.chat.completions.create({
     messages: [{ role: 'user', content: 'Say this is a test' }],
-    model: 'llama3.1',
+    model: 'llama3.2',
 })
 
 const response = await openai.chat.completions.create({
@@ -94,13 +94,13 @@ const response = await openai.chat.completions.create({
 })
 
 const completion = await openai.completions.create({
-    model: "llama3.1",
+    model: "llama3.2",
     prompt: "Say this is a test.",
 })
 
 const listCompletion = await openai.models.list()
 
-const model = await openai.models.retrieve("llama3.1")
+const model = await openai.models.retrieve("llama3.2")
 
 const embedding = await openai.embeddings.create({
   model: "all-minilm",
@@ -114,7 +114,7 @@ const embedding = await openai.embeddings.create({
 curl http://localhost:11434/v1/chat/completions \
     -H "Content-Type: application/json" \
     -d '{
-        "model": "llama3.1",
+        "model": "llama3.2",
         "messages": [
             {
                 "role": "system",
@@ -154,13 +154,13 @@ curl http://localhost:11434/v1/chat/completions \
 curl http://localhost:11434/v1/completions \
     -H "Content-Type: application/json" \
     -d '{
-        "model": "llama3.1",
+        "model": "llama3.2",
         "prompt": "Say this is a test"
     }'
 
 curl http://localhost:11434/v1/models
 
-curl http://localhost:11434/v1/models/llama3.1
+curl http://localhost:11434/v1/models/llama3.2
 
 curl http://localhost:11434/v1/embeddings \
     -H "Content-Type: application/json" \
@@ -274,7 +274,7 @@ curl http://localhost:11434/v1/embeddings \
 Before using a model, pull it locally `ollama pull`:
 
 ```shell
-ollama pull llama3.1
+ollama pull llama3.2
 ```
 
 ### Default model names
@@ -282,7 +282,7 @@ ollama pull llama3.1
 For tooling that relies on default OpenAI model names such as `gpt-3.5-turbo`, use `ollama cp` to copy an existing model name to a temporary name:
 
 ```
-ollama cp llama3.1 gpt-3.5-turbo
+ollama cp llama3.2 gpt-3.5-turbo
 ```
 
 Afterwards, this new model name can be specified the `model` field:

docs/template.md

@@ -33,7 +33,7 @@ Omitting a template in these models puts the responsibility of correctly templat
 To add templates in your model, you'll need to add a `TEMPLATE` command to the Modelfile. Here's an example using Meta's Llama 3.
 
 ```dockerfile
-FROM llama3.1
+FROM llama3.2
 
 TEMPLATE """{{- if .System }}<|start_header_id|>system<|end_header_id|>
 

docs/tutorials/langchainjs.md

@@ -15,7 +15,7 @@ import { Ollama } from "@langchain/community/llms/ollama";
 
 const ollama = new Ollama({
   baseUrl: "http://localhost:11434",
-  model: "llama3.1",
+  model: "llama3.2",
 });
 
 const answer = await ollama.invoke(`why is the sky blue?`);
@@ -23,7 +23,7 @@ const answer = await ollama.invoke(`why is the sky blue?`);
 console.log(answer);
 ```
 
-That will get us the same thing as if we ran `ollama run llama3.1 "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 install **Cheerio** and build that part of the app.
+That will get us the same thing as if we ran `ollama run llama3.2 "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 install **Cheerio** and build that part of the app.
 
 ```bash
 npm install cheerio

docs/windows.md

@@ -29,7 +29,7 @@ Ollama uses unicode characters for progress indication, which may render as unkn
 
 Here's a quick example showing API access from `powershell`
 ```powershell
-(Invoke-WebRequest -method POST -Body '{"model":"llama3.1", "prompt":"Why is the sky blue?", "stream": false}' -uri http://localhost:11434/api/generate ).Content | ConvertFrom-json
+(Invoke-WebRequest -method POST -Body '{"model":"llama3.2", "prompt":"Why is the sky blue?", "stream": false}' -uri http://localhost:11434/api/generate ).Content | ConvertFrom-json
 ```
 
 ## Troubleshooting

examples/go-chat/main.go

@@ -35,7 +35,7 @@ func main() {
 
 	ctx := context.Background()
 	req := &api.ChatRequest{
-		Model:    "llama3.1",
+		Model:    "llama3.2",
 		Messages: messages,
 	}
 

examples/langchain-python-rag-document/README.md

@@ -4,10 +4,10 @@ This example provides an interface for asking questions to a PDF document.
 
 ## Setup
 
-1. Ensure you have the `llama3.1` model installed:
+1. Ensure you have the `llama3.2` model installed:
 
 ```
-ollama pull llama3.1
+ollama pull llama3.2
 ```
 
 2. Install the Python Requirements.

examples/langchain-python-rag-document/main.py

@@ -51,7 +51,7 @@ while True:
         template=template,
     )
 
-    llm = Ollama(model="llama3.1", callback_manager=CallbackManager([StreamingStdOutCallbackHandler()]))
+    llm = Ollama(model="llama3.2", callback_manager=CallbackManager([StreamingStdOutCallbackHandler()]))
     qa_chain = RetrievalQA.from_chain_type(
         llm,
         retriever=vectorstore.as_retriever(),

examples/langchain-python-rag-websummary/README.md

@@ -4,10 +4,10 @@ This example summarizes the website, [https://ollama.com/blog/run-llama2-uncenso
 
 ## Running the Example
 
-1. Ensure you have the `llama3.1` model installed:
+1. Ensure you have the `llama3.2` model installed:
 
    ```bash
-   ollama pull llama3.1
+   ollama pull llama3.2
    ```
 
 2. Install the Python Requirements.

examples/langchain-python-rag-websummary/main.py

@@ -5,7 +5,7 @@ from langchain.chains.summarize import load_summarize_chain
 loader = WebBaseLoader("https://ollama.com/blog/run-llama2-uncensored-locally")
 docs = loader.load()
 
-llm = Ollama(model="llama3.1")
+llm = Ollama(model="llama3.2")
 chain = load_summarize_chain(llm, chain_type="stuff")
 
 result = chain.invoke(docs)

examples/langchain-python-simple/README.md

@@ -4,10 +4,10 @@ This example is a basic "hello world" of using LangChain with Ollama.
 
 ## Running the Example
 
-1. Ensure you have the `llama3.1` model installed:
+1. Ensure you have the `llama3.2` model installed:
 
    ```bash
-   ollama pull llama3.1
+   ollama pull llama3.2
    ```
 
 2. Install the Python Requirements.

examples/langchain-python-simple/main.py

@@ -1,6 +1,6 @@
 from langchain.llms import Ollama
 
 input = input("What is your question?")
-llm = Ollama(model="llama3.1")
+llm = Ollama(model="llama3.2")
 res = llm.predict(input)
 print (res)

examples/modelfile-mario/Modelfile

@@ -1,4 +1,4 @@
-FROM llama3.1
+FROM llama3.2
 PARAMETER temperature 1
 SYSTEM """
 You are Mario from super mario bros, acting as an assistant.

examples/modelfile-mario/readme.md

@@ -2,12 +2,12 @@
 
 # Example character: Mario
 
-This example shows how to create a basic character using Llama3.1 as the base model.
+This example shows how to create a basic character using Llama 3.2 as the base model.
 
 To run this example:
 
 1. Download the Modelfile
-2. `ollama pull llama3.1` to get the base model used in the model file.
+2. `ollama pull llama3.2` to get the base model used in the model file.
 3. `ollama create NAME -f ./Modelfile`
 4. `ollama run NAME`
 
@@ -18,7 +18,7 @@ Ask it some questions like "Who are you?" or "Is Peach in trouble again?"
 What the model file looks like:
 
 ```
-FROM llama3.1
+FROM llama3.2
 PARAMETER temperature 1
 SYSTEM """
 You are Mario from Super Mario Bros, acting as an assistant.

examples/python-grounded-factuality-rag-check/README.md

@@ -0,0 +1,93 @@
+# RAG Hallucination Checker using Bespoke-Minicheck
+
+This example allows the user to ask questions related to a document, which can be specified via an article url. Relevant chunks are retreived from the document and given to `llama3.2` as context to answer the question. Then each sentence in the answer is checked against the retrieved chunks using `bespoke-minicheck` to ensure that the answer does not contain hallucinations. 
+
+## Running the Example
+
+1. Ensure `all-minilm` (embedding) `llama3.2` (chat) and `bespoke-minicheck` (check) models installed:
+
+   ```bash
+   ollama pull all-minilm
+   ollama pull llama3.2
+   ollama pull bespoke-minicheck
+   ```
+
+2. Install the dependencies.
+
+   ```bash
+   pip install -r requirements.txt
+   ```
+
+3. Run the example:
+
+   ```bash
+   python main.py
+   ```
+
+## Expected Output
+
+```text
+Enter the URL of an article you want to chat with, or press Enter for default example:
+
+Loaded, chunked, and embedded text from https://www.theverge.com/2024/9/12/24242439/openai-o1-model-reasoning-strawberry-chatgpt.
+
+Enter your question or type quit: Who is the CEO of openai?
+
+Retrieved chunks:
+OpenAI is releasing a new model called o1 , the first in a planned series of “ reasoning ” models that have been trained to answer more complex questions , faster than a human can . It ’ s being released alongside o1-mini , a smaller , cheaper version . And yes , if you ’ re steeped in AI rumors : this is , in fact , the extremely hyped Strawberry model . For OpenAI , o1 represents a step toward its broader goal of human-like artificial intelligence .
+
+OpenAI is releasing a new model called o1 , the first in a planned series of “ reasoning ” models that have been trained to answer more complex questions , faster than a human can . It ’ s being released alongside o1-mini , a smaller , cheaper version . And yes , if you ’ re steeped in AI rumors : this is , in fact , the extremely hyped Strawberry model . For OpenAI , o1 represents a step toward its broader goal of human-like artificial intelligence . More practically , it does a better job at writing code and solving multistep problems than previous models . But it ’ s also more expensive and slower to use than GPT-4o . OpenAI is calling this release of o1 a “ preview ” to emphasize how nascent it is . ChatGPT Plus and Team users get access to both o1-preview and o1-mini starting today , while Enterprise and Edu users will get access early next week .
+
+More practically , it does a better job at writing code and solving multistep problems than previous models . But it ’ s also more expensive and slower to use than GPT-4o . OpenAI is calling this release of o1 a “ preview ” to emphasize how nascent it is . ChatGPT Plus and Team users get access to both o1-preview and o1-mini starting today , while Enterprise and Edu users will get access early next week . OpenAI says it plans to bring o1-mini access to all the free users of ChatGPT but hasn ’ t set a release date yet . Developer access to o1 is really expensive : In the API , o1-preview is $ 15 per 1 million input tokens , or chunks of text parsed by the model , and $ 60 per 1 million output tokens . For comparison , GPT-4o costs $ 5 per 1 million input tokens and $ 15 per 1 million output tokens .
+
+OpenAI says it plans to bring o1-mini access to all the free users of ChatGPT but hasn ’ t set a release date yet . Developer access to o1 is really expensive : In the API , o1-preview is $ 15 per 1 million input tokens , or chunks of text parsed by the model , and $ 60 per 1 million output tokens . For comparison , GPT-4o costs $ 5 per 1 million input tokens and $ 15 per 1 million output tokens . The training behind o1 is fundamentally different from its predecessors , OpenAI ’ s research lead , Jerry Tworek , tells me , though the company is being vague about the exact details . He says o1 “ has been trained using a completely new optimization algorithm and a new training dataset specifically tailored for it. ” Image : OpenAI OpenAI taught previous GPT models to mimic patterns from its training data .
+
+LLM Answer:
+The text does not mention the CEO of OpenAI. It only discusses the release of a new model called o1 and some details about it, but does not provide information on the company's leadership.
+
+LLM Claim: The text does not mention the CEO of OpenAI.
+Is this claim supported by the context according to bespoke-minicheck? Yes
+
+LLM Claim: It only discusses the release of a new model called o1 and some details about it, but does not provide information on the company's leadership.
+Is this claim supported by the context according to bespoke-minicheck? No
+```
+
+The second claim is unsupported since the text mentions the research lead. 
+
+Another tricky example:
+
+```text
+
+Enter your question or type quit: what sets o1 apart from gpt-4o?
+
+Retrieved chunks: 
+OpenAI says it plans to bring o1-mini access to all the free users of ChatGPT but hasn ’ t set a release date yet . Developer access to o1 is really expensive : In the API , o1-preview is $ 15 per 1 million input tokens , or chunks of text parsed by the model , and $ 60 per 1 million output tokens . For comparison , GPT-4o costs $ 5 per 1 million input tokens and $ 15 per 1 million output tokens . The training behind o1 is fundamentally different from its predecessors , OpenAI ’ s research lead , Jerry Tworek , tells me , though the company is being vague about the exact details . He says o1 “ has been trained using a completely new optimization algorithm and a new training dataset specifically tailored for it. ” Image : OpenAI OpenAI taught previous GPT models to mimic patterns from its training data .
+
+He says OpenAI also tested o1 against a qualifying exam for the International Mathematics Olympiad , and while GPT-4o only correctly solved only 13 percent of problems , o1 scored 83 percent . “ We can ’ t say we solved hallucinations ” In online programming contests known as Codeforces competitions , this new model reached the 89th percentile of participants , and OpenAI claims the next update of this model will perform “ similarly to PhD students on challenging benchmark tasks in physics , chemistry and biology. ” At the same time , o1 is not as capable as GPT-4o in a lot of areas . It doesn ’ t do as well on factual knowledge about the world .
+
+More practically , it does a better job at writing code and solving multistep problems than previous models . But it ’ s also more expensive and slower to use than GPT-4o . OpenAI is calling this release of o1 a “ preview ” to emphasize how nascent it is . ChatGPT Plus and Team users get access to both o1-preview and o1-mini starting today , while Enterprise and Edu users will get access early next week . OpenAI says it plans to bring o1-mini access to all the free users of ChatGPT but hasn ’ t set a release date yet . Developer access to o1 is really expensive : In the API , o1-preview is $ 15 per 1 million input tokens , or chunks of text parsed by the model , and $ 60 per 1 million output tokens . For comparison , GPT-4o costs $ 5 per 1 million input tokens and $ 15 per 1 million output tokens .
+
+“ We can ’ t say we solved hallucinations ” In online programming contests known as Codeforces competitions , this new model reached the 89th percentile of participants , and OpenAI claims the next update of this model will perform “ similarly to PhD students on challenging benchmark tasks in physics , chemistry and biology. ” At the same time , o1 is not as capable as GPT-4o in a lot of areas . It doesn ’ t do as well on factual knowledge about the world . It also doesn ’ t have the ability to browse the web or process files and images . Still , the company believes it represents a brand-new class of capabilities . It was named o1 to indicate “ resetting the counter back to 1. ” “ I ’ m gon na be honest : I think we ’ re terrible at naming , traditionally , ” McGrew says .
+LLM Answer: According to the text, several things set o1 apart from GPT-4o:
+
+* In online programming contests (Codeforces competitions), o1 scored 83% correct solutions compared to GPT-4o's 13%.
+* The training behind o1 is "fundamentally different" from its predecessors, including a completely new optimization algorithm and a new training dataset specifically tailored for it.
+* o1 has been shown to perform similarly to PhD students on challenging benchmark tasks in physics, chemistry, and biology, while GPT-4o does not have this level of performance.
+* However, o1 is also less capable than GPT-4o in some areas, particularly with regard to factual knowledge about the world.
+
+LLM Claim: According to the text, several things set o1 apart from GPT-4o:
+
+* In online programming contests (Codeforces competitions), o1 scored 83% correct solutions compared to GPT-4o's 13%.
+Is this claim supported by the context according to bespoke-minicheck? Yes
+
+LLM Claim: * The training behind o1 is "fundamentally different" from its predecessors, including a completely new optimization algorithm and a new training dataset specifically tailored for it.
+Is this claim supported by the context according to bespoke-minicheck? Yes
+
+LLM Claim: * o1 has been shown to perform similarly to PhD students on challenging benchmark tasks in physics, chemistry, and biology, while GPT-4o does not have this level of performance.
+Is this claim supported by the context according to bespoke-minicheck? No
+
+LLM Claim: * However, o1 is also less capable than GPT-4o in some areas, particularly with regard to factual knowledge about the world.
+Is this claim supported by the context according to bespoke-minicheck? Yes
+```
+
+We see that the third claim "* o1 has been shown to perform similarly to PhD students on challenging benchmark tasks in physics, chemistry, and biology, while GPT-4o does not have this level of performance." is not supported by the context. This is because the context only mentions that o1 "is claimed to perform" which is different from "has been shown to perform".

examples/python-grounded-factuality-rag-check/main.py

@@ -0,0 +1,137 @@
+import ollama
+import warnings
+from mattsollamatools import chunker
+from newspaper import Article
+import numpy as np
+from sklearn.neighbors import NearestNeighbors
+import nltk
+
+warnings.filterwarnings(
+    "ignore", category=FutureWarning, module="transformers.tokenization_utils_base"
+)
+nltk.download("punkt_tab", quiet=True)
+
+
+def getArticleText(url):
+    """Gets the text of an article from a URL.
+
+    Often there are a bunch of ads and menus on pages for a news article.
+    This uses newspaper3k to get just the text of just the article.
+    """
+    article = Article(url)
+    article.download()
+    article.parse()
+    return article.text
+
+
+def knn_search(question_embedding, embeddings, k=5):
+    """Performs K-nearest neighbors (KNN) search"""
+    X = np.array(
+        [item["embedding"] for article in embeddings for item in article["embeddings"]]
+    )
+    source_texts = [
+        item["source"] for article in embeddings for item in article["embeddings"]
+    ]
+
+    # Fit a KNN model on the embeddings
+    knn = NearestNeighbors(n_neighbors=k, metric="cosine")
+    knn.fit(X)
+
+    # Find the indices and distances of the k-nearest neighbors.
+    _, indices = knn.kneighbors(question_embedding, n_neighbors=k)
+
+    # Get the indices and source texts of the best matches
+    best_matches = [(indices[0][i], source_texts[indices[0][i]]) for i in range(k)]
+
+    return best_matches
+
+
+def check(document, claim):
+    """Checks if the claim is supported by the document by calling bespoke-minicheck.
+
+    Returns Yes/yes if the claim is supported by the document, No/no otherwise.
+    Support for logits will be added in the future.
+
+    bespoke-minicheck's system prompt is defined as:
+      'Determine whether the provided claim is consistent with the corresponding
+      document. Consistency in this context implies that all information presented in the claim
+      is substantiated by the document. If not, it should be considered inconsistent. Please
+      assess the claim's consistency with the document by responding with either "Yes" or "No".'
+
+    bespoke-minicheck's user prompt is defined as:
+      "Document: {document}\nClaim: {claim}"
+    """
+    prompt = f"Document: {document}\nClaim: {claim}"
+    response = ollama.generate(
+        model="bespoke-minicheck", prompt=prompt, options={"num_predict": 2, "temperature": 0.0}
+    )
+    return response["response"].strip()
+
+
+if __name__ == "__main__":
+    allEmbeddings = []
+    default_url = "https://www.theverge.com/2024/9/12/24242439/openai-o1-model-reasoning-strawberry-chatgpt"
+    user_input = input(
+        "Enter the URL of an article you want to chat with, or press Enter for default example: "
+    )
+    article_url = user_input.strip() if user_input.strip() else default_url
+    article = {}
+    article["embeddings"] = []
+    article["url"] = article_url
+    text = getArticleText(article_url)
+    chunks = chunker(text)
+
+    # Embed (batch) chunks using ollama
+    embeddings = ollama.embed(model="all-minilm", input=chunks)["embeddings"]
+
+    for chunk, embedding in zip(chunks, embeddings):
+        item = {}
+        item["source"] = chunk
+        item["embedding"] = embedding
+        item["sourcelength"] = len(chunk)
+        article["embeddings"].append(item)
+
+    allEmbeddings.append(article)
+
+    print(f"\nLoaded, chunked, and embedded text from {article_url}.\n")
+
+    while True:
+        # Input a question from the user
+        # For example, "Who is the chief research officer?"
+        question = input("Enter your question or type quit: ")
+
+        if question.lower() == "quit":
+            break
+
+        # Embed the user's question using ollama.embed
+        question_embedding = ollama.embed(model="all-minilm", input=question)[
+            "embeddings"
+        ]
+
+        # Perform KNN search to find the best matches (indices and source text)
+        best_matches = knn_search(question_embedding, allEmbeddings, k=4)
+
+        sourcetext = "\n\n".join([source_text for (_, source_text) in best_matches])
+
+        print(f"\nRetrieved chunks: \n{sourcetext}\n")
+
+        # Give the retreived chunks and question to the chat model
+        system_prompt = f"Only use the following information to answer the question. Do not use anything else: {sourcetext}"
+
+        ollama_response = ollama.generate(
+            model="llama3.2",
+            prompt=question,
+            system=system_prompt,
+            options={"stream": False},
+        )
+
+        answer = ollama_response["response"]
+        print(f"LLM Answer:\n{answer}\n")
+
+        # Check each sentence in the response for grounded factuality
+        if answer:
+            for claim in nltk.sent_tokenize(answer):
+                print(f"LLM Claim: {claim}")
+                print(
+                    f"Is this claim supported by the context according to bespoke-minicheck? {check(sourcetext, claim)}\n"
+                )

examples/python-grounded-factuality-rag-check/requirements.txt

@@ -0,0 +1,8 @@
+ollama
+lxml==5.3.0
+lxml_html_clean==0.2.2
+mattsollamatools==0.0.25
+newspaper3k==0.2.8
+nltk==3.9.1
+numpy==1.26.4
+scikit-learn==1.5.2

examples/python-grounded-factuality-simple-check/main.py

@@ -0,0 +1,53 @@
+"""Simple example to demonstrate how to use the bespoke-minicheck model."""
+
+import ollama
+
+# NOTE: ollama must be running for this to work, start the ollama app or run `ollama serve`
+
+
+def check(document, claim):
+    """Checks if the claim is supported by the document by calling bespoke-minicheck.
+
+    Returns Yes/yes if the claim is supported by the document, No/no otherwise.
+    Support for logits will be added in the future.
+
+    bespoke-minicheck's system prompt is defined as:
+      'Determine whether the provided claim is consistent with the corresponding
+      document. Consistency in this context implies that all information presented in the claim
+      is substantiated by the document. If not, it should be considered inconsistent. Please
+      assess the claim's consistency with the document by responding with either "Yes" or "No".'
+
+    bespoke-minicheck's user prompt is defined as:
+      "Document: {document}\nClaim: {claim}"
+    """
+    prompt = f"Document: {document}\nClaim: {claim}"
+    response = ollama.generate(
+        model="bespoke-minicheck", prompt=prompt, options={"num_predict": 2, "temperature": 0.0}
+    )
+    return response["response"].strip()
+
+
+def get_user_input(prompt):
+    user_input = input(prompt)
+    if not user_input:
+        exit()
+    print()
+    return user_input
+
+
+def main():
+    while True:
+        # Get a document from the user (e.g. "Ryan likes running and biking.")
+        document = get_user_input("Enter a document: ")
+        # Get a claim from the user (e.g. "Ryan likes to run.")
+        claim = get_user_input("Enter a claim: ")
+        # Check if the claim is supported by the document
+        grounded_factuality_check = check(document, claim)
+        print(
+            f"Is the claim supported by the document according to bespoke-minicheck? {grounded_factuality_check}"
+        )
+        print("\n\n")
+
+
+if __name__ == "__main__":
+    main()

examples/python-grounded-factuality-simple-check/readme.md

@@ -0,0 +1,54 @@
+# Simple Bespoke-Minicheck Example
+
+`bespoke-minicheck` is a model for checking if a claim is supported by a document. It is used through the **generate** endpoint, which is called in this example with a `prompt` that includes the expected formatting of the user input. 
+
+## Running the Example
+
+1. Ensure you have the `bespoke-minicheck` model installed:
+
+   ```bash
+   ollama pull bespoke-minicheck
+   ```
+
+2. Install the dependencies:
+
+   ```bash
+   pip install -r requirements.txt
+   ```
+
+3. Run the program:
+
+   ```bash
+   python main.py
+   ```
+
+4. Enter a document and a claim when prompted:
+
+   ```bash
+   Enter a document: Roses are red.
+
+   Enter a claim: Roses are blue. 
+   ```
+
+   The claim and document are then given to the `bespoke-minicheck` as inputs, which then generates a response (Yes or No) on whether the claim is supported by the document.
+
+   ```bash
+   Is the claim supported by the document according to bespoke-minicheck? No
+   ```
+
+## More Examples
+
+Document ([source](https://en.wikipedia.org/wiki/Apple_I)): 
+> The Apple Computer 1 (Apple-1[a]), later known predominantly as the Apple I(written with a Roman numeral),[b] is an 8-bit motherboard-only personal computer designed by Steve Wozniak[5][6] and released by the Apple Computer Company (now Apple Inc.) in 1976. The company was initially formed to sell the Apple I – its first product – and would later become the world's largest technology company.[7] The idea of starting a company and selling the computer came from Wozniak's friend and Apple co-founder Steve Jobs.[8][9] One of the main innovations of the Apple I was that it included video display terminal circuitry on its circuit board, allowing it to connect to a low-cost composite video monitor or television, instead of an expensive computer terminal, compared to most existing computers at the time.
+
+Claim: 
+>The Apple I is a 16-bit computer.
+
+Expected output:
+>Is the claim supported by the document according to bespoke-minicheck? **No**
+
+Claim: 
+>Apple was originally called the Apple Computer Company.
+
+Expected output:
+>Is the claim supported by the document according to bespoke-minicheck? **Yes**

examples/python-grounded-factuality-simple-check/requirements.txt

@@ -0,0 +1 @@
+ollama

examples/python-json-datagenerator/predefinedschema.py

@@ -2,7 +2,7 @@ import requests
 import json
 import random
 
-model = "llama3.1"
+model = "llama3.2"
 template = {
   "firstName": "",
   "lastName": "",

examples/python-json-datagenerator/randomaddresses.py

@@ -12,7 +12,7 @@ countries = [
     "France",
 ]
 country = random.choice(countries)
-model = "llama3.1"
+model = "llama3.2"
 
 prompt = f"generate one realistically believable sample data set of a persons first name, last name, address in {country}, and phone number. Do not use common names. Respond using JSON. Key names should have no backslashes, values should use plain ascii with no special characters."
 

examples/python-json-datagenerator/readme.md

@@ -6,10 +6,10 @@ There are two python scripts in this example. `randomaddresses.py` generates ran
 
 ## Running the Example
 
-1. Ensure you have the `llama3.1` model installed:
+1. Ensure you have the `llama3.2` model installed:
 
    ```bash
-   ollama pull llama3.1
+   ollama pull llama3.2
    ```
 
 2. Install the Python Requirements.

examples/python-simplechat/client.py

@@ -2,7 +2,7 @@ import json
 import requests
 
 # NOTE: ollama must be running for this to work, start the ollama app or run `ollama serve`
-model = "llama3.1"  # TODO: update this for whatever model you wish to use
+model = "llama3.2"  # TODO: update this for whatever model you wish to use
 
 
 def chat(messages):

examples/python-simplechat/readme.md

@@ -4,10 +4,10 @@ The **chat** endpoint is one of two ways to generate text from an LLM with Ollam
 
 ## Running the Example
 
-1. Ensure you have the `llama3.1` model installed:
+1. Ensure you have the `llama3.2` model installed:
 
    ```bash
-   ollama pull llama3.1
+   ollama pull llama3.2
    ```
 
 2. Install the Python Requirements.

examples/typescript-simplechat/client.ts

@@ -1,6 +1,6 @@
 import * as readline from "readline";
 
-const model = "llama3.1";
+const model = "llama3.2";
 type Message = {
   role: "assistant" | "user" | "system";
   content: string;

gpu/gpu.go

@@ -205,13 +205,16 @@ func GetGPUInfo() GpuInfoList {
 		if err != nil {
 			slog.Warn("error looking up system memory", "error", err)
 		}
+		depPath := LibraryDir()
+
 		cpus = []CPUInfo{
 			{
 				GpuInfo: GpuInfo{
-					memInfo: mem,
-					Library: "cpu",
-					Variant: cpuCapability.String(),
-					ID:      "0",
+					memInfo:        mem,
+					Library:        "cpu",
+					Variant:        cpuCapability.String(),
+					ID:             "0",
+					DependencyPath: depPath,
 				},
 			},
 		}
@@ -224,8 +227,6 @@ func GetGPUInfo() GpuInfoList {
 			return GpuInfoList{cpus[0].GpuInfo}
 		}
 
-		depPath := LibraryDir()
-
 		// Load ALL libraries
 		cHandles = initCudaHandles()
 

llm/ext_server/CMakeLists.txt

@@ -10,5 +10,6 @@ target_compile_definitions(${TARGET} PRIVATE
 target_link_libraries(${TARGET} PRIVATE ggml llama common llava ${CMAKE_THREAD_LIBS_INIT} ${LLAMA_SERVER_LDFLAGS})
 if (WIN32)
     TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
+    target_link_options(${TARGET} PRIVATE -municode -Wl,/subsystem:console)
 endif()
 target_compile_features(${TARGET} PRIVATE cxx_std_11)

llm/ext_server/server.cpp

@@ -1246,6 +1246,17 @@ struct llama_server_context
         task.type = TASK_TYPE_COMPLETION;
         task.multitask_id = multitask_id;
 
+        // Set cross attention state for mllama models
+        // TODO (jmorganca): this should be provided via the API
+        // TODO (jmorganca): generalize this beyond mllama models
+        char arch_str[256];
+        llama_model_meta_val_str(model, "general.architecture", arch_str, 256);
+        if (strcmp(arch_str, "mllama") == 0) {
+            // TODO (jmorganca): this should be passed in via the llama_decode api
+            // or similar, maybe using the llama_batch struct
+            // llama_set_cross_attn_state(ctx, (float*)cross_attn_state);
+        }
+
         // when a completion task's prompt array is not a singleton, we split it into multiple requests
         // otherwise, it's a single-prompt task, we actually queue it
         // if there's numbers in the prompt array it will be treated as an array of tokens

llm/generate/gen_windows.ps1

@@ -19,6 +19,19 @@ function amdGPUs {
 
 
 function init_vars {
+    write-host "Checking for cmake..."
+    get-command cmake
+    write-host "Checking for ninja..."
+    $d=(get-command -ea 'silentlycontinue' ninja).path
+    if ($null -eq $d) {
+        $MSVC_INSTALL=(Get-CimInstance MSFT_VSInstance -Namespace root/cimv2/vs)[0].InstallLocation
+        $matches=(gci -path $MSVC_INSTALL -r -fi ninja.exe)
+        if ($matches.count -eq 0) {
+            throw "Unable to locate ninja"
+        }
+        $ninjaDir=($matches[0].FullName | split-path -parent)
+        $env:PATH="$env:PATH;$ninjaDir"
+    }
     if (!$script:SRC_DIR) {
         $script:SRC_DIR = $(resolve-path "..\..\")
     }
@@ -145,7 +158,7 @@ function cleanup {
         }
 
         # Checkout each file
-        foreach ($file in $filePaths) {            
+        foreach ($file in $filePaths) {
             git -C "${script:llamacppDir}" checkout $file
         }
         git -C "${script:llamacppDir}" checkout CMakeLists.txt
@@ -162,12 +175,12 @@ 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
+
+        # cmake will silently fallback to msvc compilers if gcc 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 it
+        write-host "Checking for gcc..."
+        get-command  gcc
+        get-command  mingw32-make
         $oldTargets = $script:cmakeTargets
         $script:cmakeTargets = @("llama", "ggml")
         $script:cmakeDefs = @(
@@ -191,11 +204,10 @@ function build_static() {
     }
 }
 
-function build_cpu($gen_arch) {
+function build_cpu_x64 {
     if ((-not "${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, "-DGGML_AVX=off", "-DGGML_AVX2=off", "-DGGML_AVX512=off", "-DGGML_FMA=off", "-DGGML_F16C=off") + $script:cmakeDefs
+        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DGGML_AVX=off", "-DGGML_AVX2=off", "-DGGML_AVX512=off", "-DGGML_FMA=off", "-DGGML_F16C=off") + $script:cmakeDefs
         $script:buildDir="../build/windows/${script:ARCH}/cpu"
         $script:distDir="$script:DIST_BASE\cpu"
         write-host "Building LCD CPU"
@@ -207,6 +219,32 @@ function build_cpu($gen_arch) {
     }
 }
 
+function build_cpu_arm64 {
+    if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu"))) {
+        init_vars
+        write-host "Checking for clang..."
+        get-command clang
+        $env:CFLAGS="-march=armv8.7-a -fvectorize -ffp-model=fast -fno-finite-math-only"
+        $env:CXXFLAGS="$env:CFLAGS"
+        $env:LDFLAGS="-static-libstdc++"
+        $script:cmakeDefs = $script:commonCpuDefs + @(
+            "-DCMAKE_VERBOSE_MAKEFILE=on",
+            "-DCMAKE_C_COMPILER=clang.exe",
+            "-DCMAKE_CXX_COMPILER=clang++.exe",
+            "-DMSVC_RUNTIME_LIBRARY=MultiThreaded"
+        ) + $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
@@ -331,7 +369,7 @@ function build_rocm() {
         $script:buildDir="../build/windows/${script:ARCH}/rocm$script:ROCM_VARIANT"
         $script:distDir="$script:DIST_BASE\rocm$script:ROCM_VARIANT"
         $script:cmakeDefs += @(
-            "-G", "Ninja", 
+            "-G", "Ninja",
             "-DCMAKE_C_COMPILER=clang.exe",
             "-DCMAKE_CXX_COMPILER=clang++.exe",
             "-DGGML_HIPBLAS=on",
@@ -380,9 +418,9 @@ if ($($args.count) -eq 0) {
     apply_patches
     build_static
     if ($script:ARCH -eq "arm64") {
-        build_cpu("ARM64")
+        build_cpu_arm64
     } else { # amd64
-        build_cpu("x64")
+        build_cpu_x64
         build_cpu_avx
         build_cpu_avx2
         build_cuda
@@ -396,5 +434,5 @@ if ($($args.count) -eq 0) {
     for ( $i = 0; $i -lt $args.count; $i++ ) {
         write-host "performing $($args[$i])"
         & $($args[$i])
-    } 
+    }
 }

llm/llm.go

@@ -5,7 +5,7 @@ package llm
 // #cgo darwin,arm64 LDFLAGS: -L${SRCDIR}/build/darwin/arm64_static -L${SRCDIR}/build/darwin/arm64_static/src -L${SRCDIR}/build/darwin/arm64_static/ggml/src -framework Accelerate -framework Metal
 // #cgo darwin,amd64 LDFLAGS: -L${SRCDIR}/build/darwin/x86_64_static -L${SRCDIR}/build/darwin/x86_64_static/src -L${SRCDIR}/build/darwin/x86_64_static/ggml/src
 // #cgo windows,amd64 LDFLAGS: -static-libstdc++ -static-libgcc -static -L${SRCDIR}/build/windows/amd64_static -L${SRCDIR}/build/windows/amd64_static/src -L${SRCDIR}/build/windows/amd64_static/ggml/src
-// #cgo windows,arm64 LDFLAGS: -static-libstdc++ -static-libgcc -static -L${SRCDIR}/build/windows/arm64_static -L${SRCDIR}/build/windows/arm64_static/src -L${SRCDIR}/build/windows/arm64_static/ggml/src
+// #cgo windows,arm64 LDFLAGS: -lllama -lggml -static-libstdc++ -static-libgcc -static -L${SRCDIR}/build/windows/arm64_static -L${SRCDIR}/build/windows/arm64_static/src -L${SRCDIR}/build/windows/arm64_static/ggml/src
 // #cgo linux,amd64 LDFLAGS: -L${SRCDIR}/build/linux/x86_64_static -L${SRCDIR}/build/linux/x86_64_static/src -L${SRCDIR}/build/linux/x86_64_static/ggml/src
 // #cgo linux,arm64 LDFLAGS: -L${SRCDIR}/build/linux/arm64_static -L${SRCDIR}/build/linux/arm64_static/src -L${SRCDIR}/build/linux/arm64_static/ggml/src
 // #include <stdlib.h>

llm/llm_windows.go

@@ -4,7 +4,10 @@ import (
 	"syscall"
 )
 
-const CREATE_DEFAULT_ERROR_MODE = 0x04000000
+const (
+	CREATE_DEFAULT_ERROR_MODE   = 0x04000000
+	ABOVE_NORMAL_PRIORITY_CLASS = 0x00008000
+)
 
 var LlamaServerSysProcAttr = &syscall.SysProcAttr{
 	// Wire up the default error handling logic If for some reason a DLL is
@@ -12,5 +15,8 @@ var LlamaServerSysProcAttr = &syscall.SysProcAttr{
 	// the user can either fix their PATH, or report a bug. Without this
 	// setting, the process exits immediately with a generic exit status but no
 	// way to (easily) figure out what the actual missing DLL was.
-	CreationFlags: CREATE_DEFAULT_ERROR_MODE,
+	//
+	// Setting Above Normal priority class ensures when running as a "background service"
+	// with "programs" given best priority, we aren't starved of cpu cycles
+	CreationFlags: CREATE_DEFAULT_ERROR_MODE | ABOVE_NORMAL_PRIORITY_CLASS,
 }

llm/patches/0008-solar-pro.patch

@@ -0,0 +1,402 @@
+From 8313ce5f43f11f3d84f352f97f3802792e90e18c Mon Sep 17 00:00:00 2001
+From: Michael Yang <mxyng@pm.me>
+Date: Mon, 16 Sep 2024 15:53:16 -0700
+Subject: [PATCH] add solar-pro support
+
+solar-pro introduces block skip connections where blocks are connected
+to other, non-sequential blocks with a scale multiple
+
+this change adds 4 new keys to store the skip connections and one new
+tensor to store the scalar. the scalar is implemented a 1-dimensional
+tensor with 2 elements dervied from the model's bskcn_tv configuration.
+in general, the values are (bskcn_tv, 1 - bskcn_tv)
+---
+ src/llama.cpp | 267 +++++++++++++++++++++++++++++++++++++++++++++++---
+ 1 file changed, 254 insertions(+), 13 deletions(-)
+
+diff --git a/src/llama.cpp b/src/llama.cpp
+index f79bd782..b7771f53 100644
+--- a/src/llama.cpp
++++ b/src/llama.cpp
+@@ -213,6 +213,7 @@ enum llm_arch {
+     LLM_ARCH_NEMOTRON,
+     LLM_ARCH_EXAONE,
+     LLM_ARCH_RWKV6,
++    LLM_ARCH_SOLAR,
+     LLM_ARCH_UNKNOWN,
+ };
+ 
+@@ -261,6 +262,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
+     { LLM_ARCH_NEMOTRON,        "nemotron"     },
+     { LLM_ARCH_EXAONE,          "exaone"       },
+     { LLM_ARCH_RWKV6,           "rwkv6"        },
++    { LLM_ARCH_SOLAR,           "solar"        },
+     { LLM_ARCH_UNKNOWN,         "(unknown)"    },
+ };
+ 
+@@ -314,6 +316,7 @@ enum llm_kv {
+     LLM_KV_ATTENTION_KV_LORA_RANK,
+     LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
+     LLM_KV_ATTENTION_SLIDING_WINDOW,
++    LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,
+ 
+     LLM_KV_ROPE_DIMENSION_COUNT,
+     LLM_KV_ROPE_FREQ_BASE,
+@@ -405,19 +408,20 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
+     { LLM_KV_TIME_MIX_EXTRA_DIM,                "%s.time_mix_extra_dim"                },
+     { LLM_KV_TIME_DECAY_EXTRA_DIM,              "%s.time_decay_extra_dim"              },
+ 
+-    { LLM_KV_ATTENTION_HEAD_COUNT,             "%s.attention.head_count"             },
+-    { LLM_KV_ATTENTION_HEAD_COUNT_KV,          "%s.attention.head_count_kv"          },
+-    { LLM_KV_ATTENTION_MAX_ALIBI_BIAS,         "%s.attention.max_alibi_bias"         },
+-    { LLM_KV_ATTENTION_CLAMP_KQV,              "%s.attention.clamp_kqv"              },
+-    { LLM_KV_ATTENTION_KEY_LENGTH,             "%s.attention.key_length"             },
+-    { LLM_KV_ATTENTION_VALUE_LENGTH,           "%s.attention.value_length"           },
+-    { LLM_KV_ATTENTION_LAYERNORM_EPS,          "%s.attention.layer_norm_epsilon"     },
+-    { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,      "%s.attention.layer_norm_rms_epsilon" },
+-    { LLM_KV_ATTENTION_CAUSAL,                 "%s.attention.causal"                 },
+-    { LLM_KV_ATTENTION_Q_LORA_RANK,            "%s.attention.q_lora_rank"            },
+-    { LLM_KV_ATTENTION_KV_LORA_RANK,           "%s.attention.kv_lora_rank"           },
+-    { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" },
+-    { LLM_KV_ATTENTION_SLIDING_WINDOW,         "%s.attention.sliding_window"         },
++    { LLM_KV_ATTENTION_HEAD_COUNT,             "%s.attention.head_count"               },
++    { LLM_KV_ATTENTION_HEAD_COUNT_KV,          "%s.attention.head_count_kv"            },
++    { LLM_KV_ATTENTION_MAX_ALIBI_BIAS,         "%s.attention.max_alibi_bias"           },
++    { LLM_KV_ATTENTION_CLAMP_KQV,              "%s.attention.clamp_kqv"                },
++    { LLM_KV_ATTENTION_KEY_LENGTH,             "%s.attention.key_length"               },
++    { LLM_KV_ATTENTION_VALUE_LENGTH,           "%s.attention.value_length"             },
++    { LLM_KV_ATTENTION_LAYERNORM_EPS,          "%s.attention.layer_norm_epsilon"       },
++    { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,      "%s.attention.layer_norm_rms_epsilon"   },
++    { LLM_KV_ATTENTION_CAUSAL,                 "%s.attention.causal"                   },
++    { LLM_KV_ATTENTION_Q_LORA_RANK,            "%s.attention.q_lora_rank"              },
++    { LLM_KV_ATTENTION_KV_LORA_RANK,           "%s.attention.kv_lora_rank"             },
++    { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count"   },
++    { LLM_KV_ATTENTION_SLIDING_WINDOW,         "%s.attention.sliding_window"           },
++    { LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,  "%s.attention.block_skip_connection.%d" },
+ 
+     { LLM_KV_ROPE_DIMENSION_COUNT,          "%s.rope.dimension_count"                 },
+     { LLM_KV_ROPE_FREQ_BASE,                "%s.rope.freq_base"                       },
+@@ -589,6 +593,7 @@ enum llm_tensor {
+     LLM_TENSOR_ENC_FFN_DOWN,
+     LLM_TENSOR_ENC_FFN_UP,
+     LLM_TENSOR_ENC_OUTPUT_NORM,
++    LLM_TENSOR_BSKCN_TV,
+ };
+ 
+ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES = {
+@@ -1408,6 +1413,24 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
+             { LLM_TENSOR_CHANNEL_MIX_RECEPTANCE,    "blk.%d.channel_mix_receptance" },
+         },
+     },
++    {
++        LLM_ARCH_SOLAR,
++        {
++            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
++            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
++            { LLM_TENSOR_OUTPUT,          "output" },
++            { 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_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_BSKCN_TV,        "bskcn_tv" },
++        },
++    },
+     {
+         LLM_ARCH_UNKNOWN,
+         {
+@@ -2237,6 +2260,7 @@ enum e_model {
+     MODEL_15B,
+     MODEL_16B,
+     MODEL_20B,
++    MODEL_22B,
+     MODEL_30B,
+     MODEL_34B,
+     MODEL_35B,
+@@ -2284,6 +2308,8 @@ struct llama_hparams {
+     std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_kv_arr;
+     std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;
+ 
++    std::array<std::array<uint32_t, LLAMA_MAX_LAYERS>, 4> n_bskcn_arr;
++
+     uint32_t n_layer_dense_lead = 0;
+     uint32_t n_lora_q = 0;
+     uint32_t n_lora_kv = 0;
+@@ -2349,6 +2375,7 @@ struct llama_hparams {
+         if (this->n_head_arr    != other.n_head_arr)    return true;
+         if (this->n_head_kv_arr != other.n_head_kv_arr) return true;
+         if (this->n_ff_arr      != other.n_ff_arr)      return true;
++        if (this->n_bskcn_arr   != other.n_bskcn_arr)   return true;
+ 
+         if (this->n_rel_attn_bkts    != other.n_rel_attn_bkts)    return true;
+         if (this->n_layer_dense_lead != other.n_layer_dense_lead) return true;
+@@ -2455,6 +2482,14 @@ struct llama_hparams {
+             return ssm_d_state * ssm_d_inner;
+         }
+     }
++
++    bool n_bskcn(uint32_t n, uint32_t il = 0) const {
++        if (il < n_layer) {
++            return n_bskcn_arr[n][il] > 0;
++        }
++
++        GGML_ABORT("fatal error");
++    }
+ };
+ 
+ static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");
+@@ -2635,6 +2670,8 @@ struct llama_layer {
+     struct ggml_tensor * ffn_gate_scale;
+     struct ggml_tensor * ffn_up_scale;
+     struct ggml_tensor * ffn_down_scale;
++
++    struct ggml_tensor * bskcn_tv;
+ };
+ 
+ // very similar to llama_batch,
+@@ -5937,6 +5974,21 @@ static void llm_load_hparams(
+                     default: model.type = e_model::MODEL_UNKNOWN;
+                 }
+             } break;
++        case LLM_ARCH_SOLAR:
++            {
++                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
++
++                for (int i = 0; i < hparams.n_bskcn_arr.max_size(); ++i) {
++                    auto & bskcn = hparams.n_bskcn_arr.at(i);
++                    bskcn.fill(0);
++                    ml.get_key_or_arr(::format(LLM_KV_NAMES.at(LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION), LLM_ARCH_NAMES.at(ml.llm_kv.arch), i), bskcn, hparams.n_layer, false);
++                }
++
++                switch (hparams.n_layer) {
++                    case 64: model.type = e_model::MODEL_22B; break;
++                    default: model.type = e_model::MODEL_UNKNOWN;
++                }
++            }
+         default: (void)0;
+     }
+ 
+@@ -8420,6 +8472,38 @@ static bool llm_load_tensors(
+                     }
+ 
+                 } break;
++            case LLM_ARCH_SOLAR:
++                {
++                    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_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
++                    }
++
++                    for (int 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 * 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 * n_head, n_embd});
++
++                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
++
++                        layer.bskcn_tv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_BSKCN_TV, "weight"), {2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
++
++                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "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_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
++                    }
++                } break;
+             default:
+                 throw std::runtime_error("unknown architecture");
+         }
+@@ -15173,6 +15257,158 @@ struct llm_build_context {
+ 
+         return gf;
+     }
++
++    ggml_cgraph * build_solar() {
++        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
++
++        // mutable variable, needed during the last layer of the computation to skip unused tokens
++        int32_t n_tokens = this->n_tokens;
++
++        const int64_t n_embd_head = hparams.n_embd_head_v;
++        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
++        GGML_ASSERT(n_embd_head == hparams.n_rot);
++
++        struct ggml_tensor * cur;
++        struct ggml_tensor * inpL;
++
++        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
++
++        // 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();
++
++        struct ggml_tensor * bskcn_1;
++        struct ggml_tensor * bskcn_2;
++
++        for (int il = 0; il < n_layer; ++il) {
++            struct ggml_tensor * inpSA = inpL;
++
++            if (hparams.n_bskcn(0, il)) {
++                bskcn_1 = inpSA;
++            }
++
++            if (hparams.n_bskcn(1, il)) {
++                bskcn_2 = inpSA;
++            }
++
++            if (hparams.n_bskcn(2, il)) {
++                inpSA = ggml_add(
++                   ctx0,
++                   ggml_mul(ctx0, bskcn_1, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, 0)),
++                   ggml_mul(ctx0, inpSA, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, ggml_element_size(model.layers[il].bskcn_tv))));
++            }
++
++            if (hparams.n_bskcn(3, il)) {
++                inpSA = ggml_add(
++                   ctx0,
++                   ggml_mul(ctx0, bskcn_2, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, 0)),
++                   ggml_mul(ctx0, inpSA, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, ggml_element_size(model.layers[il].bskcn_tv))));
++            }
++
++            // 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
++            {
++                // rope freq factors for llama3; may return nullptr for llama2 and other models
++                struct ggml_tensor * rope_factors = build_rope_factors(il);
++
++                // compute Q and K and RoPE them
++                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
++                cb(Qcur, "Qcur", il);
++                if (model.layers[il].bq) {
++                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
++                    cb(Qcur, "Qcur", il);
++                }
++
++                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
++                cb(Kcur, "Kcur", il);
++                if (model.layers[il].bk) {
++                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
++                    cb(Kcur, "Kcur", il);
++                }
++
++                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
++                cb(Vcur, "Vcur", il);
++                if (model.layers[il].bv) {
++                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
++                    cb(Vcur, "Vcur", il);
++                }
++
++                Qcur = ggml_rope_ext(
++                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors,
++                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
++                    ext_factor, attn_factor, beta_fast, beta_slow
++                );
++                cb(Qcur, "Qcur", il);
++
++                Kcur = ggml_rope_ext(
++                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors,
++                    n_rot, 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, lctx, kv_self, gf,
++                        model.layers[il].wo, model.layers[il].bo,
++                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
++            }
++
++            if (il == n_layer - 1) {
++                // skip computing output for unused tokens
++                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
++                n_tokens = n_outputs;
++                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
++                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
++            }
++
++            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
++            cb(ffn_inp, "ffn_inp", il);
++
++            // feed-forward network
++            cur = llm_build_norm(ctx0, ffn_inp, hparams,
++                    model.layers[il].ffn_norm, NULL,
++                    LLM_NORM_RMS, cb, il);
++            cb(cur, "ffn_norm", il);
++
++            cur = llm_build_ffn(ctx0, lctx, cur,
++                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
++                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
++                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
++                    NULL,
++                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
++            cb(cur, "ffn_out", il);
++
++            cur = ggml_add(ctx0, cur, ffn_inp);
++            cb(cur, "ffn_out", il);
++
++            cur = lctx.cvec.apply_to(ctx0, cur, il);
++            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 = llm_build_lora_mm(lctx, ctx0, model.output, cur);
++        cb(cur, "result_output", -1);
++
++        ggml_build_forward_expand(gf, cur);
++
++        return gf;
++    }
+ };
+ 
+ static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector<uint32_t> & ids) {
+@@ -15423,6 +15659,10 @@ static struct ggml_cgraph * llama_build_graph(
+             {
+                 result = llm.build_rwkv6();
+             } break;
++        case LLM_ARCH_SOLAR:
++            {
++                result = llm.build_solar();
++            } break;
+         default:
+             GGML_ABORT("fatal error");
+     }
+@@ -18503,6 +18743,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
+         case LLM_ARCH_ARCTIC:
+         case LLM_ARCH_DEEPSEEK2:
+         case LLM_ARCH_CHATGLM:
++        case LLM_ARCH_SOLAR:
+             return LLAMA_ROPE_TYPE_NORM;
+ 
+         // the pairs of head values are offset by n_rot/2
+-- 
+2.46.0
+

llm/patches/0009-mllama.patch

@@ -0,0 +1,693 @@
+From 52f526a86b6fdd50784678c02d8212edc2412a5b Mon Sep 17 00:00:00 2001
+From: jmorganca <jmorganca@gmail.com>
+Date: Tue, 24 Sep 2024 11:53:40 -0700
+Subject: [PATCH] add mllama support
+
+mllama adds cross-attention layers to the standard llama architecture
+it also requires a way to input a new tensor: cross_attention_state
+once per generation
+
+cross-attention layers don't change and so they are cached in the
+kv cache once per run
+
+remaining is to implement the cross attention mask
+---
+ include/llama.h |   4 +
+ src/llama.cpp   | 456 ++++++++++++++++++++++++++++++++++++++++++++++--
+ 2 files changed, 447 insertions(+), 13 deletions(-)
+
+diff --git a/include/llama.h b/include/llama.h
+index bfc37e88..792520cc 100644
+--- a/include/llama.h
++++ b/include/llama.h
+@@ -449,6 +449,10 @@ extern "C" {
+                      struct llama_model * model,
+             struct llama_context_params   params);
+ 
++    // TODO (jmorganca): this should most likely be passed in as part of a batch
++    // and not set on the context for all batches.
++    LLAMA_API void llama_set_cross_attn_state(struct llama_context * ctx, float * cross_attn_state);
++
+     // Frees all allocated memory
+     LLAMA_API void llama_free(struct llama_context * ctx);
+ 
+diff --git a/src/llama.cpp b/src/llama.cpp
+index b7771f53..cf70ea90 100644
+--- a/src/llama.cpp
++++ b/src/llama.cpp
+@@ -170,6 +170,7 @@ static std::string format(const char * fmt, ...) {
+ 
+ enum llm_arch {
+     LLM_ARCH_LLAMA,
++    LLM_ARCH_MLLAMA,
+     LLM_ARCH_FALCON,
+     LLM_ARCH_BAICHUAN,
+     LLM_ARCH_GROK,
+@@ -219,6 +220,7 @@ enum llm_arch {
+ 
+ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
+     { LLM_ARCH_LLAMA,           "llama"        },
++    { LLM_ARCH_MLLAMA,          "mllama"       },
+     { LLM_ARCH_FALCON,          "falcon"       },
+     { LLM_ARCH_GROK,            "grok"         },
+     { LLM_ARCH_GPT2,            "gpt2"         },
+@@ -317,6 +319,7 @@ enum llm_kv {
+     LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
+     LLM_KV_ATTENTION_SLIDING_WINDOW,
+     LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,
++    LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS,
+ 
+     LLM_KV_ROPE_DIMENSION_COUNT,
+     LLM_KV_ROPE_FREQ_BASE,
+@@ -422,6 +425,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
+     { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count"   },
+     { LLM_KV_ATTENTION_SLIDING_WINDOW,         "%s.attention.sliding_window"           },
+     { LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,  "%s.attention.block_skip_connection.%d" },
++    { LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS, "%s.attention.cross_attention_layers"   },
+ 
+     { LLM_KV_ROPE_DIMENSION_COUNT,          "%s.rope.dimension_count"                 },
+     { LLM_KV_ROPE_FREQ_BASE,                "%s.rope.freq_base"                       },
+@@ -594,6 +598,14 @@ enum llm_tensor {
+     LLM_TENSOR_ENC_FFN_UP,
+     LLM_TENSOR_ENC_OUTPUT_NORM,
+     LLM_TENSOR_BSKCN_TV,
++    LLM_TENSOR_CROSS_ATTN_K_NORM,
++    LLM_TENSOR_CROSS_ATTN_K_PROJ,
++    LLM_TENSOR_CROSS_ATTN_O_PROJ,
++    LLM_TENSOR_CROSS_ATTN_Q_NORM,
++    LLM_TENSOR_CROSS_ATTN_Q_PROJ,
++    LLM_TENSOR_CROSS_ATTN_V_PROJ,
++    LLM_TENSOR_CROSS_ATTN_ATTN_GATE,
++    LLM_TENSOR_CROSS_ATTN_MLP_GATE,
+ };
+ 
+ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES = {
+@@ -623,6 +635,40 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
+             { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+         },
+     },
++    {
++        LLM_ARCH_MLLAMA,
++        {
++            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
++            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
++            { LLM_TENSOR_OUTPUT,          "output" },
++            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
++            { 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_ROT_EMBD,   "blk.%d.attn_rot_embd" },
++            { LLM_TENSOR_FFN_GATE_INP,    "blk.%d.ffn_gate_inp" },
++            { 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_GATE_EXP,    "blk.%d.ffn_gate.%d" },
++            { LLM_TENSOR_FFN_DOWN_EXP,    "blk.%d.ffn_down.%d" },
++            { LLM_TENSOR_FFN_UP_EXP,      "blk.%d.ffn_up.%d" },
++            { LLM_TENSOR_FFN_GATE_EXPS,   "blk.%d.ffn_gate_exps" },
++            { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
++            { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
++            { LLM_TENSOR_CROSS_ATTN_K_NORM,    "blk.%d.cross_attn_k_norm" },
++            { LLM_TENSOR_CROSS_ATTN_K_PROJ,    "blk.%d.cross_attn_k_proj" },
++            { LLM_TENSOR_CROSS_ATTN_O_PROJ,    "blk.%d.cross_attn_o_proj" },
++            { LLM_TENSOR_CROSS_ATTN_Q_NORM,    "blk.%d.cross_attn_q_norm" },
++            { LLM_TENSOR_CROSS_ATTN_Q_PROJ,    "blk.%d.cross_attn_q_proj" },
++            { LLM_TENSOR_CROSS_ATTN_V_PROJ,    "blk.%d.cross_attn_v_proj" },
++            { LLM_TENSOR_CROSS_ATTN_ATTN_GATE, "blk.%d.cross_attn_attn_gate" },
++            { LLM_TENSOR_CROSS_ATTN_MLP_GATE,  "blk.%d.cross_attn_mlp_gate" },
++        },
++    },
+     {
+         LLM_ARCH_BAICHUAN,
+         {
+@@ -2267,6 +2313,7 @@ enum e_model {
+     MODEL_40B,
+     MODEL_65B,
+     MODEL_70B,
++    MODEL_90B,
+     MODEL_236B,
+     MODEL_314B,
+     MODEL_SMALL,
+@@ -2309,6 +2356,7 @@ struct llama_hparams {
+     std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;
+ 
+     std::array<std::array<uint32_t, LLAMA_MAX_LAYERS>, 4> n_bskcn_arr;
++    std::array<uint32_t, LLAMA_MAX_LAYERS> cross_attn_layers;
+ 
+     uint32_t n_layer_dense_lead = 0;
+     uint32_t n_lora_q = 0;
+@@ -2372,10 +2420,11 @@ struct llama_hparams {
+         if (this->n_expert      != other.n_expert)      return true;
+         if (this->n_expert_used != other.n_expert_used) return true;
+ 
+-        if (this->n_head_arr    != other.n_head_arr)    return true;
+-        if (this->n_head_kv_arr != other.n_head_kv_arr) return true;
+-        if (this->n_ff_arr      != other.n_ff_arr)      return true;
+-        if (this->n_bskcn_arr   != other.n_bskcn_arr)   return true;
++        if (this->n_head_arr        != other.n_head_arr)        return true;
++        if (this->n_head_kv_arr     != other.n_head_kv_arr)     return true;
++        if (this->n_ff_arr          != other.n_ff_arr)          return true;
++        if (this->n_bskcn_arr       != other.n_bskcn_arr)       return true;
++        if (this->cross_attn_layers != other.cross_attn_layers) return true;
+ 
+         if (this->n_rel_attn_bkts    != other.n_rel_attn_bkts)    return true;
+         if (this->n_layer_dense_lead != other.n_layer_dense_lead) return true;
+@@ -2490,6 +2539,10 @@ struct llama_hparams {
+ 
+         GGML_ABORT("fatal error");
+     }
++
++    bool cross_attention_layer(uint32_t il) const {
++        return std::find(cross_attn_layers.begin(), cross_attn_layers.end(), il) != cross_attn_layers.end();
++    }
+ };
+ 
+ static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");
+@@ -2672,6 +2725,16 @@ struct llama_layer {
+     struct ggml_tensor * ffn_down_scale;
+ 
+     struct ggml_tensor * bskcn_tv;
++
++    // cross attention
++    struct ggml_tensor * cross_attn_k_norm;
++    struct ggml_tensor * cross_attn_k_proj;
++    struct ggml_tensor * cross_attn_o_proj;
++    struct ggml_tensor * cross_attn_q_norm;
++    struct ggml_tensor * cross_attn_q_proj;
++    struct ggml_tensor * cross_attn_v_proj;
++    struct ggml_tensor * cross_attn_attn_gate;
++    struct ggml_tensor * cross_attn_mlp_gate;
+ };
+ 
+ // very similar to llama_batch,
+@@ -3317,6 +3380,12 @@ struct llama_context {
+     struct ggml_tensor * inp_pos_bucket;    // I32 [n_batch|n_kv, n_batch]
+     struct ggml_tensor * inp_embd_enc;      // F32 [n_embd, n_outputs_enc]
+     struct ggml_tensor * inp_KQ_mask_cross; // F32 [n_outputs_enc, n_batch]
++
++    // TODO (jmorganca): this should most likely be passed in as part of a batch
++    // and not set on the context for all batches.
++    float * cross_attn_state = nullptr;
++    bool cross_attn_state_first_pass = true;
++    struct ggml_tensor * inp_cross_attn_state; // F32 [4, n_embd, 1061]
+ };
+ 
+ struct llama_lora_weight {
+@@ -3543,6 +3612,18 @@ static bool llama_kv_cache_init(
+     cache.v_l.reserve(n_layer);
+ 
+     for (int i = 0; i < (int) n_layer; i++) {
++        // for cross attention layers
++        if (model.arch == LLM_ARCH_MLLAMA && hparams.cross_attention_layer(i)) {
++            struct ggml_context * ctx = offload ? ctx_map.at(model.buft_layer[i].buft) : cache.ctxs.front();
++            ggml_tensor * k = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hparams.n_embd_head_k, 6404, hparams.n_head_kv(i));
++            ggml_tensor * v = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hparams.n_embd_head_v, 6404, hparams.n_head_kv(i));
++            ggml_format_name(k, "cache_k_l%d", i);
++            ggml_format_name(v, "cache_v_l%d", i);
++            cache.k_l.push_back(k);
++            cache.v_l.push_back(v);
++            continue;
++        }
++
+         const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i) + hparams.n_embd_k_s();
+         const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i) + hparams.n_embd_v_s();
+ 
+@@ -5312,12 +5393,14 @@ static void llm_load_hparams(
+     }
+ 
+     // zero-out the per-layer hparams
+-    std::fill(hparams.n_head_arr.begin(),    hparams.n_head_arr.end(),    0);
+-    std::fill(hparams.n_head_kv_arr.begin(), hparams.n_head_kv_arr.end(), 0);
+-    std::fill(hparams.n_ff_arr.begin(),      hparams.n_ff_arr.end(),      0);
++    std::fill(hparams.n_head_arr.begin(),             hparams.n_head_arr.end(),        0);
++    std::fill(hparams.n_head_kv_arr.begin(),          hparams.n_head_kv_arr.end(),     0);
++    std::fill(hparams.n_ff_arr.begin(),               hparams.n_ff_arr.end(),          0);
++    std::fill(hparams.cross_attn_layers.begin(),      hparams.cross_attn_layers.end(), -1);
+ 
+-    ml.get_key_or_arr(LLM_KV_FEED_FORWARD_LENGTH,  hparams.n_ff_arr,   hparams.n_layer);
+-    ml.get_key_or_arr(LLM_KV_ATTENTION_HEAD_COUNT, hparams.n_head_arr, hparams.n_layer);
++    ml.get_key_or_arr(LLM_KV_FEED_FORWARD_LENGTH,       hparams.n_ff_arr,          hparams.n_layer);
++    ml.get_key_or_arr(LLM_KV_ATTENTION_HEAD_COUNT,      hparams.n_head_arr,        hparams.n_layer);
++    ml.get_arr(LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS, hparams.cross_attn_layers, false);
+ 
+     // n_head_kv is optional, default to n_head
+     hparams.n_head_kv_arr = hparams.n_head_arr;
+@@ -5366,7 +5449,7 @@ static void llm_load_hparams(
+ 
+         ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot, false);
+ 
+-        if (model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_FALCON) {
++        if (model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_MLLAMA || model.arch == LLM_ARCH_FALCON) {
+             if (hparams.n_rot != hparams.n_embd_head_k) {
+                 throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd_head_k));
+             }
+@@ -5404,6 +5487,16 @@ static void llm_load_hparams(
+                     }
+                 }
+             } break;
++        case LLM_ARCH_MLLAMA:
++            {
++                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
++
++                switch (hparams.n_layer) {
++                    case 40: model.type = e_model::MODEL_11B; break;
++                    case 100: model.type = e_model::MODEL_90B; break;
++                    default: model.type = e_model::MODEL_UNKNOWN;
++                }
++            } break;
+         case LLM_ARCH_MINICPM:
+             {
+                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+@@ -6918,6 +7011,55 @@ static bool llm_load_tensors(
+                         }
+                     }
+                 } break;
++            case LLM_ARCH_MLLAMA:
++                {
++                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab+8});
++
++                    // output
++                    {
++                        model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
++                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
++
++                        // if output is NULL, init from the input tok embed
++                        if (model.output == NULL) {
++                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
++                        }
++                    }
++
++                    for (int 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];
++
++                        if (hparams.cross_attention_layer(i)) {
++                            layer.cross_attn_k_norm = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_K_NORM,   "weight", i), {128});
++                            layer.cross_attn_k_proj = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_K_PROJ,   "weight", i), {n_embd, 1024});
++                            layer.cross_attn_o_proj = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_O_PROJ,   "weight", i), {n_embd, n_embd});
++                            layer.cross_attn_q_norm = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_Q_NORM, "weight", i), {128});
++                            layer.cross_attn_q_proj = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_Q_PROJ, "weight", i), {n_embd, n_embd});
++                            layer.cross_attn_v_proj = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_V_PROJ, "weight", i), {n_embd, 1024});
++                            layer.cross_attn_attn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_ATTN_GATE, i), {1});
++                            layer.cross_attn_mlp_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_MLP_GATE, i), {1});
++                            layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
++                            layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, 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_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
++                        } else {
++                            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 * 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 * n_head, n_embd});
++                            layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
++                            layer.rope_freqs = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FREQS, "weight"), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
++                            layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "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_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
++                        }
++                    }
++                } break;
+             case LLM_ARCH_GROK:
+                 {
+                     if (n_expert == 0) {
+@@ -8678,7 +8820,7 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
+ 
+         if (model.vocab.type != LLAMA_VOCAB_TYPE_NONE &&
+             model.hparams.n_vocab != model.vocab.id_to_token.size()) {
+-            throw std::runtime_error("vocab size mismatch");
++            LLAMA_LOG_WARN("%s: vocab mismatch %u !- %zu ...\n", __func__, model.hparams.n_vocab, model.vocab.id_to_token.size());
+         }
+ 
+         if (params.vocab_only) {
+@@ -8759,7 +8901,7 @@ static struct ggml_tensor * llm_build_inp_embd(
+ 
+         inpL = ggml_get_rows(ctx, tok_embd, lctx.inp_tokens);
+     } else {
+-       lctx.inp_embd = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, batch.n_tokens);
++        lctx.inp_embd = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, batch.n_tokens);
+         inpL = lctx.inp_embd;
+         ggml_set_input(lctx.inp_embd);
+     }
+@@ -8769,6 +8911,22 @@ static struct ggml_tensor * llm_build_inp_embd(
+     return inpL;
+ }
+ 
++static struct ggml_tensor * llm_build_inp_cross_attn_state(
++        struct ggml_context * ctx,
++       struct llama_context & lctx,
++        const llama_hparams & hparams,
++         const llm_build_cb & cb) {
++    const int64_t n_embd = hparams.n_embd;
++
++    struct ggml_tensor * inpCAS;
++    lctx.inp_cross_attn_state = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, n_embd, 1601, 4);
++    cb(lctx.inp_cross_attn_state, "inp_cross_attn_state", -1);
++    ggml_set_input(lctx.inp_cross_attn_state);
++    inpCAS = lctx.inp_cross_attn_state;
++
++    return inpCAS;
++}
++
+ static void llm_build_kv_store(
+         struct ggml_context * ctx,
+         const llama_hparams & hparams,
+@@ -9743,6 +9901,7 @@ struct llm_build_context {
+         lctx.inp_pos_bucket    = nullptr;
+         lctx.inp_embd_enc      = nullptr;
+         lctx.inp_KQ_mask_cross = nullptr;
++        lctx.inp_cross_attn_state = nullptr;
+     }
+ 
+     void free() {
+@@ -10158,6 +10317,253 @@ struct llm_build_context {
+                 LLM_NORM_RMS, cb, -1);
+         cb(cur, "result_norm", -1);
+ 
++        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
++        cb(cur, "result_output", -1);
++
++        ggml_build_forward_expand(gf, cur);
++
++        return gf;
++    }
++
++    struct ggml_cgraph * build_mllama() {
++        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
++
++        // mutable variable, needed during the last layer of the computation to skip unused tokens
++        int32_t n_tokens = this->n_tokens;
++
++        const int64_t n_embd_head = hparams.n_embd_head_v;
++        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
++        GGML_ASSERT(n_embd_head == hparams.n_rot);
++ 
++        struct ggml_tensor * cur;
++        struct ggml_tensor * inpL;
++        struct ggml_tensor * inpCAS;
++
++        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
++        inpCAS = llm_build_inp_cross_attn_state(ctx0, lctx, hparams, cb);
++
++        // 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) {
++            struct ggml_tensor * inpSA = inpL;
++
++            // norm
++            cur = llm_build_norm(ctx0, inpL, hparams,
++                    model.layers[il].attn_norm, NULL,
++                    LLM_NORM_RMS, cb, il);
++            cb(cur, "attn_norm", il);
++
++            if (hparams.cross_attention_layer(il)) {
++                if (!lctx.cross_attn_state) {
++                    continue;
++                }
++
++                // cross attention layer
++                struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_q_proj, cur);
++                cb(Qcur, "Qcur", il);
++
++                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
++                cb(Qcur, "Qcur", il);
++
++                Qcur = ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
++                cb(Qcur, "Qcur", il);
++
++                // TODO: is this required?
++                Qcur = ggml_cont(ctx0, Qcur);
++                cb(Qcur, "Qcur", il);
++
++                Qcur = llm_build_norm(ctx0, Qcur, hparams, model.layers[il].cross_attn_q_norm, NULL, LLM_NORM_RMS, cb, il);
++                cb(Qcur, "Qcur", il);
++
++                struct ggml_tensor * Kcur;
++                if (lctx.cross_attn_state_first_pass) {
++                    Kcur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_k_proj, inpCAS);
++                    cb(Kcur, "Kcur", il);
++
++                    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, 6404);
++                    cb(Kcur, "Kcur", il);
++
++                    Kcur = ggml_permute(ctx0, Kcur, 0, 2, 1, 3);
++                    cb(Kcur, "Kcur", il);
++
++                    // TODO: is this required?
++                    Kcur = ggml_cont(ctx0, Kcur);
++                    cb(Kcur, "Kcur", il);
++
++                    Kcur = llm_build_norm(ctx0, Kcur, hparams, model.layers[il].cross_attn_k_norm, NULL, LLM_NORM_RMS, cb, il);
++                    cb(Kcur, "Kcur", il);
++
++                    ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, kv_self.k_l[il]));
++                } else {
++                    Kcur = ggml_view_tensor(ctx0, kv_self.k_l[il]);
++                    cb(Kcur, "Kcur (view)", il);
++                }
++
++                struct ggml_tensor * Vcur;
++                if (lctx.cross_attn_state_first_pass) {
++                    Vcur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_v_proj, inpCAS);
++                    cb(Vcur, "Vcur", il);
++
++                    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, 6404);
++                    cb(Vcur, "Vcur", il);
++
++                    Vcur = ggml_permute(ctx0, Vcur, 0, 2, 1, 3);
++                    cb(Vcur, "Vcur", il);
++
++                    ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, kv_self.v_l[il]));
++                } else {
++                    Vcur = ggml_view_tensor(ctx0, kv_self.v_l[il]);
++                    cb(Vcur, "Vcur (view)", il);
++                }
++
++                struct ggml_tensor * kq = ggml_mul_mat(ctx0, Kcur, Qcur);
++                cb(kq, "kq", il);
++
++                kq = ggml_scale_inplace(ctx0, kq, 1.0f/sqrtf(float(n_embd_head)));
++                cb(kq, "kq_scaled", il);
++
++                // TODO: apply causal masks
++                struct ggml_tensor * kq_soft_max = ggml_soft_max_inplace(ctx0, kq);
++                cb(kq_soft_max, "kq_soft_max", il);
++
++                Vcur = ggml_cont(ctx0, ggml_transpose(ctx0, Vcur));
++                cb(Vcur, "Vcur", il);
++
++                struct ggml_tensor * kqv = ggml_mul_mat(ctx0, Vcur, kq_soft_max);
++                cb(kqv, "kqv", il);
++
++                struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
++                cb(kqv_merged, "kqv_merged", il);
++
++                cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_head_v*n_head, n_tokens);
++                cb(cur, "kqv_merged_cont", il);
++
++                cur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_o_proj, cur);
++                cb(cur, "cur", il);
++
++                // TODO: do this in place once?
++                cur = ggml_mul(ctx0, cur, ggml_tanh(ctx0, model.layers[il].cross_attn_attn_gate));
++
++                struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
++                cb(ffn_inp, "ffn_inp", il);
++
++                // feed-forward network
++                cur = llm_build_norm(ctx0, ffn_inp, hparams,
++                        model.layers[il].ffn_norm, NULL,
++                        LLM_NORM_RMS, cb, il);
++                cb(cur, "ffn_norm", il);
++
++                cur = llm_build_ffn(ctx0, lctx, cur,
++                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
++                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
++                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
++                        NULL,
++                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
++                cb(cur, "ffn_out", il);
++
++                // TODO: do this inplace once?
++                cur = ggml_add_inplace(ctx0, ggml_mul_inplace(ctx0, cur, ggml_tanh(ctx0, model.layers[il].cross_attn_mlp_gate)), ffn_inp);
++                cb(cur, "ffn_out", il);
++
++                cur = lctx.cvec.apply_to(ctx0, cur, il);
++                cb(cur, "l_out", il);
++
++                // input for next layer
++                inpL = cur;
++            } else {
++                // self attention layer
++
++                // rope freq factors for llama3; may return nullptr for llama2 and other models
++                struct ggml_tensor * rope_factors = build_rope_factors(il);
++
++                // compute Q and K and RoPE them
++                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
++                cb(Qcur, "Qcur", il);
++                if (model.layers[il].bq) {
++                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
++                    cb(Qcur, "Qcur", il);
++                }
++
++                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
++                cb(Kcur, "Kcur", il);
++                if (model.layers[il].bk) {
++                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
++                    cb(Kcur, "Kcur", il);
++                }
++
++                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
++                cb(Vcur, "Vcur", il);
++                if (model.layers[il].bv) {
++                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
++                    cb(Vcur, "Vcur", il);
++                }
++
++                Qcur = ggml_rope_ext(
++                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors,
++                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
++                    ext_factor, attn_factor, beta_fast, beta_slow
++                );
++                cb(Qcur, "Qcur", il);
++
++                Kcur = ggml_rope_ext(
++                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors,
++                    n_rot, 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, lctx, kv_self, gf,
++                        model.layers[il].wo, model.layers[il].bo,
++                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
++
++
++                if (il == n_layer - 1) {
++                    // skip computing output for unused tokens
++                    struct ggml_tensor * inp_out_ids = build_inp_out_ids();
++                    n_tokens = n_outputs;
++                    cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
++                    inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
++                }
++
++                struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
++                cb(ffn_inp, "ffn_inp", il);
++
++                // feed-forward network
++                cur = llm_build_norm(ctx0, ffn_inp, hparams,
++                        model.layers[il].ffn_norm, NULL,
++                        LLM_NORM_RMS, cb, il);
++                cb(cur, "ffn_norm", il);
++
++                cur = llm_build_ffn(ctx0, lctx, cur,
++                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
++                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
++                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
++                        NULL,
++                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
++                cb(cur, "ffn_out", il);
++
++                cur = ggml_add(ctx0, cur, ffn_inp);
++                cb(cur, "ffn_out", il);
++
++                cur = lctx.cvec.apply_to(ctx0, cur, il);
++                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 = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+         cb(cur, "result_output", -1);
+@@ -15493,6 +15899,10 @@ static struct ggml_cgraph * llama_build_graph(
+             {
+                 result = llm.build_llama();
+             } break;
++        case LLM_ARCH_MLLAMA:
++            {
++                result = llm.build_mllama();
++            } break;
+         case LLM_ARCH_BAICHUAN:
+             {
+                 result = llm.build_baichuan();
+@@ -15753,6 +16163,14 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
+         ggml_backend_tensor_set(lctx.inp_pos, batch.pos, 0, n_tokens*ggml_element_size(lctx.inp_pos));
+     }
+ 
++    // TODO (jmorganca): this might copy a lot of data on every request of a
++    // single generation even though it doesn't change, so we should
++    // find a way to not set this more than one time per image
++    if (lctx.inp_cross_attn_state &&
++        lctx.inp_cross_attn_state->buffer) {
++        ggml_backend_tensor_set(lctx.inp_cross_attn_state, lctx.cross_attn_state, 0, hparams.n_embd * 1601 * 4 * ggml_element_size(lctx.inp_cross_attn_state));
++    }
++
+     if (hparams.causal_attn || cparams.pooling_type == LLAMA_POOLING_TYPE_NONE) {
+         GGML_ASSERT(lctx.inp_out_ids && "every model that can must skip unused outputs");
+         const int64_t n_tokens = batch.n_tokens;
+@@ -16430,6 +16848,10 @@ static int llama_decode_internal(
+ 
+         llama_set_inputs(lctx, ubatch);
+ 
++        // TODO: replace with something better to find out if its
++        // our first actual pass
++        lctx.cross_attn_state_first_pass = false;
++
+         llama_graph_compute(lctx, gf, n_threads, threadpool);
+ 
+         // update the kv ring buffer
+@@ -17586,7 +18008,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
+         if (llama_model_has_encoder(&model)) {
+             n_attn_layer *= 3;
+         }
+-        GGML_ASSERT((qs.n_attention_wv == n_attn_layer) && "n_attention_wv is unexpected");
++        if (qs.n_attention_wv != n_attn_layer) {
++            LLAMA_LOG_WARN("%s: n_attention_wv is unexpected, expected: %d, found: %d\n", __func__, n_attn_layer, qs.n_attention_wv);
++        }
+     }
+ 
+     size_t total_size_org = 0;
+@@ -18681,6 +19105,11 @@ struct llama_context * llama_new_context_with_model(
+     return ctx;
+ }
+ 
++void llama_set_cross_attn_state(struct llama_context * ctx, float * cross_attn_state) {
++    ctx->cross_attn_state_first_pass = true;
++    ctx->cross_attn_state = cross_attn_state;
++}
++
+ void llama_free(struct llama_context * ctx) {
+     delete ctx;
+ }
+@@ -18731,6 +19160,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
+ 
+         // use what we call a normal RoPE, operating on pairs of consecutive head values
+         case LLM_ARCH_LLAMA:
++        case LLM_ARCH_MLLAMA:
+         case LLM_ARCH_BAICHUAN:
+         case LLM_ARCH_STARCODER:
+         case LLM_ARCH_PLAMO:
+-- 
+2.39.3 (Apple Git-146)
+

macapp/src/app.tsx

@@ -19,7 +19,7 @@ export default function () {
   const [step, setStep] = useState<Step>(Step.WELCOME)
   const [commandCopied, setCommandCopied] = useState<boolean>(false)
 
-  const command = 'ollama run llama3.1'
+  const command = 'ollama run llama3.2'
 
   return (
     <div className='drag'>

readline/errors.go

@@ -4,10 +4,7 @@ import (
 	"errors"
 )
 
-var (
-	ErrInterrupt       = errors.New("Interrupt")
-	ErrNewLineDetected = errors.New("new line detected")
-)
+var ErrInterrupt = errors.New("Interrupt")
 
 type InterruptError struct {
 	Line []rune

readline/readline.go

@@ -225,9 +225,6 @@ func (i *Instance) Readline() (string, error) {
 			buf.MoveToEnd()
 			fmt.Println()
 
-			if r == CharCtrlJ {
-				return output, ErrNewLineDetected
-			}
 			return output, nil
 		default:
 			if metaDel {

scripts/build_windows.ps1

@@ -7,12 +7,22 @@
 $ErrorActionPreference = "Stop"
 
 function checkEnv() {
-    $script:ARCH = $Env:PROCESSOR_ARCHITECTURE.ToLower()
-    $script:TARGET_ARCH=$Env:PROCESSOR_ARCHITECTURE.ToLower()
+    if ($null -ne $env:ARCH ) {
+        $script:ARCH = $env:ARCH
+    } else {
+        $arch=([System.Runtime.InteropServices.RuntimeInformation]::OSArchitecture)
+        if ($null -ne $arch) {
+            $script:ARCH = ($arch.ToString().ToLower()).Replace("x64", "amd64")
+        } else {
+            write-host "WARNING: old powershell detected, assuming amd64 architecture - set `$env:ARCH to override"
+            $script:ARCH="amd64"
+        }
+    }
+    $script:TARGET_ARCH=$script:ARCH
     Write-host "Building for ${script:TARGET_ARCH}"
     write-host "Locating required tools and paths"
     $script:SRC_DIR=$PWD
-    if (!$env:VCToolsRedistDir) {
+    if ($null -eq $env:VCToolsRedistDir) {
         $MSVC_INSTALL=(Get-CimInstance MSFT_VSInstance -Namespace root/cimv2/vs)[0].InstallLocation
         $env:VCToolsRedistDir=(get-item "${MSVC_INSTALL}\VC\Redist\MSVC\*")[0]
     }
@@ -28,9 +38,12 @@ function checkEnv() {
         $script:CUDA_DIRS=$cudaList
     }
     
-    $script:INNO_SETUP_DIR=(get-item "C:\Program Files*\Inno Setup*\")[0]
+    $inoSetup=(get-item "C:\Program Files*\Inno Setup*\")
+    if ($inoSetup.length -gt 0) {
+        $script:INNO_SETUP_DIR=$inoSetup[0]
+    }
 
-    $script:DEPS_DIR="${script:SRC_DIR}\dist\windows-${script:TARGET_ARCH}"
+    $script:DIST_DIR="${script:SRC_DIR}\dist\windows-${script:TARGET_ARCH}"
     $env:CGO_ENABLED="1"
     Write-Output "Checking version"
     if (!$env:VERSION) {
@@ -67,7 +80,6 @@ function checkEnv() {
 
 
 function buildOllama() {
-    write-host "Building ollama CLI"
     if ($null -eq ${env:OLLAMA_SKIP_GENERATE}) {
         Remove-Item -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}"
 
@@ -75,15 +87,16 @@ function buildOllama() {
         #        which targets to build
 
         # Start by skipping CUDA to build everything else
-        pwsh -Command { $env:OLLAMA_SKIP_CUDA_GENERATE="1"; & go generate ./... }
+        write-host "Building ollama runners"
+        powershell -Command { $env:OLLAMA_SKIP_CUDA_GENERATE="1"; & go generate ./... }
         if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}    
 
         # Then skip everyhting else and build all the CUDA variants
         foreach ($env:CUDA_LIB_DIR in $script:CUDA_DIRS) {
-            write-host "Building CUDA ${env:CUDA_LIB_DIR}"
+            write-host "Building CUDA ${env:CUDA_LIB_DIR} runner"
 
             if ($env:CUDA_LIB_DIR.Contains("v12")) {
-                pwsh -Command {
+                powershell -Command {
                     $env:OLLAMA_SKIP_CUDA_GENERATE=""
                     $env:OLLAMA_SKIP_STATIC_GENERATE="1"
                     $env:OLLAMA_SKIP_CPU_GENERATE="1"
@@ -96,7 +109,7 @@ function buildOllama() {
                     & go generate ./...
                 }
             } else {
-                pwsh -Command {
+                powershell -Command {
                     $env:OLLAMA_SKIP_CUDA_GENERATE=""
                     $env:OLLAMA_SKIP_STATIC_GENERATE="1"
                     $env:OLLAMA_SKIP_CPU_GENERATE="1"
@@ -115,6 +128,7 @@ function buildOllama() {
     } else {
         write-host "Skipping generate step with OLLAMA_SKIP_GENERATE set"
     }
+    write-host "Building ollama CLI"
     & go build -trimpath -ldflags "-s -w -X=github.com/ollama/ollama/version.Version=$script:VERSION -X=github.com/ollama/ollama/server.mode=release" .
     if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
     if ("${env:KEY_CONTAINER}") {
@@ -130,34 +144,50 @@ function buildApp() {
     write-host "Building Ollama App"
     cd "${script:SRC_DIR}\app"
     & windres -l 0 -o ollama.syso ollama.rc
-    & go build -trimpath -ldflags "-s -w -H windowsgui -X=github.com/ollama/ollama/version.Version=$script:VERSION -X=github.com/ollama/ollama/server.mode=release" .
+    & go build -trimpath -ldflags "-s -w -H windowsgui -X=github.com/ollama/ollama/version.Version=$script:VERSION -X=github.com/ollama/ollama/server.mode=release" -o "${script:SRC_DIR}\dist\windows-${script:TARGET_ARCH}-app.exe" .
     if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
     if ("${env:KEY_CONTAINER}") {
         & "${script:SignTool}" sign /v /fd sha256 /t http://timestamp.digicert.com /f "${script:OLLAMA_CERT}" `
-            /csp "Google Cloud KMS Provider" /kc ${env:KEY_CONTAINER} app.exe
+            /csp "Google Cloud KMS Provider" /kc ${env:KEY_CONTAINER} "${script:SRC_DIR}\dist\windows-${script:TARGET_ARCH}-app.exe"
         if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
     }
 }
 
 function gatherDependencies() {
-    write-host "Gathering runtime dependencies"
+    if ($null -eq $env:VCToolsRedistDir) {
+        write-error "Unable to locate VC Install location - please use a Developer shell"
+        exit 1
+    }
+    write-host "Gathering runtime dependencies from $env:VCToolsRedistDir"
     cd "${script:SRC_DIR}"
-    md "${script:DEPS_DIR}\lib\ollama" -ea 0 > $null
+    md "${script:DIST_DIR}\lib\ollama" -ea 0 > $null
 
     # TODO - this varies based on host build system and MSVC version - drive from dumpbin output
     # currently works for Win11 + MSVC 2019 + Cuda V11
-    cp "${env:VCToolsRedistDir}\x64\Microsoft.VC*.CRT\msvcp140*.dll" "${script:DEPS_DIR}\lib\ollama\"
-    cp "${env:VCToolsRedistDir}\x64\Microsoft.VC*.CRT\vcruntime140.dll" "${script:DEPS_DIR}\lib\ollama\"
-    cp "${env:VCToolsRedistDir}\x64\Microsoft.VC*.CRT\vcruntime140_1.dll" "${script:DEPS_DIR}\lib\ollama\"
-    foreach ($part in $("runtime", "stdio", "filesystem", "math", "convert", "heap", "string", "time", "locale", "environment")) {
-        cp "$env:VCToolsRedistDir\..\..\..\Tools\Llvm\x64\bin\api-ms-win-crt-${part}*.dll" "${script:DEPS_DIR}\lib\ollama\"
+    if ($script:TARGET_ARCH -eq "amd64") {
+        $depArch="x64"
+    } else {
+        $depArch=$script:TARGET_ARCH
+    }
+    if ($depArch -eq "amd64") {
+        cp "${env:VCToolsRedistDir}\${depArch}\Microsoft.VC*.CRT\msvcp140*.dll" "${script:DIST_DIR}\lib\ollama\"
+        cp "${env:VCToolsRedistDir}\${depArch}\Microsoft.VC*.CRT\vcruntime140.dll" "${script:DIST_DIR}\lib\ollama\"
+        cp "${env:VCToolsRedistDir}\${depArch}\Microsoft.VC*.CRT\vcruntime140_1.dll" "${script:DIST_DIR}\lib\ollama\"
+        $llvmCrtDir="$env:VCToolsRedistDir\..\..\..\Tools\Llvm\${depArch}\bin"
+        foreach ($part in $("runtime", "stdio", "filesystem", "math", "convert", "heap", "string", "time", "locale", "environment")) {
+            write-host "cp ${llvmCrtDir}\api-ms-win-crt-${part}*.dll ${script:DIST_DIR}\lib\ollama\"
+            cp "${llvmCrtDir}\api-ms-win-crt-${part}*.dll" "${script:DIST_DIR}\lib\ollama\"
+        }
+    } else {
+        # Carying the dll's doesn't seem to work, so use the redist installer
+        copy-item -path "${env:VCToolsRedistDir}\vc_redist.arm64.exe" -destination "${script:DIST_DIR}" -verbose
     }
 
 
     cp "${script:SRC_DIR}\app\ollama_welcome.ps1" "${script:SRC_DIR}\dist\"
     if ("${env:KEY_CONTAINER}") {
         write-host "about to sign"
-        foreach ($file in (get-childitem "${script:DEPS_DIR}\lib\ollama\cu*.dll") + @("${script:SRC_DIR}\dist\ollama_welcome.ps1")){
+        foreach ($file in (get-childitem "${script:DIST_DIR}\lib\ollama\cu*.dll") + @("${script:SRC_DIR}\dist\ollama_welcome.ps1")){
             write-host "signing $file"
             & "${script:SignTool}" sign /v /fd sha256 /t http://timestamp.digicert.com /f "${script:OLLAMA_CERT}" `
                 /csp "Google Cloud KMS Provider" /kc ${env:KEY_CONTAINER} $file
@@ -167,6 +197,10 @@ function gatherDependencies() {
 }
 
 function buildInstaller() {
+    if ($null -eq ${script:INNO_SETUP_DIR}) {
+        write-host "Inno Setup not present, skipping installer build"
+        return
+    }
     write-host "Building Ollama Installer"
     cd "${script:SRC_DIR}\app"
     $env:PKG_VERSION=$script:PKG_VERSION
@@ -183,13 +217,20 @@ function distZip() {
     Compress-Archive -Path "${script:SRC_DIR}\dist\windows-${script:TARGET_ARCH}\*" -DestinationPath "${script:SRC_DIR}\dist\ollama-windows-${script:TARGET_ARCH}.zip" -Force
 }
 
+checkEnv
 try {
-    checkEnv
-    buildOllama
-    buildApp
-    gatherDependencies
-    buildInstaller
-    distZip
+    if ($($args.count) -eq 0) {
+        buildOllama
+        buildApp
+        gatherDependencies
+        buildInstaller
+        distZip
+    } else {
+        for ( $i = 0; $i -lt $args.count; $i++ ) {
+            write-host "performing $($args[$i])"
+            & $($args[$i])
+        } 
+    }
 } catch {
     write-host "Build Failed"
     write-host $_

server/model.go

@@ -272,6 +272,30 @@ func detectContentType(r io.Reader) (string, error) {
 	return "unknown", nil
 }
 
+func parseObjects(s string) []map[string]any {
+	var objs []map[string]any
+	for offset := 0; offset < len(s); {
+		var obj map[string]any
+		decoder := json.NewDecoder(strings.NewReader(s[offset:]))
+		if err := decoder.Decode(&obj); errors.Is(err, io.EOF) || errors.Is(err, io.ErrUnexpectedEOF) {
+			break
+		} else if syntax := &(json.SyntaxError{}); errors.As(err, &syntax) {
+			// skip over any syntax errors
+			offset += int(syntax.Offset)
+		} else if unmarshalType := &(json.UnmarshalTypeError{}); errors.As(err, &unmarshalType) {
+			// skip over any unmarshalable types
+			offset += int(unmarshalType.Offset)
+		} else if err != nil {
+			return nil
+		} else {
+			offset += int(decoder.InputOffset())
+			objs = append(objs, obj)
+		}
+	}
+
+	return objs
+}
+
 // parseToolCalls attempts to parse a JSON string into a slice of ToolCalls.
 // mxyng: this only really works if the input contains tool calls in some JSON format
 func (m *Model) parseToolCalls(s string) ([]api.ToolCall, bool) {
@@ -304,16 +328,14 @@ func (m *Model) parseToolCalls(s string) ([]api.ToolCall, bool) {
 		return nil, false
 	}
 
-	var kv map[string]any
-	// execute the subtree with placeholders to identify the keys
-	// trim any commands that might exist in the template
-	if err := json.Unmarshal(bytes.TrimSuffix(b.Bytes(), []byte(",")), &kv); err != nil {
+	templateObjects := parseObjects(b.String())
+	if len(templateObjects) == 0 {
 		return nil, false
 	}
 
 	// find the keys that correspond to the name and arguments fields
 	var name, arguments string
-	for k, v := range kv {
+	for k, v := range templateObjects[0] {
 		switch v.(type) {
 		case string:
 			name = k
@@ -326,43 +348,32 @@ func (m *Model) parseToolCalls(s string) ([]api.ToolCall, bool) {
 		return nil, false
 	}
 
-	var objs []map[string]any
-	for offset := 0; offset < len(s); {
-		var obj map[string]any
-		decoder := json.NewDecoder(strings.NewReader(s[offset:]))
-		if err := decoder.Decode(&obj); errors.Is(err, io.EOF) || errors.Is(err, io.ErrUnexpectedEOF) {
-			break
-		} else if syntax := &(json.SyntaxError{}); errors.As(err, &syntax) {
-			// skip over any syntax errors
-			offset += int(syntax.Offset)
-		} else if unmarshalType := &(json.UnmarshalTypeError{}); errors.As(err, &unmarshalType) {
-			// skip over any unmarshalable types
-			offset += int(unmarshalType.Offset)
-		} else if err != nil {
-			slog.Error("parseToolCalls", "error", err)
-			return nil, false
-		} else {
-			offset += int(decoder.InputOffset())
+	responseObjects := parseObjects(s)
+	if len(responseObjects) == 0 {
+		return nil, false
+	}
 
-			// collect all nested objects
-			var collect func(any) []map[string]any
-			collect = func(obj any) (all []map[string]any) {
-				switch o := obj.(type) {
-				case map[string]any:
-					all = append(all, o)
-					for _, v := range o {
-						all = append(all, collect(v)...)
-					}
-				case []any:
-					for _, v := range o {
-						all = append(all, collect(v)...)
-					}
-				}
-
-				return all
+	// collect all nested objects
+	var collect func(any) []map[string]any
+	collect = func(obj any) (all []map[string]any) {
+		switch o := obj.(type) {
+		case map[string]any:
+			all = append(all, o)
+			for _, v := range o {
+				all = append(all, collect(v)...)
+			}
+		case []any:
+			for _, v := range o {
+				all = append(all, collect(v)...)
 			}
-			objs = append(objs, collect(obj)...)
 		}
+
+		return all
+	}
+
+	var objs []map[string]any
+	for _, p := range responseObjects {
+		objs = append(objs, collect(p)...)
 	}
 
 	var toolCalls []api.ToolCall

server/model_test.go

@@ -69,6 +69,7 @@ The temperature in San Francisco, CA is 70°F and in Toronto, Canada is 20°C.`,
 {"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}}
 </tool_call>`, true},
 		{"xlam", `{"tool_calls": [{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}},{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}}]}`, true},
+		{"nemotron", `<toolcall>{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}},{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}}]} </toolcall>`, true},
 	}
 
 	var tools []api.Tool
@@ -217,3 +218,45 @@ func TestParseLayerFromCopy(t *testing.T) {
 		t.Fatalf("got %d != want 5", len(layers))
 	}
 }
+
+func TestParseObjects(t *testing.T) {
+	tests := []struct {
+		input string
+		want  []map[string]any
+	}{
+		{
+			input: `[{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}},{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}}]`,
+			want: []map[string]any{
+				{"name": "get_current_weather", "arguments": map[string]any{"format": "fahrenheit", "location": "San Francisco, CA"}},
+				{"name": "get_current_weather", "arguments": map[string]any{"format": "celsius", "location": "Toronto, Canada"}},
+			},
+		},
+		{
+			input: `<toolcall>{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}} </toolcall>`,
+			want: []map[string]any{
+				{"name": "get_current_weather", "arguments": map[string]any{"format": "fahrenheit", "location": "San Francisco, CA"}},
+			},
+		},
+		{
+			input: `<toolcall>{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}} </toolcall> <toolcall>{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, ON"}} </toolcall>`,
+			want: []map[string]any{
+				{"name": "get_current_weather", "arguments": map[string]any{"format": "fahrenheit", "location": "San Francisco, CA"}},
+				{"name": "get_current_weather", "arguments": map[string]any{"format": "celsius", "location": "Toronto, ON"}},
+			},
+		},
+		{
+			input: `{"name": "get_current_weather", "arguments": `,
+			want:  nil,
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.input, func(t *testing.T) {
+			got := parseObjects(tc.input)
+
+			if diff := cmp.Diff(got, tc.want); diff != "" {
+				t.Errorf("mismatch (-got +want):\n%s", diff)
+			}
+		})
+	}
+}

server/sched_test.go

@@ -354,7 +354,7 @@ func TestRequestsMultipleLoadedModels(t *testing.T) {
 }
 
 func TestGetRunner(t *testing.T) {
-	ctx, done := context.WithTimeout(context.Background(), 100*time.Millisecond)
+	ctx, done := context.WithTimeout(context.Background(), 200*time.Millisecond)
 	defer done()
 
 	a := newScenarioRequest(t, ctx, "ollama-model-1a", 10, &api.Duration{Duration: 2 * time.Millisecond})
@@ -395,7 +395,7 @@ func TestGetRunner(t *testing.T) {
 	slog.Info("c")
 	successCh1c, errCh1c := s.GetRunner(c.ctx, c.req.model, c.req.opts, c.req.sessionDuration)
 	// Starts in pending channel, then should be quickly processsed to return an error
-	time.Sleep(20 * time.Millisecond) // Long enough for the "a" model to expire and unload
+	time.Sleep(50 * time.Millisecond) // Long enough for the "a" model to expire and unload
 	require.Empty(t, successCh1c)
 	s.loadedMu.Lock()
 	require.Empty(t, s.loaded)

server/testdata/tools/nemotron.gotmpl

@@ -0,0 +1,33 @@
+{{- if (or .Tools .System) }}<extra_id_0>System
+{{ if .System }}{{ .System }}
+
+
+{{ end }}
+{{- if .Tools }}
+{{- range .Tools }}<tool> {{ . }} </tool>{{ end }}
+
+
+{{ end }}
+{{- end }}
+{{- range $i, $m := .Messages }}
+{{- $last := eq (len (slice $.Messages $i)) 1 -}}
+{{- if eq .Role "user" }}<extra_id_1>User
+{{ .Content }}
+{{- if $last }}
+<extra_id_1>Assistant
+{{- end }}
+{{ else if eq .Role "tool" }}<extra_id_1>Tool
+{{ .Content }}
+{{- if $last }}
+<extra_id_1>Assistant
+{{- end }}
+{{ else if eq .Role "assistant" }}<extra_id_1>Assistant
+{{- if .ToolCalls }}
+{{ range .ToolCalls }}<toolcall> {"name": "{{ .Function.Name }}", "arguments": {{ .Function.Arguments }}} </toolcall> {{ end }}
+{{ else }}
+{{ .Content }}
+{{- if not $last }}
+{{ end }}
+{{- end }}
+{{- end }}
+{{- end }}

server/testdata/tools/nemotron.out

@@ -0,0 +1,18 @@
+<extra_id_0>System
+You are a knowledgable assistant. You can answer questions and perform tasks.
+
+
+<tool> {"type":"function","function":{"name":"get_current_weather","description":"Get the current weather","parameters":{"type":"object","required":["location","format"],"properties":{"format":{"type":"string","description":"The temperature unit to use. Infer this from the users location.","enum":["celsius","fahrenheit"]},"location":{"type":"string","description":"The city and state, e.g. San Francisco, CA"}}}}} </tool>
+
+
+<extra_id_1>User
+What's the weather like today in Paris?
+<extra_id_1>Assistant
+<toolcall> {"name": "get_current_weather", "arguments": {"format":"celsius","location":"Paris, France"}} </toolcall> 
+<extra_id_1>Tool
+22
+<extra_id_1>Assistant
+The current temperature in Paris, France is 22 degrees Celsius.
+<extra_id_1>User
+What's the weather like today in San Francisco and Toronto?
+<extra_id_1>Assistant