Revamp ROCm support

This refines where we extract the LLM libraries to by adding a new
OLLAMA_HOME env var, that defaults to `~/.ollama` The logic was already
idempotenent, so this should speed up startups after the first time a
new release is deployed.  It also cleans up after itself.

We now build only a single ROCm version (latest major) on both windows
and linux.  Given the large size of ROCms tensor files, we split the
dependency out.  It's bundled into the installer on windows, and a
separate download on windows.  The linux install script is now smart and
detects the presence of AMD GPUs and looks to see if rocm v6 is already
present, and if not, then downloads our dependency tar file.

For Linux discovery, we now use sysfs and check each GPU against what
ROCm supports so we can degrade to CPU gracefully instead of having
llama.cpp+rocm assert/crash on us.  For Windows, we now use go's windows
dynamic library loading logic to access the amdhip64.dll APIs to query
the GPU information.

llm/generate/gen_linux.sh

@@ -179,17 +179,21 @@ fi
 
 if [ -d "${ROCM_PATH}" ]; then
     echo "ROCm libraries detected - building dynamic ROCm library"
-    if [ -f ${ROCM_PATH}/lib/librocm_smi64.so.? ]; then
-        ROCM_VARIANT=_v$(ls ${ROCM_PATH}/lib/librocm_smi64.so.? | cut -f3 -d. || true)
+    if [ -f ${ROCM_PATH}/lib/librocblas.so.*.*.????? ]; then
+        ROCM_VARIANT=_v$(ls ${ROCM_PATH}/lib/librocblas.so.*.*.????? | cut -f5 -d. || true)
     fi
     init_vars
     CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DLLAMA_HIPBLAS=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs) -DGPU_TARGETS=$(amdGPUs)"
     BUILD_DIR="${LLAMACPP_DIR}/build/linux/${ARCH}/rocm${ROCM_VARIANT}"
-    EXTRA_LIBS="-L${ROCM_PATH}/lib -L/opt/amdgpu/lib/x86_64-linux-gnu/ -Wl,-rpath,${ROCM_PATH}/lib,-rpath,/opt/amdgpu/lib/x86_64-linux-gnu/ -lhipblas -lrocblas -lamdhip64 -lrocsolver -lamd_comgr -lhsa-runtime64 -lrocsparse -ldrm -ldrm_amdgpu"
+    EXTRA_LIBS="-L${ROCM_PATH}/lib -L/opt/amdgpu/lib/x86_64-linux-gnu/ -Wl,-rpath,\$ORIGIN/../rocm/ -lhipblas -lrocblas -lamdhip64 -lrocsolver -lamd_comgr -lhsa-runtime64 -lrocsparse -ldrm -ldrm_amdgpu"
     build
 
-    # Note: the ROCM libs and runtime library files are too large to embed, so we depend on
-    #       them being present at runtime on the host
+    # Record the ROCM dependencies
+    rm -f "${BUILD_DIR}/lib/deps.txt"
+    touch "${BUILD_DIR}/lib/deps.txt"
+    for dep in $(ldd "${BUILD_DIR}/lib/libext_server.so" | grep "=>" | cut -f2 -d= | cut -f2 -d' ' | grep -e rocm -e amdgpu -e libtinfo ); do
+        echo "${dep}" >> "${BUILD_DIR}/lib/deps.txt"
+    done
     compress_libs
 fi
 

llm/generate/gen_windows.ps1

@@ -2,19 +2,52 @@
 
 $ErrorActionPreference = "Stop"
 
+function amdGPUs {
+    if ($env:AMDGPU_TARGETS) {
+        return $env:AMDGPU_TARGETS
+    }
+    # TODO - load from some common data file for linux + windows build consistency
+    $GPU_LIST = @(
+        "gfx900"
+        "gfx906:xnack-"
+        "gfx908:xnack-"
+        "gfx90a:xnack+"
+        "gfx90a:xnack-"
+        "gfx1010"
+        "gfx1012"
+        "gfx1030"
+        "gfx1100"
+        "gfx1101"
+        "gfx1102"
+    )
+    $GPU_LIST -join ';'
+}
+
 function init_vars {
+    # Verify the environment is a Developer Shell for MSVC 2019
+    write-host $env:VSINSTALLDIR
+    if (($env:VSINSTALLDIR -eq $null)) {
+        Write-Error "`r`nBUILD ERROR - YOUR DEVELOPMENT ENVIRONMENT IS NOT SET UP CORRECTLY`r`nTo build Ollama you must run from an MSVC Developer Shell`r`nSee .\docs\development.md for instructions to set up your dev environment"
+        exit 1
+    }
     $script:SRC_DIR = $(resolve-path "..\..\")
     $script:llamacppDir = "../llama.cpp"
-    $script:cmakeDefs = @("-DBUILD_SHARED_LIBS=on", "-DLLAMA_NATIVE=off",  "-A", "x64")
+    $script:cmakeDefs = @(
+        "-DBUILD_SHARED_LIBS=on",
+        "-DLLAMA_NATIVE=off"
+        )
     $script:cmakeTargets = @("ext_server")
     $script:ARCH = "amd64" # arm not yet supported.
     if ($env:CGO_CFLAGS -contains "-g") {
-        $script:cmakeDefs += @("-DCMAKE_VERBOSE_MAKEFILE=on", "-DLLAMA_SERVER_VERBOSE=on")
+        $script:cmakeDefs += @("-DCMAKE_VERBOSE_MAKEFILE=on", "-DLLAMA_SERVER_VERBOSE=on", "-DCMAKE_BUILD_TYPE=RelWithDebInfo")
         $script:config = "RelWithDebInfo"
     } else {
-        $script:cmakeDefs += @("-DLLAMA_SERVER_VERBOSE=off")
+        $script:cmakeDefs += @("-DLLAMA_SERVER_VERBOSE=off", "-DCMAKE_BUILD_TYPE=Release")
         $script:config = "Release"
     }
+    if ($null -ne $env:CMAKE_SYSTEM_VERSION) {
+        $script:cmakeDefs += @("-DCMAKE_SYSTEM_VERSION=${env:CMAKE_SYSTEM_VERSION}")
+    }
     # Try to find the CUDA dir
     if ($env:CUDA_LIB_DIR -eq $null) {
         $d=(get-command -ea 'silentlycontinue' nvcc).path
@@ -157,7 +190,7 @@ apply_patches
 $script:commonCpuDefs = @("-DCMAKE_POSITION_INDEPENDENT_CODE=on")
 
 init_vars
-$script:cmakeDefs = $script:commonCpuDefs + @("-DLLAMA_AVX=off", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
+$script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=off", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
 $script:buildDir="${script:llamacppDir}/build/windows/${script:ARCH}/cpu"
 write-host "Building LCD CPU"
 build
@@ -166,7 +199,7 @@ sign
 compress_libs
 
 init_vars
-$script:cmakeDefs = $script:commonCpuDefs + @("-DLLAMA_AVX=on", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
+$script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
 $script:buildDir="${script:llamacppDir}/build/windows/${script:ARCH}/cpu_avx"
 write-host "Building AVX CPU"
 build
@@ -175,7 +208,7 @@ sign
 compress_libs
 
 init_vars
-$script:cmakeDefs = $script:commonCpuDefs + @("-DLLAMA_AVX=on", "-DLLAMA_AVX2=on", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=on", "-DLLAMA_F16C=on") + $script:cmakeDefs
+$script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=on", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=on", "-DLLAMA_F16C=on") + $script:cmakeDefs
 $script:buildDir="${script:llamacppDir}/build/windows/${script:ARCH}/cpu_avx2"
 write-host "Building AVX2 CPU"
 build
@@ -192,18 +225,51 @@ if ($null -ne $script:CUDA_LIB_DIR) {
     }
     init_vars
     $script:buildDir="${script:llamacppDir}/build/windows/${script:ARCH}/cuda$script:CUDA_VARIANT"
-    $script:cmakeDefs += @("-DLLAMA_CUBLAS=ON", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=off", "-DCUDAToolkit_INCLUDE_DIR=$script:CUDA_INCLUDE_DIR", "-DCMAKE_CUDA_ARCHITECTURES=${script:CMAKE_CUDA_ARCHITECTURES}")
+    $script:cmakeDefs += @("-A", "x64", "-DLLAMA_CUBLAS=ON", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=off", "-DCUDAToolkit_INCLUDE_DIR=$script:CUDA_INCLUDE_DIR", "-DCMAKE_CUDA_ARCHITECTURES=${script:CMAKE_CUDA_ARCHITECTURES}")
+    write-host "Building CUDA"
     build
     install
     sign
     compress_libs
 }
-# TODO - actually implement ROCm support on windows
-$script:buildDir="${script:llamacppDir}/build/windows/${script:ARCH}/rocm"
 
-rm -ea 0 -recurse -force -path "${script:buildDir}/lib"
-md "${script:buildDir}/lib" -ea 0 > $null
-echo $null >> "${script:buildDir}/lib/.generated"
+if ($null -ne $env:HIP_PATH) {
+    $script:ROCM_VERSION=(get-item $env:HIP_PATH).Basename
+    if ($null -ne $script:ROCM_VERSION) {
+        $script:ROCM_VARIANT="_v"+$script:ROCM_VERSION
+    }
+
+    init_vars
+    $script:buildDir="${script:llamacppDir}/build/windows/${script:ARCH}/rocm$script:ROCM_VARIANT"
+    $script:cmakeDefs += @(
+        "-G", "Ninja", 
+        "-DCMAKE_C_COMPILER=clang.exe",
+        "-DCMAKE_CXX_COMPILER=clang++.exe",
+        "-DLLAMA_HIPBLAS=on",
+        "-DLLAMA_AVX=on",
+        "-DLLAMA_AVX2=off",
+        "-DCMAKE_POSITION_INDEPENDENT_CODE=on",
+        "-DAMDGPU_TARGETS=$(amdGPUs)",
+        "-DGPU_TARGETS=$(amdGPUs)"
+        )
+
+    # Make sure the ROCm binary dir is first in the path
+    $env:PATH="$env:HIP_PATH\bin;$env:VSINSTALLDIR\Common7\IDE\CommonExtensions\Microsoft\CMake\Ninja;$env:PATH"
+
+    # We have to clobber the LIB var from the developer shell for clang to work properly
+    $env:LIB=""
+
+    write-host "Building ROCm"
+    build
+    # Ninja doesn't prefix with config name
+    ${script:config}=""
+    install
+    if ($null -ne $script:DUMPBIN) {
+        & "$script:DUMPBIN" /dependents "${script:buildDir}/bin/${script:config}/ext_server.dll" | select-string ".dll"
+    }
+    sign
+    compress_libs
+}
 
 cleanup
 write-host "`ngo generate completed"

llm/llm.go

@@ -19,7 +19,7 @@ type LLM interface {
 	Close()
 }
 
-func New(workDir, model string, adapters, projectors []string, opts api.Options) (LLM, error) {
+func New(model string, adapters, projectors []string, opts api.Options) (LLM, error) {
 	if _, err := os.Stat(model); err != nil {
 		return nil, err
 	}
@@ -120,15 +120,15 @@ func New(workDir, model string, adapters, projectors []string, opts api.Options)
 
 	opts.RopeFrequencyBase = 0.0
 	opts.RopeFrequencyScale = 0.0
-	return newLlmServer(info, workDir, model, adapters, projectors, opts)
+	return newLlmServer(info, model, adapters, projectors, opts)
 }
 
 // Give any native cgo implementations an opportunity to initialize
-func Init(workdir string) error {
-	return nativeInit(workdir)
+func Init() error {
+	return nativeInit()
 }
 
-func newLlmServer(gpuInfo gpu.GpuInfo, workDir, model string, adapters, projectors []string, opts api.Options) (LLM, error) {
+func newLlmServer(gpuInfo gpu.GpuInfo, model string, adapters, projectors []string, opts api.Options) (LLM, error) {
 	dynLibs := getDynLibs(gpuInfo)
 
 	// Check to see if the user has requested a specific library instead of auto-detecting
@@ -147,7 +147,7 @@ func newLlmServer(gpuInfo gpu.GpuInfo, workDir, model string, adapters, projecto
 	_, err := os.Stat(dynLibs[0])
 	if err != nil {
 		slog.Info(fmt.Sprintf("%s has disappeared, reloading libraries", dynLibs[0]))
-		err = nativeInit(workDir)
+		err = nativeInit()
 		if err != nil {
 			return nil, err
 		}

llm/payload_common.go

@@ -103,10 +103,14 @@ func rocmDynLibPresent() bool {
 	return false
 }
 
-func nativeInit(workdir string) error {
+func nativeInit() error {
 	slog.Info("Extracting dynamic libraries...")
+	assetsDir, err := gpu.AssetsDir()
+	if err != nil {
+		return err
+	}
 	if runtime.GOOS == "darwin" {
-		err := extractPayloadFiles(workdir, "llama.cpp/ggml-metal.metal")
+		err := extractPayloadFiles(assetsDir, "llama.cpp/ggml-metal.metal")
 		if err != nil {
 			if err == payloadMissing {
 				// TODO perhaps consider this a hard failure on arm macs?
@@ -115,10 +119,10 @@ func nativeInit(workdir string) error {
 			}
 			return err
 		}
-		os.Setenv("GGML_METAL_PATH_RESOURCES", workdir)
+		os.Setenv("GGML_METAL_PATH_RESOURCES", assetsDir)
 	}
 
-	libs, err := extractDynamicLibs(workdir, "llama.cpp/build/*/*/*/lib/*")
+	libs, err := extractDynamicLibs(assetsDir, "llama.cpp/build/*/*/*/lib/*")
 	if err != nil {
 		if err == payloadMissing {
 			slog.Info(fmt.Sprintf("%s", payloadMissing))
@@ -149,17 +153,13 @@ func nativeInit(workdir string) error {
 	return nil
 }
 
-func extractDynamicLibs(workDir, glob string) ([]string, error) {
+func extractDynamicLibs(assetsDir, glob string) ([]string, error) {
 	files, err := fs.Glob(libEmbed, glob)
 	if err != nil || len(files) == 0 {
 		return nil, payloadMissing
 	}
 	libs := []string{}
 
-	// TODO consider making this idempotent with some sort of persistent directory (where we store models probably)
-	// and tracking by version so we don't reexpand the files every time
-	// Also maybe consider lazy loading only what is needed
-
 	g := new(errgroup.Group)
 	for _, file := range files {
 		pathComps := strings.Split(file, "/")
@@ -172,14 +172,14 @@ func extractDynamicLibs(workDir, glob string) ([]string, error) {
 		g.Go(func() error {
 			// llama.cpp/build/$OS/$GOARCH/$VARIANT/lib/$LIBRARY
 			// Include the variant in the path to avoid conflicts between multiple server libs
-			targetDir := filepath.Join(workDir, pathComps[pathComponentCount-3])
+			targetDir := filepath.Join(assetsDir, pathComps[pathComponentCount-3])
 			srcFile, err := libEmbed.Open(file)
 			if err != nil {
 				return fmt.Errorf("read payload %s: %v", file, err)
 			}
 			defer srcFile.Close()
 			if err := os.MkdirAll(targetDir, 0o755); err != nil {
-				return fmt.Errorf("create payload temp dir %s: %v", workDir, err)
+				return fmt.Errorf("create payload lib dir %s: %v", assetsDir, err)
 			}
 			src := io.Reader(srcFile)
 			filename := file
@@ -196,19 +196,13 @@ func extractDynamicLibs(workDir, glob string) ([]string, error) {
 				libs = append(libs, destFile)
 			}
 
-			_, err = os.Stat(destFile)
-			switch {
-			case errors.Is(err, os.ErrNotExist):
-				destFile, err := os.OpenFile(destFile, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o755)
-				if err != nil {
-					return fmt.Errorf("write payload %s: %v", file, err)
-				}
-				defer destFile.Close()
-				if _, err := io.Copy(destFile, src); err != nil {
-					return fmt.Errorf("copy payload %s: %v", file, err)
-				}
-			case err != nil:
-				return fmt.Errorf("stat payload %s: %v", file, err)
+			destFp, err := os.OpenFile(destFile, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o755)
+			if err != nil {
+				return fmt.Errorf("write payload %s: %v", file, err)
+			}
+			defer destFp.Close()
+			if _, err := io.Copy(destFp, src); err != nil {
+				return fmt.Errorf("copy payload %s: %v", file, err)
 			}
 			return nil
 		})
@@ -216,7 +210,7 @@ func extractDynamicLibs(workDir, glob string) ([]string, error) {
 	return libs, g.Wait()
 }
 
-func extractPayloadFiles(workDir, glob string) error {
+func extractPayloadFiles(assetsDir, glob string) error {
 	files, err := fs.Glob(libEmbed, glob)
 	if err != nil || len(files) == 0 {
 		return payloadMissing
@@ -228,8 +222,8 @@ func extractPayloadFiles(workDir, glob string) error {
 			return fmt.Errorf("read payload %s: %v", file, err)
 		}
 		defer srcFile.Close()
-		if err := os.MkdirAll(workDir, 0o755); err != nil {
-			return fmt.Errorf("create payload temp dir %s: %v", workDir, err)
+		if err := os.MkdirAll(assetsDir, 0o755); err != nil {
+			return fmt.Errorf("create payload lib dir %s: %v", assetsDir, err)
 		}
 		src := io.Reader(srcFile)
 		filename := file
@@ -241,20 +235,22 @@ func extractPayloadFiles(workDir, glob string) error {
 			filename = strings.TrimSuffix(filename, ".gz")
 		}
 
-		destFile := filepath.Join(workDir, filepath.Base(filename))
+		destFile := filepath.Join(assetsDir, filepath.Base(filename))
 		_, err = os.Stat(destFile)
 		switch {
 		case errors.Is(err, os.ErrNotExist):
-			destFile, err := os.OpenFile(destFile, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o755)
+			destFp, err := os.OpenFile(destFile, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o755)
 			if err != nil {
 				return fmt.Errorf("write payload %s: %v", file, err)
 			}
-			defer destFile.Close()
-			if _, err := io.Copy(destFile, src); err != nil {
+			defer destFp.Close()
+			if _, err := io.Copy(destFp, src); err != nil {
 				return fmt.Errorf("copy payload %s: %v", file, err)
 			}
 		case err != nil:
 			return fmt.Errorf("stat payload %s: %v", file, err)
+		case err == nil:
+			slog.Debug("payload already exists: " + destFile)
 		}
 	}
 	return nil

llm/payload_linux.go

@@ -4,5 +4,5 @@ import (
 	"embed"
 )
 
-//go:embed llama.cpp/build/linux/*/*/lib/*.so*
+//go:embed llama.cpp/build/linux/*/*/lib/*
 var libEmbed embed.FS

scripts/build_linux.sh

@@ -22,5 +22,6 @@ for TARGETARCH in ${BUILD_ARCH}; do
         .
     docker create --platform linux/$TARGETARCH --name builder-$TARGETARCH builder:$TARGETARCH
     docker cp builder-$TARGETARCH:/go/src/github.com/jmorganca/ollama/ollama ./dist/ollama-linux-$TARGETARCH
+    docker cp builder-$TARGETARCH:/go/src/github.com/jmorganca/ollama/dist/deps/ ./dist/
     docker rm builder-$TARGETARCH
 done

server/routes.go

@@ -66,8 +66,6 @@ var defaultSessionDuration = 5 * time.Minute
 
 // load a model into memory if it is not already loaded, it is up to the caller to lock loaded.mu before calling this function
 func load(c *gin.Context, model *Model, opts api.Options, sessionDuration time.Duration) error {
-	workDir := c.GetString("workDir")
-
 	needLoad := loaded.runner == nil || // is there a model loaded?
 		loaded.ModelPath != model.ModelPath || // has the base model changed?
 		!reflect.DeepEqual(loaded.AdapterPaths, model.AdapterPaths) || // have the adapters changed?
@@ -82,7 +80,7 @@ func load(c *gin.Context, model *Model, opts api.Options, sessionDuration time.D
 			loaded.Options = nil
 		}
 
-		llmRunner, err := llm.New(workDir, model.ModelPath, model.AdapterPaths, model.ProjectorPaths, opts)
+		llmRunner, err := llm.New(model.ModelPath, model.AdapterPaths, model.ProjectorPaths, opts)
 		if err != nil {
 			// some older models are not compatible with newer versions of llama.cpp
 			// show a generalized compatibility error until there is a better way to
@@ -1035,7 +1033,7 @@ func Serve(ln net.Listener) error {
 		os.Exit(0)
 	}()
 
-	if err := llm.Init(s.WorkDir); err != nil {
+	if err := llm.Init(); err != nil {
 		return fmt.Errorf("unable to initialize llm library %w", err)
 	}
 	if runtime.GOOS == "linux" { // TODO - windows too