build: Make target improvements (#7499)

* llama: wire up builtin runner

This adds a new entrypoint into the ollama CLI to run the cgo built runner.
On Mac arm64, this will have GPU support, but on all other platforms it will
be the lowest common denominator CPU build.  After we fully transition
to the new Go runners more tech-debt can be removed and we can stop building
the "default" runner via make and rely on the builtin always.

* build: Make target improvements

Add a few new targets and help for building locally.
This also adjusts the runner lookup to favor local builds, then
runners relative to the executable, and finally payloads.

* Support customized CPU flags for runners

This implements a simplified custom CPU flags pattern for the runners.
When built without overrides, the runner name contains the vector flag
we check for (AVX) to ensure we don't try to run on unsupported systems
and crash.  If the user builds a customized set, we omit the naming
scheme and don't check for compatibility.  This avoids checking
requirements at runtime, so that logic has been removed as well.  This
can be used to build GPU runners with no vector flags, or CPU/GPU
runners with additional flags (e.g. AVX512) enabled.

* Use relative paths

If the user checks out the repo in a path that contains spaces, make gets
really confused so use relative paths for everything in-repo to avoid breakage.

* Remove payloads from main binary

* install: clean up prior libraries

This removes support for v0.3.6 and older versions (before the tar bundle)
and ensures we clean up prior libraries before extracting the bundle(s).
Without this change, runners and dependent libraries could leak when we
update and lead to subtle runtime errors.

.github/workflows/release.yaml

@@ -85,13 +85,12 @@ jobs:
           import-module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
           Enter-VsDevShell -vsinstallpath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -skipautomaticlocation -DevCmdArguments '-arch=x64 -no_logo'
           if (!(gcc --version | select-string -quiet clang)) { throw "wrong gcc compiler detected - must be clang" }
-          make
+          make dist
         name: make
       - uses: actions/upload-artifact@v4
         with:
           name: generate-windows-cpu
           path: |
-            build/**/*
             dist/windows-amd64/**
 
   # ROCm generation step
@@ -143,13 +142,12 @@ jobs:
           import-module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
           Enter-VsDevShell -vsinstallpath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -skipautomaticlocation -DevCmdArguments '-arch=x64 -no_logo'
           if (!(gcc --version | select-string -quiet clang)) { throw "wrong gcc compiler detected - must be clang" }
-          make -C llama print-HIP_PATH print-HIP_LIB_DIR
-          make rocm
+          make help-runners
+          make dist_rocm
       - uses: actions/upload-artifact@v4
         with:
           name: generate-windows-rocm
           path: |
-            build/**/*
             dist/windows-amd64/**
 
   # CUDA generation step
@@ -226,12 +224,11 @@ jobs:
           import-module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
           Enter-VsDevShell -vsinstallpath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -skipautomaticlocation -DevCmdArguments '-arch=x64 -no_logo'
           if (!(gcc --version | select-string -quiet clang)) { throw "wrong gcc compiler detected - must be clang" }
-          make cuda_v$(($env:CUDA_PATH | split-path -leaf) -replace 'v(\d+).*', '$1')
+          make dist_cuda_v$(($env:CUDA_PATH | split-path -leaf) -replace 'v(\d+).*', '$1')
       - uses: actions/upload-artifact@v4
         with:
           name: generate-windows-cuda-${{ matrix.cuda.version }}
           path: |
-            build/**/*
             dist/windows-amd64/**
 
   # windows arm64 generate, go build, and zip file (no installer)

.gitignore

@@ -10,9 +10,6 @@ ollama
 .idea
 test_data
 *.crt
-llm/build
-build/*/*/*
-!build/**/placeholder
 llama/build
 __debug_bin*
 llama/vendor
\ No newline at end of file

Dockerfile

 ARG GOLANG_VERSION=1.22.8
-ARG CMAKE_VERSION=3.22.1
 ARG CUDA_VERSION_11=11.3.1
-ARG CUDA_V11_ARCHITECTURES="50;52;53;60;61;62;70;72;75;80;86"
 ARG CUDA_VERSION_12=12.4.0
-ARG CUDA_V12_ARCHITECTURES="60;61;62;70;72;75;80;86;87;89;90;90a"
 ARG ROCM_VERSION=6.1.2
 ARG JETPACK_6=r36.2.0
 ARG JETPACK_5=r35.4.1
@@ -15,24 +12,22 @@ ARG JETPACK_5=r35.4.1
 #
 ### Then incremental builds will be much faster in this container
 #
-# make -j 10 && go build -trimpath -o dist/linux-amd64/ollama .
+# make -j 10 dist
 #
 FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCM_VERSION}-complete AS unified-builder-amd64
-ARG CMAKE_VERSION
 ARG GOLANG_VERSION
 ARG CUDA_VERSION_11
 ARG CUDA_VERSION_12
 COPY ./scripts/rh_linux_deps.sh /
 ENV PATH /opt/rh/devtoolset-10/root/usr/bin:/usr/local/cuda/bin:$PATH
 ENV LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/cuda/lib64
-ENV LIBRARY_PATH=/usr/local/cuda/lib64/stubs:/opt/amdgpu/lib64
-RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
+RUN GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
 RUN yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo && \
     dnf clean all && \
     dnf install -y \
     zsh \
-    cuda-$(echo ${CUDA_VERSION_11} | cut -f1-2 -d. | sed -e "s/\./-/g") \
-    cuda-$(echo ${CUDA_VERSION_12} | cut -f1-2 -d. | sed -e "s/\./-/g")
+    cuda-toolkit-$(echo ${CUDA_VERSION_11} | cut -f1-2 -d. | sed -e "s/\./-/g") \
+    cuda-toolkit-$(echo ${CUDA_VERSION_12} | cut -f1-2 -d. | sed -e "s/\./-/g")
 # TODO intel oneapi goes here...
 ENV GOARCH amd64
 ENV CGO_ENABLED 1
@@ -46,12 +41,11 @@ ENTRYPOINT [ "zsh" ]
 # docker run --platform linux/arm64 --rm -it -v $(pwd):/go/src/github.com/ollama/ollama/ builder-arm64
 #
 FROM --platform=linux/arm64 rockylinux:8 AS unified-builder-arm64
-ARG CMAKE_VERSION
 ARG GOLANG_VERSION
 ARG CUDA_VERSION_11
 ARG CUDA_VERSION_12
 COPY ./scripts/rh_linux_deps.sh /
-RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
+RUN GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
 RUN yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/sbsa/cuda-rhel8.repo && \
     dnf config-manager --set-enabled appstream && \
     dnf clean all && \
@@ -62,37 +56,28 @@ RUN yum-config-manager --add-repo https://developer.download.nvidia.com/compute/
 ENV PATH /opt/rh/gcc-toolset-10/root/usr/bin:$PATH:/usr/local/cuda/bin
 ENV LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/cuda/lib64
 ENV LIBRARY_PATH=/usr/local/cuda/lib64/stubs:/opt/amdgpu/lib64
-ENV GOARCH amd64
+ENV GOARCH arm64
 ENV CGO_ENABLED 1
 WORKDIR /go/src/github.com/ollama/ollama/
 ENTRYPOINT [ "zsh" ]
 
-FROM --platform=linux/amd64 unified-builder-amd64 AS runners-amd64
+FROM --platform=linux/amd64 unified-builder-amd64 AS build-amd64
 COPY . .
 ARG OLLAMA_SKIP_CUDA_GENERATE
-ARG OLLAMA_SKIP_CUDA_11_GENERATE
-ARG OLLAMA_SKIP_CUDA_12_GENERATE
 ARG OLLAMA_SKIP_ROCM_GENERATE
-ARG CUDA_V11_ARCHITECTURES
-ARG CUDA_V12_ARCHITECTURES
 ARG OLLAMA_FAST_BUILD
 RUN --mount=type=cache,target=/root/.ccache \
     if grep "^flags" /proc/cpuinfo|grep avx>/dev/null; then \
-        make -j $(expr $(nproc) / 2 ) ; \
+        make -j $(expr $(nproc) / 2 ) dist ; \
     else \
-        make -j 5 ; \
+        make -j 5 dist ; \
+    fi
+RUN cd dist/linux-$GOARCH && \
+    tar -cf - . | pigz --best > ../ollama-linux-$GOARCH.tgz
+RUN if [ -z ${OLLAMA_SKIP_ROCM_GENERATE} ] ; then \
+    cd dist/linux-$GOARCH-rocm && \
+    tar -cf - . | pigz --best > ../ollama-linux-$GOARCH-rocm.tgz ;\
     fi
-
-FROM --platform=linux/arm64 unified-builder-arm64 AS runners-arm64
-COPY . .
-ARG OLLAMA_SKIP_CUDA_GENERATE
-ARG OLLAMA_SKIP_CUDA_11_GENERATE
-ARG OLLAMA_SKIP_CUDA_12_GENERATE
-ARG CUDA_V11_ARCHITECTURES
-ARG CUDA_V12_ARCHITECTURES
-ARG OLLAMA_FAST_BUILD
-RUN --mount=type=cache,target=/root/.ccache \
-    make -j 5
 
 # Jetsons need to be built in discrete stages
 FROM --platform=linux/arm64 nvcr.io/nvidia/l4t-jetpack:${JETPACK_5} AS runners-jetpack5-arm64
@@ -107,10 +92,9 @@ COPY . .
 ARG CGO_CFLAGS
 ENV GOARCH arm64
 RUN --mount=type=cache,target=/root/.ccache \
-    make -j 5 cuda_v11 \
+    make -j 5 dist_cuda_v11 \
         CUDA_ARCHITECTURES="72;87" \
         GPU_RUNNER_VARIANT=_jetpack5 \
-        CGO_EXTRA_LDFLAGS_LINUX=-L/usr/local/cuda/lib64/stubs \
         DIST_LIB_DIR=/go/src/github.com/ollama/ollama/dist/linux-arm64-jetpack5/lib/ollama \
         DIST_GPU_RUNNER_DEPS_DIR=/go/src/github.com/ollama/ollama/dist/linux-arm64-jetpack5/lib/ollama/cuda_jetpack5
 
@@ -126,69 +110,26 @@ COPY . .
 ARG CGO_CFLAGS
 ENV GOARCH arm64
 RUN --mount=type=cache,target=/root/.ccache \
-    make -j 5 cuda_v12 \
+    make -j 5 dist_cuda_v12 \
         CUDA_ARCHITECTURES="87" \
         GPU_RUNNER_VARIANT=_jetpack6 \
-        CGO_EXTRA_LDFLAGS_LINUX=-L/usr/local/cuda/lib64/stubs \
         DIST_LIB_DIR=/go/src/github.com/ollama/ollama/dist/linux-arm64-jetpack6/lib/ollama \
         DIST_GPU_RUNNER_DEPS_DIR=/go/src/github.com/ollama/ollama/dist/linux-arm64-jetpack6/lib/ollama/cuda_jetpack6
 
-
-# Intermediate stages used for ./scripts/build_linux.sh
-FROM --platform=linux/amd64 centos:7 AS builder-amd64
-ARG CMAKE_VERSION
-ARG GOLANG_VERSION
-COPY ./scripts/rh_linux_deps.sh /
-RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
-ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
-ENV CGO_ENABLED 1
-ENV GOARCH amd64
-WORKDIR /go/src/github.com/ollama/ollama
-
-FROM --platform=linux/amd64 builder-amd64 AS build-amd64
+FROM --platform=linux/arm64 unified-builder-arm64 AS build-arm64
 COPY . .
-COPY --from=runners-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
-COPY --from=runners-amd64 /go/src/github.com/ollama/ollama/build/ build/
-ARG GOFLAGS
-ARG CGO_CFLAGS
-ARG OLLAMA_SKIP_ROCM_GENERATE
+ARG OLLAMA_SKIP_CUDA_GENERATE
+ARG OLLAMA_FAST_BUILD
 RUN --mount=type=cache,target=/root/.ccache \
-    go build -trimpath -o dist/linux-amd64/bin/ollama .
-RUN cd dist/linux-$GOARCH && \
-    tar --exclude runners -cf - . | pigz --best > ../ollama-linux-$GOARCH.tgz
-RUN if [ -z ${OLLAMA_SKIP_ROCM_GENERATE} ] ; then \
-    cd dist/linux-$GOARCH-rocm && \
-    tar -cf - . | pigz --best > ../ollama-linux-$GOARCH-rocm.tgz ;\
-    fi
-
-FROM --platform=linux/arm64 rockylinux:8 AS builder-arm64
-ARG CMAKE_VERSION
-ARG GOLANG_VERSION
-COPY ./scripts/rh_linux_deps.sh /
-RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
-ENV PATH /opt/rh/gcc-toolset-10/root/usr/bin:$PATH
-ENV CGO_ENABLED 1
-ENV GOARCH arm64
-WORKDIR /go/src/github.com/ollama/ollama
-
-FROM --platform=linux/arm64 builder-arm64 AS build-arm64
-COPY . .
-COPY --from=runners-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
-COPY --from=runners-arm64 /go/src/github.com/ollama/ollama/build/ build/
+    make -j 5 dist
 COPY --from=runners-jetpack5-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
-COPY --from=runners-jetpack5-arm64 /go/src/github.com/ollama/ollama/build/ build/
 COPY --from=runners-jetpack6-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
-COPY --from=runners-jetpack6-arm64 /go/src/github.com/ollama/ollama/build/ build/
-ARG GOFLAGS
-ARG CGO_CFLAGS
-RUN --mount=type=cache,target=/root/.ccache \
-    go build -trimpath -o dist/linux-arm64/bin/ollama .
 RUN cd dist/linux-$GOARCH && \
-    tar --exclude runners -cf - . | pigz --best > ../ollama-linux-$GOARCH.tgz
+    tar -cf - . | pigz --best > ../ollama-linux-$GOARCH.tgz
 RUN cd dist/linux-$GOARCH-jetpack5 && \
-    tar --exclude runners -cf - . | pigz --best > ../ollama-linux-$GOARCH-jetpack5.tgz
+    tar -cf - . | pigz --best > ../ollama-linux-$GOARCH-jetpack5.tgz
 RUN cd dist/linux-$GOARCH-jetpack6 && \
-    tar --exclude runners -cf - . | pigz --best > ../ollama-linux-$GOARCH-jetpack6.tgz
+    tar -cf - . | pigz --best > ../ollama-linux-$GOARCH-jetpack6.tgz
 
 FROM --platform=linux/amd64 scratch AS dist-amd64
 COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/ollama-linux-*.tgz /
@@ -197,30 +138,13 @@ COPY --from=build-arm64 /go/src/github.com/ollama/ollama/dist/ollama-linux-*.tgz
 FROM dist-$TARGETARCH AS dist
 
 
-# Optimized container images do not cary nested payloads
-FROM --platform=linux/amd64 builder-amd64 AS container-build-amd64
-WORKDIR /go/src/github.com/ollama/ollama
-COPY . .
-ARG GOFLAGS
-ARG CGO_CFLAGS
-RUN --mount=type=cache,target=/root/.ccache \
-    go build -trimpath -o dist/linux-amd64/bin/ollama .
-
-FROM --platform=linux/arm64 builder-arm64 AS container-build-arm64
-WORKDIR /go/src/github.com/ollama/ollama
-COPY . .
-ARG GOFLAGS
-ARG CGO_CFLAGS
-RUN --mount=type=cache,target=/root/.ccache \
-    go build -trimpath -o dist/linux-arm64/bin/ollama .
-
 # For amd64 container images, filter out cuda/rocm to minimize size
-FROM runners-amd64 AS runners-cuda-amd64
+FROM build-amd64 AS runners-cuda-amd64
 RUN rm -rf \
     ./dist/linux-amd64/lib/ollama/libggml_hipblas.so \
     ./dist/linux-amd64/lib/ollama/runners/rocm*
 
-FROM runners-amd64 AS runners-rocm-amd64
+FROM build-amd64 AS runners-rocm-amd64
 RUN rm -rf \
     ./dist/linux-amd64/lib/ollama/libggml_cuda*.so \
     ./dist/linux-amd64/lib/ollama/libcu*.so* \
@@ -230,15 +154,15 @@ FROM --platform=linux/amd64 ubuntu:22.04 AS runtime-amd64
 RUN apt-get update && \
     apt-get install -y ca-certificates && \
     apt-get clean && rm -rf /var/lib/apt/lists/*
-COPY --from=container-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
+COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
 COPY --from=runners-cuda-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
 
 FROM --platform=linux/arm64 ubuntu:22.04 AS runtime-arm64
 RUN apt-get update && \
     apt-get install -y ca-certificates && \
     apt-get clean && rm -rf /var/lib/apt/lists/*
-COPY --from=container-build-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/bin/ /bin/
-COPY --from=runners-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/lib/ /lib/
+COPY --from=build-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/bin/ /bin/
+COPY --from=build-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/lib/ /lib/
 COPY --from=runners-jetpack5-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64-jetpack5/lib/ /lib/
 COPY --from=runners-jetpack6-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64-jetpack6/lib/ /lib/
 
@@ -251,7 +175,7 @@ COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64-rocm/l
 RUN apt-get update && \
     apt-get install -y ca-certificates && \
     apt-get clean && rm -rf /var/lib/apt/lists/*
-COPY --from=container-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
+COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
 COPY --from=runners-rocm-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
 
 EXPOSE 11434

Makefile

-GOALS := $(or $(MAKECMDGOALS),all)
-.PHONY: $(GOALS)
-$(GOALS):
-	$(MAKE) -C llama $@
\ No newline at end of file
+# top level makefile for Ollama
+include make/common-defs.make
+
+
+# Determine which if any GPU runners we should build
+include make/cuda-v11-defs.make
+include make/cuda-v12-defs.make
+include make/rocm-defs.make
+
+ifeq ($(CUSTOM_CPU_FLAGS),)
+ifneq ($(OS),darwin)
+ifeq ($(ARCH),amd64)
+	RUNNER_TARGETS=cpu
+endif
+endif
+# Without CUSTOM_CPU_FLAGS we default to build both v11 and v12 if present
+ifeq ($(OLLAMA_SKIP_CUDA_GENERATE),)
+ifneq ($(CUDA_11_COMPILER),)
+	RUNNER_TARGETS += cuda_v11
+endif
+ifneq ($(CUDA_12_COMPILER),)
+	RUNNER_TARGETS += cuda_v12
+endif
+endif
+else # CUSTOM_CPU_FLAGS is set, we'll build only the latest cuda version detected
+ifneq ($(CUDA_12_COMPILER),)
+	RUNNER_TARGETS += cuda_v12
+else ifneq ($(CUDA_11_COMPILER),)
+	RUNNER_TARGETS += cuda_v11
+endif
+endif
+
+ifeq ($(OLLAMA_SKIP_ROCM_GENERATE),)
+ifneq ($(HIP_COMPILER),)
+	RUNNER_TARGETS += rocm
+endif
+endif
+
+
+all: runners exe
+
+dist: $(addprefix dist_, $(RUNNER_TARGETS)) dist_exe
+
+dist_%:
+	@$(MAKE) --no-print-directory -f make/Makefile.$* dist
+
+runners: $(RUNNER_TARGETS)
+
+$(RUNNER_TARGETS):
+	@$(MAKE) --no-print-directory -f make/Makefile.$@
+
+exe dist_exe:
+	@$(MAKE) --no-print-directory -f make/Makefile.ollama $@
+
+help-sync apply-patches create-patches sync sync-clean:
+	@$(MAKE) --no-print-directory -f make/Makefile.sync $@
+
+test integration lint:
+	@$(MAKE) --no-print-directory -f make/Makefile.test $@
+
+clean:
+	rm -rf $(BUILD_DIR) $(DIST_LIB_DIR) $(OLLAMA_EXE) $(DIST_OLLAMA_EXE)
+	go clean -cache
+
+help:
+	@echo "The following make targets will help you build Ollama"
+	@echo ""
+	@echo "	make all   		# (default target) Build Ollama llm subprocess runners, and the primary ollama executable"
+	@echo "	make runners		# Build Ollama llm subprocess runners; after you may use 'go build .' to build the primary ollama exectuable"
+	@echo "	make <runner>		# Build specific runners. Enabled: '$(RUNNER_TARGETS)'"
+	@echo "	make dist		# Build the runners and primary ollama executable for distribution"
+	@echo "	make help-sync 		# Help information on vendor update targets"
+	@echo "	make help-runners 	# Help information on runner targets"
+	@echo ""
+	@echo "The following make targets will help you test Ollama"
+	@echo ""
+	@echo "	make test   		# Run unit tests"
+	@echo "	make integration	# Run integration tests.  You must 'make all' first"
+	@echo "	make lint   		# Run lint and style tests"
+	@echo ""
+	@echo "For more information see 'docs/development.md'"
+	@echo ""
+
+
+help-runners:
+	@echo "The following runners will be built based on discovered GPU libraries: '$(RUNNER_TARGETS)'"
+	@echo ""
+	@echo "GPU Runner CPU Flags: '$(GPU_RUNNER_CPU_FLAGS)'  (Override with CUSTOM_CPU_FLAGS)"
+	@echo ""
+	@echo "# CUDA_PATH sets the location where CUDA toolkits are present"
+	@echo "CUDA_PATH=$(CUDA_PATH)"
+	@echo "	CUDA_11_PATH=$(CUDA_11_PATH)"
+	@echo "	CUDA_11_COMPILER=$(CUDA_11_COMPILER)"
+	@echo "	CUDA_12_PATH=$(CUDA_12_PATH)"
+	@echo "	CUDA_12_COMPILER=$(CUDA_12_COMPILER)"
+	@echo ""
+	@echo "# HIP_PATH sets the location where the ROCm toolkit is present"
+	@echo "HIP_PATH=$(HIP_PATH)"
+	@echo "	HIP_COMPILER=$(HIP_COMPILER)"
+
+.PHONY: all exe dist help help-sync help-runners test integration lint runners clean $(RUNNER_TARGETS)
+
+# Handy debugging for make variables
+print-%:
+	@echo '$*=$($*)'

build/darwin/amd64/placeholder

-This is here to make sure the build/ directory exists for the go:embed command

build/darwin/arm64/placeholder

-This is here to make sure the build/ directory exists for the go:embed command

build/embed_darwin_amd64.go

-package build
-
-import "embed"
-
-// Darwin payloads separated by architecture to avoid duplicate payloads when cross compiling
-
-//go:embed darwin/amd64/*
-var EmbedFS embed.FS

build/embed_darwin_arm64.go

-package build
-
-import "embed"
-
-// Darwin payloads separated by architecture to avoid duplicate payloads when cross compiling
-
-//go:embed darwin/arm64/*
-var EmbedFS embed.FS

build/embed_linux.go

-package build
-
-import "embed"
-
-//go:embed linux/*
-var EmbedFS embed.FS

build/embed_unused.go

-//go:build !linux && !darwin
-
-package build
-
-import "embed"
-
-// unused on windows
-var EmbedFS embed.FS

build/linux/amd64/placeholder

-This is here to make sure the build/ directory exists for the go:embed command

build/linux/arm64/placeholder

-This is here to make sure the build/ directory exists for the go:embed command

cmd/cmd.go

@@ -37,6 +37,8 @@ import (
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
+	"github.com/ollama/ollama/llama"
+	"github.com/ollama/ollama/llama/runner"
 	"github.com/ollama/ollama/parser"
 	"github.com/ollama/ollama/progress"
 	"github.com/ollama/ollama/server"
@@ -1420,6 +1422,19 @@ func NewCLI() *cobra.Command {
 		RunE:    DeleteHandler,
 	}
 
+	runnerCmd := &cobra.Command{
+		Use:    "runner",
+		Short:  llama.PrintSystemInfo(),
+		Hidden: true,
+		RunE: func(cmd *cobra.Command, args []string) error {
+			return runner.Execute(os.Args[1:])
+		},
+		FParseErrWhitelist: cobra.FParseErrWhitelist{UnknownFlags: true},
+	}
+	runnerCmd.SetHelpFunc(func(cmd *cobra.Command, args []string) {
+		_ = runner.Execute(args[1:])
+	})
+
 	envVars := envconfig.AsMap()
 
 	envs := []envconfig.EnvVar{envVars["OLLAMA_HOST"]}
@@ -1476,6 +1491,7 @@ func NewCLI() *cobra.Command {
 		psCmd,
 		copyCmd,
 		deleteCmd,
+		runnerCmd,
 	)
 
 	return rootCmd

cmd/runner/main.go

+package main
+
+import (
+	"fmt"
+	"os"
+
+	"github.com/ollama/ollama/llama/runner"
+)
+
+func main() {
+	if err := runner.Execute(os.Args[1:]); err != nil {
+		fmt.Fprintf(os.Stderr, "error: %s\n", err)
+		os.Exit(1)
+	}
+}

discover/amd_linux.go

@@ -77,6 +77,7 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 
 	gfxOverride := envconfig.HsaOverrideGfxVersion()
 	var supported []string
+	depPaths := LibraryDirs()
 	libDir := ""
 
 	// The amdgpu driver always exposes the host CPU(s) first, but we have to skip them and subtract
@@ -300,8 +301,11 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 			})
 			continue
 		}
-
-		if int(major) < RocmComputeMin {
+		minVer, err := strconv.Atoi(RocmComputeMajorMin)
+		if err != nil {
+			slog.Error("invalid RocmComputeMajorMin setting", "value", RocmComputeMajorMin, "error", err)
+		}
+		if int(major) < minVer {
 			reason := fmt.Sprintf("amdgpu too old gfx%d%x%x", major, minor, patch)
 			slog.Warn(reason, "gpu", gpuID)
 			unsupportedGPUs = append(unsupportedGPUs, UnsupportedGPUInfo{
@@ -349,8 +353,9 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 				})
 				return nil, err
 			}
+			depPaths = append(depPaths, libDir)
 		}
-		gpuInfo.DependencyPath = []string{libDir}
+		gpuInfo.DependencyPath = depPaths
 
 		if gfxOverride == "" {
 			// Only load supported list once

discover/amd_windows.go

@@ -50,12 +50,14 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 		slog.Info(err.Error())
 		return nil, err
 	}
+	depPaths := LibraryDirs()
 	libDir, err := AMDValidateLibDir()
 	if err != nil {
 		err = fmt.Errorf("unable to verify rocm library: %w", err)
 		slog.Warn(err.Error())
 		return nil, err
 	}
+	depPaths = append(depPaths, libDir)
 
 	var supported []string
 	gfxOverride := envconfig.HsaOverrideGfxVersion()
@@ -111,7 +113,7 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 				UnreliableFreeMemory: true,
 
 				ID:             strconv.Itoa(i), // TODO this is probably wrong if we specify visible devices
-				DependencyPath: []string{libDir},
+				DependencyPath: depPaths,
 				MinimumMemory:  rocmMinimumMemory,
 				Name:           name,
 				Compute:        gfx,

discover/cpu_common.go

@@ -5,21 +5,8 @@ import (
 	"path/filepath"
 	"runtime"
 	"strings"
-
-	"golang.org/x/sys/cpu"
 )
 
-func GetCPUCapability() CPUCapability {
-	if cpu.X86.HasAVX2 {
-		return CPUCapabilityAVX2
-	}
-	if cpu.X86.HasAVX {
-		return CPUCapabilityAVX
-	}
-	// else LCD
-	return CPUCapabilityNone
-}
-
 func IsNUMA() bool {
 	if runtime.GOOS != "linux" {
 		// numa support in llama.cpp is linux only

discover/gpu.go

@@ -16,12 +16,14 @@ import (
 	"os"
 	"path/filepath"
 	"runtime"
+	"strconv"
 	"strings"
 	"sync"
 	"unsafe"
 
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
+	"github.com/ollama/ollama/runners"
 )
 
 type cudaHandles struct {
@@ -45,7 +47,6 @@ const (
 var (
 	gpuMutex      sync.Mutex
 	bootstrapped  bool
-	cpuCapability CPUCapability
 	cpus          []CPUInfo
 	cudaGPUs      []CudaGPUInfo
 	nvcudaLibPath string
@@ -64,9 +65,13 @@ var (
 )
 
 // With our current CUDA compile flags, older than 5.0 will not work properly
-var CudaComputeMin = [2]C.int{5, 0}
+// (string values used to allow ldflags overrides at build time)
+var (
+	CudaComputeMajorMin = "5"
+	CudaComputeMinorMin = "0"
+)
 
-var RocmComputeMin = 9
+var RocmComputeMajorMin = "9"
 
 // TODO find a better way to detect iGPU instead of minimum memory
 const IGPUMemLimit = 1 * format.GibiByte // 512G is what they typically report, so anything less than 1G must be iGPU
@@ -101,9 +106,9 @@ func initCudaHandles() *cudaHandles {
 		localAppData := os.Getenv("LOCALAPPDATA")
 		cudartMgmtPatterns = []string{filepath.Join(localAppData, "Programs", "Ollama", CudartMgmtName)}
 	}
-	libDir := LibraryDir()
-	if libDir != "" {
-		cudartMgmtPatterns = []string{filepath.Join(libDir, CudartMgmtName)}
+	libDirs := LibraryDirs()
+	for _, d := range libDirs {
+		cudartMgmtPatterns = append(cudartMgmtPatterns, filepath.Join(d, CudartMgmtName))
 	}
 	cudartMgmtPatterns = append(cudartMgmtPatterns, CudartGlobs...)
 
@@ -219,16 +224,23 @@ func GetGPUInfo() GpuInfoList {
 
 	if !bootstrapped {
 		slog.Info("looking for compatible GPUs")
+		cudaComputeMajorMin, err := strconv.Atoi(CudaComputeMajorMin)
+		if err != nil {
+			slog.Error("invalid CudaComputeMajorMin setting", "value", CudaComputeMajorMin, "error", err)
+		}
+		cudaComputeMinorMin, err := strconv.Atoi(CudaComputeMinorMin)
+		if err != nil {
+			slog.Error("invalid CudaComputeMinorMin setting", "value", CudaComputeMinorMin, "error", err)
+		}
 		bootstrapErrors = []error{}
 		needRefresh = false
-		cpuCapability = GetCPUCapability()
 		var memInfo C.mem_info_t
 
 		mem, err := GetCPUMem()
 		if err != nil {
 			slog.Warn("error looking up system memory", "error", err)
 		}
-		depPath := LibraryDir()
+		depPaths := LibraryDirs()
 		details, err := GetCPUDetails()
 		if err != nil {
 			slog.Warn("failed to lookup CPU details", "error", err)
@@ -238,24 +250,14 @@ func GetGPUInfo() GpuInfoList {
 				GpuInfo: GpuInfo{
 					memInfo:        mem,
 					Library:        "cpu",
-					Variant:        cpuCapability.String(),
+					Variant:        runners.GetCPUCapability().String(),
 					ID:             "0",
-					DependencyPath: []string{depPath},
+					DependencyPath: depPaths,
 				},
 				CPUs: details,
 			},
 		}
 
-		// Fallback to CPU mode if we're lacking required vector extensions on x86
-		if cpuCapability < GPURunnerCPUCapability && runtime.GOARCH == "amd64" {
-			err := fmt.Errorf("CPU does not have minimum vector extensions, GPU inference disabled.  Required:%s  Detected:%s", GPURunnerCPUCapability, cpuCapability)
-			slog.Warn(err.Error())
-			bootstrapErrors = append(bootstrapErrors, err)
-			bootstrapped = true
-			// No need to do any GPU discovery, since we can't run on them
-			return GpuInfoList{cpus[0].GpuInfo}
-		}
-
 		// Load ALL libraries
 		cHandles = initCudaHandles()
 
@@ -292,19 +294,23 @@ func GetGPUInfo() GpuInfoList {
 				gpuInfo.DriverMajor = driverMajor
 				gpuInfo.DriverMinor = driverMinor
 				variant := cudaVariant(gpuInfo)
-				if depPath != "" {
-					gpuInfo.DependencyPath = []string{depPath}
+				if depPaths != nil {
+					gpuInfo.DependencyPath = depPaths
 					// Check for variant specific directory
 					if variant != "" {
-						if _, err := os.Stat(filepath.Join(depPath, "cuda_"+variant)); err == nil {
-							gpuInfo.DependencyPath = []string{filepath.Join(depPath, "cuda_"+variant), depPath}
+						for _, d := range depPaths {
+							if _, err := os.Stat(filepath.Join(d, "cuda_"+variant)); err == nil {
+								// Put the variant directory first in the search path to avoid runtime linking to the wrong library
+								gpuInfo.DependencyPath = append([]string{filepath.Join(d, "cuda_"+variant)}, gpuInfo.DependencyPath...)
+								break
+							}
 						}
 					}
 				}
 				gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
 				gpuInfo.Variant = variant
 
-				if memInfo.major < CudaComputeMin[0] || (memInfo.major == CudaComputeMin[0] && memInfo.minor < CudaComputeMin[1]) {
+				if int(memInfo.major) < cudaComputeMajorMin || (int(memInfo.major) == cudaComputeMajorMin && int(memInfo.minor) < cudaComputeMinorMin) {
 					unsupportedGPUs = append(unsupportedGPUs,
 						UnsupportedGPUInfo{
 							GpuInfo: gpuInfo.GpuInfo,
@@ -370,7 +376,7 @@ func GetGPUInfo() GpuInfoList {
 						gpuInfo.FreeMemory = uint64(memInfo.free)
 						gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
 						gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
-						gpuInfo.DependencyPath = []string{depPath}
+						gpuInfo.DependencyPath = depPaths
 						oneapiGPUs = append(oneapiGPUs, gpuInfo)
 					}
 				}
@@ -385,6 +391,8 @@ func GetGPUInfo() GpuInfoList {
 		if len(cudaGPUs) == 0 && len(rocmGPUs) == 0 && len(oneapiGPUs) == 0 {
 			slog.Info("no compatible GPUs were discovered")
 		}
+
+		// TODO verify we have runners for the discovered GPUs, filter out any that aren't supported with good error messages
 	}
 
 	// For detected GPUs, load library if not loaded
@@ -509,7 +517,10 @@ func FindGPULibs(baseLibName string, defaultPatterns []string) []string {
 	slog.Debug("Searching for GPU library", "name", baseLibName)
 
 	// Start with our bundled libraries
-	patterns := []string{filepath.Join(LibraryDir(), baseLibName)}
+	patterns := []string{}
+	for _, d := range LibraryDirs() {
+		patterns = append(patterns, filepath.Join(d, baseLibName))
+	}
 
 	switch runtime.GOOS {
 	case "windows":
@@ -705,32 +716,26 @@ func (l GpuInfoList) GetVisibleDevicesEnv() (string, string) {
 	}
 }
 
-func LibraryDir() string {
-	// On Windows/linux we bundle the dependencies at the same level as the executable
+func LibraryDirs() []string {
+	// dependencies can exist wherever we found the runners (e.g. build tree for developers) and relative to the executable
+	// This can be simplified once we no longer carry runners as payloads
+	paths := []string{}
 	appExe, err := os.Executable()
 	if err != nil {
 		slog.Warn("failed to lookup executable path", "error", err)
+	} else {
+		appRelative := filepath.Join(filepath.Dir(appExe), envconfig.LibRelativeToExe(), "lib", "ollama")
+		if _, err := os.Stat(appRelative); err == nil {
+			paths = append(paths, appRelative)
+		}
 	}
-	cwd, err := os.Getwd()
+	rDir := runners.Locate()
 	if err != nil {
-		slog.Warn("failed to lookup working directory", "error", err)
-	}
-	// Scan for any of our dependeices, and pick first match
-	for _, root := range []string{filepath.Dir(appExe), filepath.Join(filepath.Dir(appExe), envconfig.LibRelativeToExe()), cwd} {
-		libDep := filepath.Join("lib", "ollama")
-		if _, err := os.Stat(filepath.Join(root, libDep)); err == nil {
-			return filepath.Join(root, libDep)
-		}
-		// Developer mode, local build
-		if _, err := os.Stat(filepath.Join(root, runtime.GOOS+"-"+runtime.GOARCH, libDep)); err == nil {
-			return filepath.Join(root, runtime.GOOS+"-"+runtime.GOARCH, libDep)
-		}
-		if _, err := os.Stat(filepath.Join(root, "dist", runtime.GOOS+"-"+runtime.GOARCH, libDep)); err == nil {
-			return filepath.Join(root, "dist", runtime.GOOS+"-"+runtime.GOARCH, libDep)
-		}
+		slog.Warn("unable to locate gpu dependency libraries", "error", err)
+	} else {
+		paths = append(paths, filepath.Dir(rDir))
 	}
-	slog.Warn("unable to locate gpu dependency libraries")
-	return ""
+	return paths
 }
 
 func GetSystemInfo() SystemInfo {

discover/gpu_darwin.go

@@ -15,6 +15,7 @@ import (
 	"syscall"
 
 	"github.com/ollama/ollama/format"
+	"github.com/ollama/ollama/runners"
 )
 
 const (
@@ -27,7 +28,7 @@ func GetGPUInfo() GpuInfoList {
 		return []GpuInfo{
 			{
 				Library: "cpu",
-				Variant: GetCPUCapability().String(),
+				Variant: runners.GetCPUCapability().String(),
 				memInfo: mem,
 			},
 		}
@@ -50,7 +51,7 @@ func GetCPUInfo() GpuInfoList {
 	return []GpuInfo{
 		{
 			Library: "cpu",
-			Variant: GetCPUCapability().String(),
+			Variant: runners.GetCPUCapability().String(),
 			memInfo: mem,
 		},
 	}

discover/types.go

@@ -5,6 +5,7 @@ import (
 	"log/slog"
 
 	"github.com/ollama/ollama/format"
+	"github.com/ollama/ollama/runners"
 )
 
 type memInfo struct {
@@ -47,6 +48,13 @@ type GpuInfo struct { // TODO better name maybe "InferenceProcessor"?
 	// TODO other performance capability info to help in scheduling decisions
 }
 
+func (gpu GpuInfo) RunnerName() string {
+	if gpu.Variant != "" {
+		return gpu.Library + "_" + gpu.Variant
+	}
+	return gpu.Library
+}
+
 type CPUInfo struct {
 	GpuInfo
 	CPUs []CPU
@@ -99,7 +107,7 @@ func (l GpuInfoList) ByLibrary() []GpuInfoList {
 	for _, info := range l {
 		found := false
 		requested := info.Library
-		if info.Variant != CPUCapabilityNone.String() {
+		if info.Variant != runners.CPUCapabilityNone.String() {
 			requested += "_" + info.Variant
 		}
 		for i, lib := range libs {
@@ -140,29 +148,6 @@ func (a ByFreeMemory) Len() int           { return len(a) }
 func (a ByFreeMemory) Swap(i, j int)      { a[i], a[j] = a[j], a[i] }
 func (a ByFreeMemory) Less(i, j int) bool { return a[i].FreeMemory < a[j].FreeMemory }
 
-type CPUCapability uint32
-
-// Override at build time when building base GPU runners
-var GPURunnerCPUCapability = CPUCapabilityAVX
-
-const (
-	CPUCapabilityNone CPUCapability = iota
-	CPUCapabilityAVX
-	CPUCapabilityAVX2
-	// TODO AVX512
-)
-
-func (c CPUCapability) String() string {
-	switch c {
-	case CPUCapabilityAVX:
-		return "avx"
-	case CPUCapabilityAVX2:
-		return "avx2"
-	default:
-		return "no vector extensions"
-	}
-}
-
 type SystemInfo struct {
 	System          CPUInfo              `json:"system"`
 	GPUs            []GpuInfo            `json:"gpus"`