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.
2025-12-11 00:07:07 +00:00 · 2024-12-10 09:47:19 -08:00
parent 63269668c0
commit 4879a234c4
58 changed files with 877 additions and 1168 deletions
--- a/make/Makefile.cpu
+++ b/make/Makefile.cpu
@@ -0,0 +1,40 @@
+# Build the discrete cpu runner(s) for the platform which do not rely on 3rd party GPU libraries
+
+include make/common-defs.make
+
+CPU_GOFLAGS="-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=$(VERSION)\" \"-X=github.com/ollama/ollama/llama.CpuFeatures=$(subst $(space),$(comma),$(TARGET_CPU_FLAGS))\" $(TARGET_LDFLAGS)"
+ifeq ($(ARCH),amd64)
+ifeq ($(origin CUSTOM_CPU_FLAGS),undefined)
+	RUNNERS = cpu_avx cpu_avx2
+endif
+endif
+
+DIST_RUNNERS = $(addprefix $(RUNNERS_DIST_DIR)/,$(addsuffix /ollama_llama_server$(EXE_EXT),$(RUNNERS)))
+BUILD_RUNNERS = $(addprefix $(RUNNERS_BUILD_DIR)/,$(addsuffix /ollama_llama_server$(EXE_EXT),$(RUNNERS)))
+
+cpu: $(BUILD_RUNNERS)
+
+dist: $(DIST_RUNNERS)
+
+$(RUNNERS_BUILD_DIR)/cpu_avx/ollama_llama_server$(EXE_EXT): TARGET_CPU_FLAGS="avx"
+$(RUNNERS_BUILD_DIR)/cpu_avx/ollama_llama_server$(EXE_EXT): ./llama/*.go ./llama/runner/*.go $(COMMON_SRCS) $(COMMON_HDRS)
+	@-mkdir -p $(dir $@)
+	GOARCH=$(ARCH) go build -buildmode=pie $(CPU_GOFLAGS) -trimpath -tags $(subst $(space),$(comma),$(TARGET_CPU_FLAGS)) -o $@ ./cmd/runner
+
+$(RUNNERS_BUILD_DIR)/cpu_avx2/ollama_llama_server$(EXE_EXT): TARGET_CPU_FLAGS="avx avx2"
+$(RUNNERS_BUILD_DIR)/cpu_avx2/ollama_llama_server$(EXE_EXT): ./llama/*.go ./llama/runner/*.go $(COMMON_SRCS) $(COMMON_HDRS)
+	@-mkdir -p $(dir $@)
+	GOARCH=$(ARCH) go build -buildmode=pie $(CPU_GOFLAGS) -trimpath -tags $(subst $(space),$(comma),$(TARGET_CPU_FLAGS)) -o $@ ./cmd/runner
+
+$(RUNNERS_DIST_DIR)/%: $(RUNNERS_BUILD_DIR)/%
+	@-mkdir -p $(dir $@)
+	cp $< $@
+
+clean:
+	rm -f $(BUILD_RUNNERS) $(DIST_RUNNERS)
+
+.PHONY: clean cpu dist
+
+# Handy debugging for make variables
+print-%:
+	@echo '$*=$($*)'
--- a/make/Makefile.cuda_v11
+++ b/make/Makefile.cuda_v11
@@ -0,0 +1,13 @@
+# Build rules for CUDA v11 runner
+
+include make/common-defs.make
+include make/cuda-v11-defs.make
+
+GPU_RUNNER_VARIANT := _v11
+GPU_COMPILER=$(CUDA_11_COMPILER)
+CUDA_ARCHITECTURES?=50;52;53;60;61;62;70;72;75;80;86
+GPU_LIB_DIR = $(CUDA_11_LIB_DIR)
+CGO_EXTRA_LDFLAGS = $(CUDA_11_CGO_EXTRA_LDFLAGS)
+
+include make/cuda.make
+include make/gpu.make
--- a/make/Makefile.cuda_v12
+++ b/make/Makefile.cuda_v12
@@ -0,0 +1,13 @@
+# Build rules for CUDA v12 runner
+
+include make/common-defs.make
+include make/cuda-v12-defs.make
+
+GPU_RUNNER_VARIANT := _v12
+GPU_COMPILER=$(CUDA_12_COMPILER)
+CUDA_ARCHITECTURES?=60;61;62;70;72;75;80;86;87;89;90;90a
+GPU_LIB_DIR = $(CUDA_12_LIB_DIR)
+CGO_EXTRA_LDFLAGS = $(CUDA_12_CGO_EXTRA_LDFLAGS)
+
+include make/cuda.make
+include make/gpu.make
--- a/make/Makefile.ollama
+++ b/make/Makefile.ollama
@@ -0,0 +1,19 @@
+# Makefile for building top-level ollama binary
+
+include make/common-defs.make
+
+exe: $(OLLAMA_EXE)
+dist_exe dist_ollama: $(DIST_OLLAMA_EXE)
+
+GO_DEPS=$(foreach dir,$(shell go list -deps -f '{{.Dir}}' . ),$(wildcard $(dir)/*.go))
+CPU_GOFLAGS="-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=$(VERSION)\" \"-X=github.com/ollama/ollama/llama.CpuFeatures=$(subst $(space),$(comma),$(TARGET_CPU_FLAGS))\" $(EXTRA_GOLDLAGS) $(TARGET_LDFLAGS)"
+
+$(OLLAMA_EXE) $(DIST_OLLAMA_EXE): TARGET_CPU_FLAGS=$(CUSTOM_CPU_FLAGS)
+$(OLLAMA_EXE) $(DIST_OLLAMA_EXE): $(COMMON_SRCS) $(COMMON_HDRS) $(GO_DEPS)
+	GOARCH=$(ARCH) go build -buildmode=pie $(CPU_GOFLAGS) -trimpath $(if $(CUSTOM_CPU_FLAGS),-tags $(subst $(space),$(comma),$(CUSTOM_CPU_FLAGS))) -o $@ .
+
+.PHONY: ollama dist_ollama exe dist_exe
+
+# Handy debugging for make variables
+print-%:
+	@echo '$*=$($*)'
--- a/make/Makefile.rocm
+++ b/make/Makefile.rocm
@@ -0,0 +1,117 @@
+# Build rules for ROCm runner
+#
+# Note: at present we only support a single ROCm version (whichever is default on the build system)
+# unlike CUDA where we'll build both a v11 and v12 variant.
+
+include make/common-defs.make
+include make/rocm-defs.make
+
+HIP_ARCHS_COMMON := gfx900 gfx940 gfx941 gfx942 gfx1010 gfx1012 gfx1030 gfx1100 gfx1101 gfx1102
+HIP_ARCHS_LINUX := gfx906:xnack- gfx908:xnack- gfx90a:xnack+ gfx90a:xnack-
+
+ifeq ($(OS),windows)
+	GPU_LIB_DIR := $(shell cygpath -m -s "$(HIP_PATH)/bin")
+	CGO_EXTRA_LDFLAGS := -L$(shell cygpath -m -s "$(HIP_PATH)/lib")
+	HIP_ARCHS?=$(HIP_ARCHS_COMMON)
+	GPU_COMPILER_CFLAGS = $(CFLAGS) -D_WIN32_WINNT=0x602
+	GPU_COMPILER_CXXFLAGS = $(CXXFLAGS) -D_WIN32_WINNT=0x602
+else ifeq ($(OS),linux)
+	GPU_LIB_DIR := $(strip $(shell ls -d $(HIP_PATH)/lib64 2>/dev/null || ls -d $(HIP_PATH)/lib 2>/dev/null))
+	CGO_EXTRA_LDFLAGS := -L$(GPU_LIB_DIR)
+	HIP_ARCHS?=$(HIP_ARCHS_COMMON) $(HIP_ARCHS_LINUX)
+	GPU_COMPILER_CFLAGS = $(CFLAGS) -fPIC -D_GNU_SOURCE
+	GPU_COMPILER_CXXFLAGS = $(CXXFLAGS) -fPIC -D_GNU_SOURCE
+endif
+GPU_COMPILER=$(HIP_COMPILER)
+
+# TODO future multi-variant support for ROCm
+# ROCM_VERSION = $(subst $(space),.,$(wordlist 1,2,$(subst .,$(space),$(word 3,$(subst -,$(space),$(filter HIP version: %,$(shell $(GPU_COMPILER) --version)))))))
+# ifneq (,$(ROCM_VERSION))
+# 	GPU_RUNNER_VARIANT = _v$(ROCM_VERSION)
+# endif
+
+GPU_RUNNER_GO_TAGS := rocm
+GPU_RUNNER_NAME := rocm$(GPU_RUNNER_VARIANT)
+GPU_RUNNER_DRIVER_LIB_LINK := -lamdhip64
+GPU_RUNNER_LIBS_SHORT := hipblas rocblas
+
+# Note: ROCm requires an extra step of discovering and copying the transitive dependencies on linux
+ifeq ($(OS),windows)
+	ROCM_DIST_DEPS_DIR = ./dist/$(OS)-$(ARCH)/lib/ollama
+	GPU_LIBS = $(sort $(wildcard $(addsuffix *.$(SHARED_EXT),$(addprefix $(GPU_LIB_DIR)/$(SHARED_PREFIX),$(GPU_RUNNER_LIBS_SHORT)))))
+else ifeq ($(OS),linux)
+	ROCM_DIST_DEPS_DIR = ./dist/$(OS)-$(ARCH)-rocm/lib/ollama
+	GPU_LIBS = $(sort $(wildcard $(addsuffix *.$(SHARED_EXT).*,$(addprefix $(GPU_LIB_DIR)/$(SHARED_PREFIX),$(GPU_RUNNER_LIBS_SHORT)))))
+	ROCM_TRANSITIVE_LIBS_INITIAL = $(sort $(shell ldd $(GPU_LIBS) | grep "=>" | cut -f2 -d= | cut -f2 -d' '  | grep -e rocm -e amdgpu -e libtinfo -e libnuma -e libelf))
+	GPU_TRANSITIVE_LIBS = $(sort $(shell readlink -f $(ROCM_TRANSITIVE_LIBS_INITIAL)) $(ROCM_TRANSITIVE_LIBS_INITIAL))
+	FILTERED_GPU_TRANSITIVE_LIBS=$(sort $(filter-out $(addprefix %,$(notdir $(GPU_LIBS))), $(GPU_TRANSITIVE_LIBS)))
+	GPU_DIST_TRANSITIVE_LIB_DEPS = $(sort $(addprefix $(ROCM_DIST_DEPS_DIR)/,$(notdir $(FILTERED_GPU_TRANSITIVE_LIBS))))
+endif
+GPU_DIST_LIB_DEPS= $(sort $(addprefix $(ROCM_DIST_DEPS_DIR)/,$(notdir $(GPU_LIBS))))
+ROCBLAS_DIST_DEP_MANIFEST = $(ROCM_DIST_DEPS_DIR)/rocblas/library/TensileManifest.txt
+
+ifeq ($(OS),linux)
+	GPU_COMPILER_FPIC := -fPIC -Wno-unused-function -std=gnu++11
+else ifeq ($(OS),windows)
+	GPU_COMPILER_FPIC := -Xclang --dependent-lib=msvcrt
+endif
+GPU_RUNNER_ARCH_FLAGS := $(foreach arch,$(subst ;,$(space),$(HIP_ARCHS)),--offload-arch=$(arch))
+
+# HIPCC uses clang which requires avx512 -> -mavx512f -mavx512dq -mavx512bw
+GPU_VECTOR_FLAGS=$(if $(filter avx512,$(GPU_RUNNER_CPU_FLAGS)),avx512f avx512dq avx512bw) $(filter-out avx512,$(GPU_RUNNER_CPU_FLAGS))
+
+GPU_COMPILER_CUFLAGS = \
+	$(GPU_COMPILER_FPIC) \
+	$(addprefix -m,$(GPU_VECTOR_FLAGS)) \
+	-mf16c \
+	-mfma \
+	-c \
+	-O3 \
+	-DGGML_USE_CUDA \
+	-DGGML_BUILD=1 \
+	-DGGML_SHARED=1 \
+	-DGGML_CUDA_DMMV_X=32 \
+	-DGGML_CUDA_MMV_Y=1 \
+	-DGGML_SCHED_MAX_COPIES=4 \
+	-DGGML_USE_HIPBLAS \
+	-DGGML_USE_LLAMAFILE \
+	-DHIP_FAST_MATH \
+	-D__HIP_PLATFORM_AMD__=1 \
+	-D__HIP_ROCclr__=1 \
+	-DNDEBUG \
+	-DK_QUANTS_PER_ITERATION=2 \
+	-D_CRT_SECURE_NO_WARNINGS \
+	-D_GNU_SOURCE \
+	-D_XOPEN_SOURCE=600 \
+	-DUSE_PROF_API=1 \
+	-std=gnu++14 \
+	-x hip \
+	-mllvm=-amdgpu-early-inline-all=true \
+	-mllvm=-amdgpu-function-calls=false \
+	-Wno-expansion-to-defined \
+	-Wno-invalid-noreturn \
+	-Wno-ignored-attributes \
+	-Wno-pass-failed \
+	-Wno-deprecated-declarations \
+	-Wno-unused-result \
+	-I./llama/
+
+# Workaround buggy P2P copy on some windows multi-GPU setups
+# This workaround breaks linux systems with small system RAM, so only enable on windows
+ifeq ($(OS),windows)
+	GPU_COMPILER_CUFLAGS += -DGGML_CUDA_NO_PEER_COPY=1
+endif
+
+include make/gpu.make
+
+# Adjust the rules from gpu.make to handle the ROCm dependencies properly
+$(RUNNERS_DIST_DIR)/$(GPU_RUNNER_NAME)$(GPU_RUNNER_EXTRA_VARIANT)/ollama_llama_server$(EXE_EXT): $(ROCBLAS_DIST_DEP_MANIFEST) $(GPU_DIST_TRANSITIVE_LIB_DEPS)
+$(ROCBLAS_DIST_DEP_MANIFEST):
+	@-mkdir -p $(dir $@)
+	@echo "Copying rocblas library..."
+	(cd $(GPU_LIB_DIR)/rocblas/library/ && tar cf - . ) | (cd $(dir $@) && tar xf - )
+	@echo "rocblas library copy complete"
+
+$(GPU_DIST_TRANSITIVE_LIB_DEPS):
+	@-mkdir -p $(dir $@)
+	$(CP) $(dir $(filter %$(notdir $@),$(GPU_TRANSITIVE_LIBS)))/$(notdir $@) $(dir $@)
--- a/make/Makefile.sync
+++ b/make/Makefile.sync
@@ -0,0 +1,196 @@
+# Helpers for managing our vendored llama.cpp repo and patch set
+
+REPO_ROOT:=./
+DEST_DIR:=./llama/
+
+include $(DEST_DIR)vendoring
+
+LLAMACPP_REPO := ./llama/vendor/
+
+# Relative to the vendor dir
+VENDOR_RELATIVE_PATCH_DIR := ../patches/
+
+
+help-sync:
+	@echo "The following make targets will help you update llama.cpp to a new base commit, or work on new features/fixes"
+	@echo ""
+	@echo "	make apply-patches	# Establish the tracking repo if not already present, reset to the base commit, and apply our patch set"
+	@echo "	make sync		# Vendor llama.cpp and ggml from the tracking repo working tree"
+	@echo "	make sync-clean		# Remove all vendored files"
+	@echo "	make create-patches	# Generate the patch set based on the current commits in the tracking repo since the base commit"
+	@echo ""
+	@echo "For more details on the workflow, see the Vendoring section in 'docs/development.md'"
+
+apply-patches: $(LLAMACPP_REPO)
+	@if ! git -C $(LLAMACPP_REPO) --no-pager diff --exit-code ; then \
+  		echo "ERROR: Your llama.cpp repo is dirty.  The apply-patches target requires a clean working tree"; \
+		echo "To clobber: git -C $(LLAMACPP_REPO) reset --hard HEAD" ; \
+  		exit 1; \
+	fi
+	@echo "Checking out $(LLAMACPP_BASE_COMMIT)"
+	@git -C $(LLAMACPP_REPO) checkout -q $(LLAMACPP_BASE_COMMIT) || \
+		git -C $(LLAMACPP_REPO) fetch --all && git -C $(LLAMACPP_REPO) checkout -q $(LLAMACPP_BASE_COMMIT)
+	@echo "Applying ollama patches..."
+	@cd $(LLAMACPP_REPO) && git -c 'user.name=nobody' -c 'user.email=<>' am -3 $(VENDOR_RELATIVE_PATCH_DIR)*.patch || \
+		echo "Please resolve the conflicts in $(LLAMACPP_REPO), and run 'git am --continue' to continue applying subsequent patches"
+	@echo ""
+	@echo "The tracking repo $(LLAMACPP_REPO) is now in a detached state with all patches applied."
+	@echo "Don't forget to commit any changes you make and run 'make create-patches' "
+	
+$(LLAMACPP_REPO):
+	@echo "Cloning llama.cpp to $(LLAMACPP_REPO)"
+	git clone https://github.com/ggerganov/llama.cpp.git $@
+
+create-patches: $(LLAMACPP_REPO)
+	@if ! git -C $(LLAMACPP_REPO) --no-pager diff --exit-code ; then \
+  		echo "ERROR: Your llama.cpp repo is dirty.  You must commit any pending changes for format-patch to generate patches"; \
+  		exit 1; \
+	fi
+	@cd $(LLAMACPP_REPO) && git format-patch --no-signature --no-numbered --zero-commit -o $(VENDOR_RELATIVE_PATCH_DIR) $(LLAMACPP_BASE_COMMIT)
+
+# Vendoring template logic
+EXCLUDED_FILES=sgemm.cpp sgemm.h sampling_ext.cpp sampling_ext.h stb_image.h json.hpp llama_darwin.c base64.hpp
+OLLAMA_NATIVE_FILES=mllama.cpp mllama.h llama_darwin.c sampling_ext.cpp sampling_ext.h
+define vendor_file
+$(strip $(addprefix $(2),$(notdir $1))) : $(addprefix $(LLAMACPP_REPO),$(1))
+ifneq ($$(filter-out $(EXCLUDED_FILES),$(notdir $1)),)
+	@echo "vendoring $1"; \
+		mkdir -p $$(dir $$@) && \
+		echo "/**" > $$@ && \
+		echo " * llama.cpp - commit $$(LLAMACPP_BASE_COMMIT) - do not edit this file" >> $$@ && \
+		echo " *" >> $$@ && \
+		sed 's/^/ * /' <$(LLAMACPP_REPO)/LICENSE | sed 's/ *$$$$//' >> $$@ && \
+		echo " */" >> $$@ && \
+		echo "" >> $$@ && \
+		cat $$< >> $$@
+else
+	@echo "vendoring $1"; \
+		mkdir -p $$(dir $$@) && \
+		cat $$< > $$@
+endif
+VENDORED_FILES += $(strip $(addprefix $(2),$(notdir $1)))
+endef
+
+# llama.cpp files -> llama/
+LLAMACPP_FILES=\
+	src/unicode.cpp \
+	src/unicode.h \
+	src/unicode-data.cpp \
+	src/unicode-data.h \
+	src/llama.cpp \
+	src/llama-impl.h \
+	src/llama-vocab.cpp \
+	src/llama-vocab.h \
+	src/llama-grammar.cpp \
+	src/llama-grammar.h \
+	src/llama-sampling.cpp \
+	src/llama-sampling.h \
+	include/llama.h \
+	ggml/src/llamafile/sgemm.cpp \
+	ggml/src/llamafile/sgemm.h
+$(foreach name,$(LLAMACPP_FILES),$(eval $(call vendor_file,$(name),$(DEST_DIR))))
+
+# llama.cpp files -> llama/llamafile
+LLAMAFILE_FILES= \
+	ggml/src/llamafile/sgemm.h
+$(foreach name,$(LLAMAFILE_FILES),$(eval $(call vendor_file,$(name),$(DEST_DIR)llamafile/)))
+
+# ggml files -> llama/
+GGML_FILES= \
+	ggml/src/ggml.c \
+	ggml/include/ggml.h \
+	ggml/src/ggml-quants.c \
+	ggml/src/ggml-quants.h \
+	ggml/src/ggml-metal.metal \
+	ggml/include/ggml-metal.h \
+	ggml/src/ggml-impl.h \
+	ggml/include/ggml-cuda.h \
+	ggml/src/ggml-cuda.cu \
+	ggml/src/ggml-common.h \
+	ggml/include/ggml-backend.h \
+	ggml/src/ggml-backend.c \
+	ggml/src/ggml-backend-impl.h \
+	ggml/include/ggml-alloc.h \
+	ggml/src/ggml-alloc.c \
+	ggml/src/ggml-aarch64.h \
+	ggml/src/ggml-aarch64.c \
+	ggml/src/ggml-cpu-impl.h \
+	ggml/include/ggml-blas.h \
+	ggml/src/ggml-blas.cpp
+$(foreach name,$(GGML_FILES),$(eval $(call vendor_file,$(name),$(DEST_DIR))))
+
+# TODO generalize renaming pattern if we have more of these
+$(DEST_DIR)ggml-metal_darwin_arm64.m : $(LLAMACPP_REPO)ggml/src/ggml-metal.m
+	@echo "vendoring $(subst $(LLAMACPP_REPO),,$<)"; \
+		mkdir -p $(dir $@) && \
+		echo "/**" > $@ && \
+		echo " * llama.cpp - commit $(LLAMACPP_BASE_COMMIT) - do not edit this file" >> $@ && \
+		echo " *" >> $@ && \
+		sed 's/^/ * /' <$(LLAMACPP_REPO)/LICENSE | sed 's/ *$$//' >> $@ && \
+		echo " */" >> $@ && \
+		echo "" >> $@ && \
+		cat $< >> $@
+VENDORED_FILES += $(DEST_DIR)ggml-metal_darwin_arm64.m
+
+# ggml-cuda -> llama/ggml-cuda/
+GGML_CUDA_FILES= ggml/src/ggml-cuda/*.cu ggml/src/ggml-cuda/*.cuh
+GGML_CUDA_FILES_EXPANDED = $(addprefix ggml/src/ggml-cuda/,$(notdir $(wildcard $(addprefix $(LLAMACPP_REPO),$(GGML_CUDA_FILES)))))
+$(foreach name,$(GGML_CUDA_FILES_EXPANDED),$(eval $(call vendor_file,$(name),$(DEST_DIR)ggml-cuda/)))
+
+GGML_TEMPLATE_FILES= ggml/src/ggml-cuda/template-instances/*.cu
+GGML_TEMPLATE_FILES_EXPANDED = 	$(addprefix ggml/src/ggml-cuda/template-instances/,$(notdir $(wildcard $(addprefix $(LLAMACPP_REPO),$(GGML_TEMPLATE_FILES)))))
+$(foreach name,$(GGML_TEMPLATE_FILES_EXPANDED),$(eval $(call vendor_file,$(name),$(DEST_DIR)ggml-cuda/template-instances/)))
+
+GGML_VENDOR_FILES= ggml/src/ggml-cuda/vendors/*.h
+GGML_VENDOR_FILES_EXPANDED=$(addprefix ggml/src/ggml-cuda/vendors/,$(notdir $(wildcard $(addprefix $(LLAMACPP_REPO),$(GGML_VENDOR_FILES)))))
+$(foreach name,$(GGML_VENDOR_FILES_EXPANDED),$(eval $(call vendor_file,$(name),$(DEST_DIR)ggml-cuda/vendors/)))
+
+# llava -> llama/
+LAVA_FILES= \
+	examples/llava/clip.cpp \
+	examples/llava/clip.h \
+	examples/llava/llava.cpp \
+	examples/llava/llava.h \
+	common/log.h \
+	common/log.cpp \
+	common/stb_image.h
+# These files are mostly used by the llava code
+# and shouldn't be necessary once we use clip.cpp directly
+LAVA_FILES+= \
+	common/common.cpp \
+	common/common.h \
+	common/sampling.cpp \
+	common/sampling.h \
+	common/json.hpp \
+	common/json-schema-to-grammar.cpp \
+	common/json-schema-to-grammar.h \
+	common/base64.hpp
+$(foreach name,$(LAVA_FILES),$(eval $(call vendor_file,$(name),$(DEST_DIR))))
+
+$(DEST_DIR)build-info.cpp:
+	@echo "Generating $@"
+	@echo "int LLAMA_BUILD_NUMBER = 0;" > $@
+	@echo "char const *LLAMA_COMMIT = \"$(LLAMACPP_BASE_COMMIT)\";" >> $@
+	@echo "char const *LLAMA_COMPILER = \"\";" >> $@
+	@echo "char const *LLAMA_BUILD_TARGET = \"\";" >> $@
+VENDORED_FILES += $(DEST_DIR)build-info.cpp
+
+
+sync: $(LLAMACPP_REPO) .WAIT $(VENDORED_FILES) .WAIT remove-stale-files
+
+sync-clean:
+	rm -f $(VENDORED_FILES) $(EXTRA_NATIVE_FILES)
+
+PATS=*.c *.h *.cpp *.m *.metal *.cu *.cuh
+NATIVE_DIRS=$(DEST_DIR) $(DEST_DIR)llamafile/ $(DEST_DIR)ggml-cuda/ $(DEST_DIR)ggml-cuda/template-instances/ $(DEST_DIR)ggml-cuda/vendors/
+ALL_NATIVE_FILES=$(foreach dir,$(NATIVE_DIRS),$(wildcard $(addprefix $(dir),$(PATS))))
+EXTRA_NATIVE_FILES=$(filter-out $(VENDORED_FILES) $(addprefix $(DEST_DIR),$(OLLAMA_NATIVE_FILES)), $(ALL_NATIVE_FILES))
+remove-stale-files:
+	@rm -f $(EXTRA_NATIVE_FILES)
+
+.PHONY: help-sync apply-patches sync create-patches remove-stale-fails .WAIT
+
+
+# Handy debugging for make variables
+print-%:
+	@echo '$*=$($*)'
--- a/make/Makefile.test
+++ b/make/Makefile.test
@@ -0,0 +1,19 @@
+# Targets to assist in running tests
+
+include make/common-defs.make
+
+test:
+	cd .. && go test ./... 
+
+integration: $(OLLAMA_EXE)
+	cd .. && go test --tags=integration ./integration -v
+
+lint:
+	cd .. && golangci-lint run -v
+
+# Note: in this makefile we error instead of building to allow more fine-grain control of testing flows
+$(OLLAMA_EXE):
+	@echo ""
+	@echo "ERROR: You must build ollama first - use 'make all' to build the ollama binaries"
+	@echo ""
+	@exit 1
--- a/make/common-defs.make
+++ b/make/common-defs.make
@@ -0,0 +1,91 @@
+# Common definitions for the various Makefiles
+# No rules are defined here so this is safe to include at the beginning of other makefiles
+
+OS := $(shell uname -s)
+ARCH ?= $(subst aarch64,arm64,$(subst x86_64,amd64,$(shell uname -m)))
+ifneq (,$(findstring MINGW,$(OS))$(findstring MSYS,$(OS)))
+	OS := windows
+	ARCH := $(shell systeminfo 2>/dev/null | grep "System Type" | grep ARM64 > /dev/null && echo "arm64" || echo "amd64" )
+else ifeq ($(OS),Linux)
+	OS := linux
+else ifeq ($(OS),Darwin)
+	OS := darwin
+endif
+comma:= ,
+empty:=
+space:= $(empty) $(empty)
+uc = $(subst a,A,$(subst b,B,$(subst c,C,$(subst d,D,$(subst e,E,$(subst f,F,$(subst g,G,$(subst h,H,$(subst i,I,$(subst j,J,$(subst k,K,$(subst l,L,$(subst m,M,$(subst n,N,$(subst o,O,$(subst p,P,$(subst q,Q,$(subst r,R,$(subst s,S,$(subst t,T,$(subst u,U,$(subst v,V,$(subst w,W,$(subst x,X,$(subst y,Y,$(subst z,Z,$1))))))))))))))))))))))))))
+
+export CGO_CFLAGS_ALLOW = -mfma|-mf16c
+export CGO_CXXFLAGS_ALLOW = -mfma|-mf16c
+export HIP_PLATFORM = amd
+export CGO_ENABLED=1
+
+BUILD_DIR = ./llama/build/$(OS)-$(ARCH)
+DIST_BASE = ./dist/$(OS)-$(ARCH)
+
+ifeq ($(OS),windows)
+	# Absolute paths with cygpath to convert to 8.3 without spaces
+	PWD="$(shell pwd)"
+	DIST_OLLAMA_EXE=$(DIST_BASE)/ollama$(EXE_EXT)
+else
+	CCACHE:=$(shell command -v ccache 2>/dev/null || echo "")
+	DIST_OLLAMA_EXE=$(DIST_BASE)/bin/ollama$(EXE_EXT)
+endif
+DIST_LIB_DIR = $(DIST_BASE)/lib/ollama
+RUNNERS_DIST_DIR = $(DIST_LIB_DIR)/runners
+RUNNERS_BUILD_DIR = $(BUILD_DIR)/runners
+VERSION?=$(shell git describe --tags --first-parent --abbrev=7 --long --dirty --always | sed -e "s/^v//g")
+
+# Conditionally enable ccache for cgo builds too
+ifneq ($(CCACHE),)
+	CC?=$(CCACHE) gcc
+	CXX?=$(CCACHE) g++
+	export CC
+	export CXX
+endif
+
+
+# Override in environment to tune CPU vector flags
+ifeq ($(ARCH),amd64)
+ifeq ($(origin CUSTOM_CPU_FLAGS),undefined)
+	GPU_RUNNER_CPU_FLAGS=avx
+	GPU_RUNNER_EXTRA_VARIANT=_avx
+else
+	GPU_RUNNER_CPU_FLAGS=$(subst $(comma),$(space),$(CUSTOM_CPU_FLAGS))
+endif
+endif
+
+ifeq ($(OS),windows)
+	CP := cp
+	OBJ_EXT := obj
+	SHARED_EXT := dll
+	EXE_EXT := .exe
+	SHARED_PREFIX := 
+	CPU_FLAG_PREFIX := /arch:
+ifneq ($(HIP_PATH),)
+	# If HIP_PATH has spaces, hipcc trips over them when subprocessing
+	HIP_PATH := $(shell cygpath -m -s "$(patsubst %\,%,$(HIP_PATH))")
+	export HIP_PATH
+endif
+else ifeq ($(OS),linux)
+	CP := cp -df
+	OBJ_EXT := o
+	SHARED_EXT := so
+	SHARED_PREFIX := lib
+	CPU_FLAG_PREFIX := -m
+else
+	OBJ_EXT := o
+	SHARED_EXT := so
+	CPU_FLAG_PREFIX := -m
+	CP := cp -df
+endif
+
+COMMON_SRCS := \
+	$(wildcard ./llama/*.c) \
+	$(wildcard ./llama/*.cpp)
+COMMON_HDRS := \
+	$(wildcard ./llama/*.h) \
+	$(wildcard ./llama/*.hpp)
+
+OLLAMA_EXE=./ollama$(EXE_EXT)
--- a/make/cuda-v11-defs.make
+++ b/make/cuda-v11-defs.make
@@ -0,0 +1,17 @@
+# Common definitions for the various Makefiles which set cuda settings
+# No rules are defined here so this is safe to include at the beginning of other makefiles
+
+ifeq ($(OS),windows)
+	CUDA_PATH?=$(shell cygpath -m -s "C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\" 2>/dev/null)unknown
+	CUDA_BASE_DIR := $(dir $(shell cygpath -m -s "$(CUDA_PATH)\\.." 2>/dev/null))
+	CUDA_11_PATH:=$(shell ls -d $(CUDA_BASE_DIR)/v11.? 2>/dev/null)
+	CUDA_11_COMPILER:=$(wildcard $(CUDA_11_PATH)/bin/nvcc.exe)
+	CUDA_11_LIB_DIR = $(strip $(shell ls -d $(CUDA_11_PATH)/bin 2>/dev/null))
+	CUDA_11_CGO_EXTRA_LDFLAGS = -L"$(CUDA_11_PATH)/lib/x64"
+else ifeq ($(OS),linux)
+	CUDA_PATH?=/usr/local/cuda
+	CUDA_11_PATH:=$(shell ls -d $(CUDA_PATH)-11 2>/dev/null)
+	CUDA_11_COMPILER:=$(wildcard $(CUDA_11_PATH)/bin/nvcc)
+	CUDA_11_LIB_DIR=$(strip $(shell ls -d $(CUDA_11_PATH)/lib64 2>/dev/null || ls -d $(CUDA_11_PATH)/lib 2>/dev/null))
+	CUDA_11_CGO_EXTRA_LDFLAGS = -L"$(CUDA_11_LIB_DIR)" -L"$(CUDA_11_LIB_DIR)/stubs"
+endif
--- a/make/cuda-v12-defs.make
+++ b/make/cuda-v12-defs.make
@@ -0,0 +1,17 @@
+# Common definitions for the various Makefiles which set cuda settings
+# No rules are defined here so this is safe to include at the beginning of other makefiles
+
+ifeq ($(OS),windows)
+	CUDA_PATH?=$(shell cygpath -m -s "C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\" 2>/dev/null)unknown
+	CUDA_BASE_DIR := $(dir $(shell cygpath -m -s "$(CUDA_PATH)\\.." 2>/dev/null))
+	CUDA_12_PATH:=$(shell ls -d $(CUDA_BASE_DIR)/v12.? 2>/dev/null)
+	CUDA_12_COMPILER:=$(wildcard $(CUDA_12_PATH)/bin/nvcc.exe)
+	CUDA_12_LIB_DIR = $(strip $(shell ls -d $(CUDA_12_PATH)/bin 2>/dev/null))
+	CUDA_12_CGO_EXTRA_LDFLAGS = -L"$(CUDA_12_PATH)/lib/x64"
+else ifeq ($(OS),linux)
+	CUDA_PATH?=/usr/local/cuda
+	CUDA_12_PATH:=$(shell ls -d $(CUDA_PATH)-12 2>/dev/null)
+	CUDA_12_COMPILER:=$(wildcard $(CUDA_12_PATH)/bin/nvcc)
+	CUDA_12_LIB_DIR=$(strip $(shell ls -d $(CUDA_12_PATH)/lib64 2>/dev/null || ls -d $(CUDA_12_PATH)/lib 2>/dev/null))
+	CUDA_12_CGO_EXTRA_LDFLAGS = -L"$(CUDA_12_LIB_DIR)" -L"$(CUDA_12_LIB_DIR)/stubs" 
+endif
--- a/make/cuda.make
+++ b/make/cuda.make
@@ -0,0 +1,54 @@
+# Common definitions for all cuda versions
+
+ifndef GPU_RUNNER_VARIANT
+dummy:
+	$(error This makefile is not meant to build directly, but instead included in other Makefiles that set required variables)
+endif
+
+
+GPU_RUNNER_NAME := cuda$(GPU_RUNNER_VARIANT)
+GPU_RUNNER_GO_TAGS := cuda cuda$(GPU_RUNNER_VARIANT)
+GPU_RUNNER_DRIVER_LIB_LINK := -lcuda
+GPU_RUNNER_LIBS_SHORT := cublas cudart cublasLt
+
+ifeq ($(OS),windows)
+	# On windows, nvcc uses msvc which does not support avx512vbmi avx512vnni avx512bf16, but macros can turn them on
+	GPU_VECTOR_FLAGS=$(call uc,$(filter-out avx512bf16,$(filter-out avx512vnni,$(filter-out avx512vbmi,$(GPU_RUNNER_CPU_FLAGS)))))
+	GPU_COMPILER_EXTRA_FLAGS=$(if $(filter avx512vbmi,$(GPU_RUNNER_CPU_FLAGS)),-D__AVX512VBMI__)
+	GPU_COMPILER_EXTRA_FLAGS+=$(if $(filter avx512vnni,$(GPU_RUNNER_CPU_FLAGS)),-D__AVX512VNNI__)
+	GPU_COMPILER_EXTRA_FLAGS+=$(if $(filter avx512bf16,$(GPU_RUNNER_CPU_FLAGS)),-D__AVX512BF16__)
+	GPU_LIBS = $(sort $(wildcard $(addsuffix *.$(SHARED_EXT),$(addprefix $(GPU_LIB_DIR)/$(SHARED_PREFIX),$(GPU_RUNNER_LIBS_SHORT)))))
+	GPU_COMPILER_CFLAGS = $(CFLAGS) -D_WIN32_WINNT=0x602
+	GPU_COMPILER_CXXFLAGS = $(CXXFLAGS) -D_WIN32_WINNT=0x602
+else ifeq ($(OS),linux)
+	# On linux, nvcc requires avx512 -> -mavx512f -mavx512dq -mavx512bw
+	GPU_VECTOR_FLAGS=$(if $(filter avx512,$(GPU_RUNNER_CPU_FLAGS)),avx512f avx512dq avx512bw) $(filter-out avx512,$(GPU_RUNNER_CPU_FLAGS))
+	GPU_COMPILER_EXTRA_FLAGS = -fPIC -Wno-unused-function -std=c++11
+	GPU_LIBS = $(sort $(wildcard $(addsuffix *.$(SHARED_EXT).*,$(addprefix $(GPU_LIB_DIR)/$(SHARED_PREFIX),$(GPU_RUNNER_LIBS_SHORT)))))
+	GPU_COMPILER_CFLAGS = $(CFLAGS) -Xcompiler -fPIC -D_GNU_SOURCE
+	GPU_COMPILER_CXXFLAGS = $(CXXFLAGS) -Xcompiler -fPIC -D_GNU_SOURCE
+endif
+GPU_DIST_LIB_DEPS= $(sort $(addprefix $(DIST_GPU_RUNNER_DEPS_DIR)/,$(notdir $(GPU_LIBS))))
+
+GPU_RUNNER_ARCH_FLAGS := $(foreach arch,$(subst ;,$(space),$(CUDA_ARCHITECTURES)),--generate-code=arch=compute_$(arch)$(comma)code=[compute_$(arch)$(comma)sm_$(arch)]) \
+	-DGGML_CUDA_USE_GRAPHS=1
+GPU_COMPILER_CUFLAGS = \
+	$(GPU_COMPILER_EXTRA_FLAGS) \
+	-Xcompiler "$(addprefix $(CPU_FLAG_PREFIX),$(GPU_VECTOR_FLAGS))" \
+	-t2 \
+	-DGGML_CUDA_DMMV_X=32 \
+	-DGGML_CUDA_MMV_Y=1 \
+	-DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 \
+	-DGGML_USE_CUDA=1 \
+	-DGGML_SHARED=1 \
+	-DGGML_BUILD=1 \
+	-DGGML_USE_LLAMAFILE \
+	-DK_QUANTS_PER_ITERATION=2 \
+	-DNDEBUG \
+	-D_GNU_SOURCE \
+	-D_XOPEN_SOURCE=600 \
+	-Wno-deprecated-gpu-targets \
+	--forward-unknown-to-host-compiler \
+	-use_fast_math \
+	-I./llama/  \
+	-O3
--- a/make/gpu.make
+++ b/make/gpu.make
@@ -0,0 +1,90 @@
+# Generalized GPU runner build
+
+ifndef GPU_RUNNER_NAME
+dummy:
+	$(error This makefile is not meant to build directly, but instead included in other Makefiles that set required variables)
+endif
+
+GPU_GOFLAGS="-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=$(VERSION)\" \"-X=github.com/ollama/ollama/llama.CpuFeatures=$(subst $(space),$(comma),$(GPU_RUNNER_CPU_FLAGS))\" $(EXTRA_GOLDLAGS) $(TARGET_LDFLAGS)"
+
+# TODO Unify how we handle dependencies in the dist/packaging and install flow
+# today, cuda is bundled, but rocm is split out.  Should split them each out by runner
+DIST_GPU_RUNNER_DEPS_DIR = $(DIST_LIB_DIR)
+
+
+GPU_RUNNER_LIBS = $(wildcard $(addsuffix .$(SHARED_EXT).*,$(addprefix $(GPU_LIB_DIR)/$(SHARED_PREFIX),$(GPU_RUNNER_LIBS_SHORT))))
+
+GPU_RUNNER_SRCS := \
+	llama/ggml-cuda.cu \
+	$(filter-out $(wildcard llama/ggml-cuda/fattn*.cu),$(wildcard llama/ggml-cuda/*.cu)) \
+	$(wildcard llama/ggml-cuda/template-instances/mmq*.cu) \
+	llama/ggml.c llama/ggml-backend.c llama/ggml-alloc.c llama/ggml-quants.c llama/sgemm.cpp llama/ggml-aarch64.c
+GPU_RUNNER_HDRS := \
+	$(wildcard llama/ggml-cuda/*.cuh)
+
+
+# Conditional flags and components to speed up developer builds
+ifneq ($(OLLAMA_FAST_BUILD),)
+	GPU_COMPILER_CUFLAGS += 	\
+		-DGGML_DISABLE_FLASH_ATTN
+else
+	GPU_RUNNER_SRCS += \
+		$(wildcard llama/ggml-cuda/fattn*.cu) \
+		$(wildcard llama/ggml-cuda/template-instances/fattn-wmma*.cu) \
+		$(wildcard llama/ggml-cuda/template-instances/fattn-vec*q4_0-q4_0.cu) \
+		$(wildcard llama/ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu) \
+		$(wildcard llama/ggml-cuda/template-instances/fattn-vec*f16-f16.cu)
+endif
+
+GPU_RUNNER_OBJS := $(GPU_RUNNER_SRCS:.cu=.$(GPU_RUNNER_NAME).$(OBJ_EXT))
+GPU_RUNNER_OBJS := $(GPU_RUNNER_OBJS:.c=.$(GPU_RUNNER_NAME).$(OBJ_EXT))
+GPU_RUNNER_OBJS := $(addprefix $(BUILD_DIR)/,$(GPU_RUNNER_OBJS:.cpp=.$(GPU_RUNNER_NAME).$(OBJ_EXT)))
+
+DIST_RUNNERS = $(addprefix $(RUNNERS_DIST_DIR)/,$(addsuffix /ollama_llama_server$(EXE_EXT),$(GPU_RUNNER_NAME)$(GPU_RUNNER_EXTRA_VARIANT)))
+BUILD_RUNNERS = $(addprefix $(RUNNERS_BUILD_DIR)/,$(addsuffix /ollama_llama_server$(EXE_EXT),$(GPU_RUNNER_NAME)$(GPU_RUNNER_EXTRA_VARIANT)))
+
+
+$(GPU_RUNNER_NAME): $(BUILD_RUNNERS) 
+
+dist: $(DIST_RUNNERS)
+
+# Build targets
+$(BUILD_DIR)/%.$(GPU_RUNNER_NAME).$(OBJ_EXT): %.cu
+	@-mkdir -p $(dir $@)
+	$(CCACHE) $(GPU_COMPILER) -c $(GPU_COMPILER_CFLAGS) $(GPU_COMPILER_CUFLAGS) $(GPU_RUNNER_ARCH_FLAGS) -o $@ $<
+$(BUILD_DIR)/%.$(GPU_RUNNER_NAME).$(OBJ_EXT): %.c
+	@-mkdir -p $(dir $@)
+	$(CCACHE) $(GPU_COMPILER) -c $(GPU_COMPILER_CFLAGS) -o $@ $<
+$(BUILD_DIR)/%.$(GPU_RUNNER_NAME).$(OBJ_EXT): %.cpp
+	@-mkdir -p $(dir $@)
+	$(CCACHE) $(GPU_COMPILER) -c $(GPU_COMPILER_CXXFLAGS) -o $@ $<
+$(RUNNERS_BUILD_DIR)/$(GPU_RUNNER_NAME)$(GPU_RUNNER_EXTRA_VARIANT)/ollama_llama_server$(EXE_EXT): TARGET_CGO_LDFLAGS = $(CGO_EXTRA_LDFLAGS) -L"$(RUNNERS_BUILD_DIR)/$(GPU_RUNNER_NAME)$(GPU_RUNNER_EXTRA_VARIANT)/"
+$(RUNNERS_BUILD_DIR)/$(GPU_RUNNER_NAME)$(GPU_RUNNER_EXTRA_VARIANT)/ollama_llama_server$(EXE_EXT): $(RUNNERS_BUILD_DIR)/$(GPU_RUNNER_NAME)$(GPU_RUNNER_EXTRA_VARIANT)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).$(SHARED_EXT) ./llama/*.go ./llama/runner/*.go $(COMMON_SRCS) $(COMMON_HDRS)
+	@-mkdir -p $(dir $@)
+	GOARCH=$(ARCH) CGO_LDFLAGS="$(TARGET_CGO_LDFLAGS)" go build -buildmode=pie $(GPU_GOFLAGS) -trimpath -tags $(subst $(space),$(comma),$(GPU_RUNNER_CPU_FLAGS) $(GPU_RUNNER_GO_TAGS)) -o $@ ./cmd/runner
+$(RUNNERS_BUILD_DIR)/$(GPU_RUNNER_NAME)$(GPU_RUNNER_EXTRA_VARIANT)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).$(SHARED_EXT): $(GPU_RUNNER_OBJS) $(COMMON_HDRS) $(GPU_RUNNER_HDRS)
+	@-mkdir -p $(dir $@)
+	$(CCACHE) $(GPU_COMPILER) --shared -L$(GPU_LIB_DIR) $(GPU_RUNNER_DRIVER_LIB_LINK) -L${DIST_GPU_RUNNER_DEPS_DIR} $(foreach lib, $(GPU_RUNNER_LIBS_SHORT), -l$(lib)) $(GPU_RUNNER_OBJS) -o $@
+
+# Distribution targets
+$(RUNNERS_DIST_DIR)/%: $(RUNNERS_BUILD_DIR)/%
+	@-mkdir -p $(dir $@)
+	$(CP) $< $@
+$(RUNNERS_DIST_DIR)/$(GPU_RUNNER_NAME)$(GPU_RUNNER_EXTRA_VARIANT)/ollama_llama_server$(EXE_EXT): $(RUNNERS_DIST_DIR)/$(GPU_RUNNER_NAME)$(GPU_RUNNER_EXTRA_VARIANT)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).$(SHARED_EXT) $(GPU_DIST_LIB_DEPS)
+$(RUNNERS_DIST_DIR)/$(GPU_RUNNER_NAME)$(GPU_RUNNER_EXTRA_VARIANT)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).$(SHARED_EXT): $(RUNNERS_BUILD_DIR)/$(GPU_RUNNER_NAME)$(GPU_RUNNER_EXTRA_VARIANT)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).$(SHARED_EXT)
+	@-mkdir -p $(dir $@)
+	$(CP) $< $@
+$(GPU_DIST_LIB_DEPS):
+	@-mkdir -p $(dir $@)
+	$(CP) $(GPU_LIB_DIR)/$(notdir $@) $(dir $@)
+
+clean: 
+	rm -f $(GPU_RUNNER_OBJS) $(BUILD_RUNNERS) $(DIST_RUNNERS)
+
+.PHONY: clean $(GPU_RUNNER_NAME)
+
+
+# Handy debugging for make variables
+print-%:
+	@echo '$*=$($*)'
+
--- a/make/rocm-defs.make
+++ b/make/rocm-defs.make
@@ -0,0 +1,9 @@
+# Common definitions for the various Makefiles which set cuda settings
+# No rules are defined here so this is safe to include at the beginning of other makefiles
+
+ifeq ($(OS),windows)
+	HIP_COMPILER:=$(wildcard $(HIP_PATH)/bin/hipcc.bin.exe)
+else ifeq ($(OS),linux)
+	HIP_PATH?=$(shell ls -d /opt/rocm 2>/dev/null)
+	HIP_COMPILER:=$(wildcard $(HIP_PATH)/bin/hipcc)
+endif