Improve dependency gathering logic (#7345)

This unfies the rocm/cuda dependency logic into the makefile
and fixes a missing define which broke windows rocm

llama/Makefile

@@ -9,8 +9,7 @@ ifeq ($(OS),windows)
 	CUDA_BASE_DIR := $(dir $(shell cygpath -m -s "$(CUDA_PATH)\\.." 2>/dev/null))
 	CUDA_11:=$(shell ls -d $(CUDA_BASE_DIR)/v11.? 2>/dev/null)
 	CUDA_12:=$(shell ls -d $(CUDA_BASE_DIR)/v12.? 2>/dev/null)
-	HIP_PATH_83 := $(shell cygpath -m -s "$(subst \,/,$(HIP_PATH))" 2>/dev/null)
-	HIP_LIB_DIR := $(shell ls -d $(HIP_PATH_83)/lib 2>/dev/null)
+	HIP_LIB_DIR := $(shell ls -d $(HIP_PATH)/lib 2>/dev/null)
 else ifeq ($(OS),linux)
 	HIP_PATH?=/opt/rocm
 	HIP_LIB_DIR := $(shell ls -d $(HIP_PATH)/lib 2>/dev/null)

llama/llama.go

@@ -44,8 +44,8 @@ package llama
 #cgo rocm CFLAGS: -DGGML_USE_CUDA -DGGML_USE_HIPBLAS -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_CUDA_MMV_Y=1 -DGGML_BUILD=1
 #cgo rocm CXXFLAGS: -DGGML_USE_CUDA -DGGML_USE_HIPBLAS -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_CUDA_MMV_Y=1 -DGGML_BUILD=1
 #cgo rocm LDFLAGS: -L${SRCDIR} -lggml_rocm -lhipblas -lamdhip64 -lrocblas
-#cgo windows CFLAGS: -Wno-discarded-qualifiers
-#cgo windows CFLAGS: -Wno-discarded-qualifiers
+#cgo windows CFLAGS: -Wno-discarded-qualifiers -D_WIN32_WINNT=0x602
+#cgo windows CXXFLAGS: -D_WIN32_WINNT=0x602
 #cgo windows LDFLAGS: -lmsvcrt
 #cgo windows LDFLAGS: -lmsvcrt -static-libstdc++ -static-libgcc -static
 #cgo windows,amd64 LDFLAGS: -L${SRCDIR}/build/Windows/amd64

llama/make/Makefile.default

@@ -24,17 +24,17 @@ all: $(BUILD_RUNNERS) $(DIST_RUNNERS) $(PAYLOAD_RUNNERS)
 $(RUNNERS_BUILD_DIR)/$(DEFAULT_RUNNER)/ollama_llama_server$(EXE_EXT): TARGET_CPU_FLAGS=$(CUSTOM_CPU_FLAGS)
 $(RUNNERS_BUILD_DIR)/$(DEFAULT_RUNNER)/ollama_llama_server$(EXE_EXT): *.go ./runner/*.go $(COMMON_SRCS) $(COMMON_HDRS)
 	@-mkdir -p $(dir $@)
-	GOARCH=$(ARCH) go build -buildmode=pie $(CPU_GOFLAGS) $(if $(CUSTOM_CPU_FLAGS),-tags $(subst $(space),$(comma),$(CUSTOM_CPU_FLAGS))) -o $@ ./runner
+	GOARCH=$(ARCH) go build -buildmode=pie $(CPU_GOFLAGS) -trimpath $(if $(CUSTOM_CPU_FLAGS),-tags $(subst $(space),$(comma),$(CUSTOM_CPU_FLAGS)))  -o $@ ./runner
 
 $(RUNNERS_BUILD_DIR)/cpu_avx/ollama_llama_server$(EXE_EXT): TARGET_CPU_FLAGS="avx"
 $(RUNNERS_BUILD_DIR)/cpu_avx/ollama_llama_server$(EXE_EXT): *.go ./runner/*.go $(COMMON_SRCS) $(COMMON_HDRS)
 	@-mkdir -p $(dir $@)
-	GOARCH=$(ARCH) go build -buildmode=pie $(CPU_GOFLAGS) -tags $(subst $(space),$(comma),$(TARGET_CPU_FLAGS)) -o $@ ./runner
+	GOARCH=$(ARCH) go build -buildmode=pie $(CPU_GOFLAGS) -trimpath -tags $(subst $(space),$(comma),$(TARGET_CPU_FLAGS)) -o $@ ./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): *.go ./runner/*.go $(COMMON_SRCS) $(COMMON_HDRS)
 	@-mkdir -p $(dir $@)
-	GOARCH=$(ARCH) go build -buildmode=pie $(CPU_GOFLAGS) -tags $(subst $(space),$(comma),$(TARGET_CPU_FLAGS)) -o $@ ./runner
+	GOARCH=$(ARCH) go build -buildmode=pie $(CPU_GOFLAGS) -trimpath -tags $(subst $(space),$(comma),$(TARGET_CPU_FLAGS)) -o $@ ./runner
 
 $(RUNNERS_DIST_DIR)/%: $(RUNNERS_BUILD_DIR)/%
 	@-mkdir -p $(dir $@)
@@ -49,3 +49,6 @@ clean:
 
 .PHONY: clean all
 
+# Handy debugging for make variables
+print-%:
+	@echo '$*=$($*)'

llama/make/Makefile.rocm

@@ -9,20 +9,16 @@ HIP_ARCHS_COMMON := gfx900 gfx940 gfx941 gfx942 gfx1010 gfx1012 gfx1030 gfx1100
 HIP_ARCHS_LINUX := gfx906:xnack- gfx908:xnack- gfx90a:xnack+ gfx90a:xnack-
 
 ifeq ($(OS),windows)
-	GPU_LIB_DIR_WIN := $(shell cygpath -m -s "$(HIP_PATH)\bin")
-	# If HIP_PATH has spaces, hipcc trips over them when subprocessing
-	HIP_PATH := $(shell cygpath -m -s "$(HIP_PATH)\")
-	CGO_EXTRA_LDFLAGS_WIN := -L$(shell cygpath -m -s "$(HIP_PATH)\lib")
-	export HIP_PATH
-	GPU_COMPILER_WIN := $(HIP_PATH)bin/hipcc.bin.exe
+	GPU_LIB_DIR_WIN := $(shell cygpath -m -s "$(HIP_PATH)/bin")
+	CGO_EXTRA_LDFLAGS_WIN := -L$(shell cygpath -m -s "$(HIP_PATH)/lib")
+	GPU_COMPILER_WIN := $(HIP_PATH)/bin/hipcc.bin.exe
 	GPU_COMPILER:=$(GPU_COMPILER_WIN)
 else ifeq ($(OS),linux)
-	HIP_PATH?=/opt/rocm
 	GPU_LIB_DIR_LINUX := $(HIP_PATH)/lib
 	GPU_COMPILER_LINUX := $(shell X=$$(which hipcc 2>/dev/null) && echo $$X)
 	GPU_COMPILER:=$(GPU_COMPILER_LINUX)
-	ROCM_TRANSITIVE_LIBS_INITIAL = $(sort $(shell ldd $(ROCM_LIBS) | grep "=>" | cut -f2 -d= | cut -f2 -d' '  | grep -e rocm -e amdgpu -e libtinfo -e libnuma -e libelf))
-	ROCM_TRANSITIVE_LIBS = $(sort $(shell readlink -f $(ROCM_TRANSITIVE_LIBS_INITIAL)) $(ROCM_TRANSITIVE_LIBS_INITIAL))
+	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))
 endif
 
 # TODO future multi-variant support for ROCm
@@ -42,13 +38,13 @@ GPU_COMPILER_CFLAGS_LINUX = $(CFLAGS) -fPIC -D_GNU_SOURCE
 GPU_COMPILER_CXXFLAGS_WIN = $(CXXFLAGS) -D_WIN32_WINNT=0x602
 GPU_COMPILER_CXXFLAGS_LINUX = $(CXXFLAGS) -fPIC -D_GNU_SOURCE
 
-ROCM_LIBS = $(wildcard $(addsuffix .$(SHARED_EXT),$(addprefix $(GPU_LIB_DIR)/$(SHARED_PREFIX),$(GPU_RUNNER_LIBS_SHORT))))
+GPU_LIBS = $(wildcard $(addsuffix .$(SHARED_EXT),$(addprefix $(GPU_LIB_DIR)/$(SHARED_PREFIX),$(GPU_RUNNER_LIBS_SHORT))))
 ifeq ($(OS),windows)
 	ROCM_DIST_DEPS_DIR = $(abspath $(SRC_DIR)/../dist/$(OS)-$(ARCH))/lib/ollama
 else ifeq ($(OS),linux)
 	ROCM_DIST_DEPS_DIR = $(abspath $(SRC_DIR)/../dist/$(OS)-$(ARCH)-rocm)/lib/ollama
 endif
-ROCM_DIST_DEPS_LIBS= $(sort $(addprefix $(ROCM_DIST_DEPS_DIR)/,$(notdir $(ROCM_LIBS)) $(notdir $(ROCM_TRANSITIVE_LIBS))))
+GPU_DIST_DEPS_LIBS= $(sort $(addprefix $(ROCM_DIST_DEPS_DIR)/,$(notdir $(GPU_LIBS)) $(notdir $(GPU_TRANSITIVE_LIBS))))
 ROCBLAS_DIST_DEP_MANIFEST = $(ROCM_DIST_DEPS_DIR)/rocblas/library/TensileManifest.txt
 
 ifeq ($(OS),linux)
@@ -94,12 +90,9 @@ GPU_COMPILER_CUFLAGS = \
 include make/gpu.make
 
 # Adjust the rules from gpu.make to handle the ROCm dependencies properly
-$(RUNNERS_DIST_DIR)/$(GPU_RUNNER_NAME)/ollama_llama_server$(EXE_EXT): $(ROCBLAS_DIST_DEP_MANIFEST) $(ROCM_DIST_DEPS_LIBS)
+$(RUNNERS_DIST_DIR)/$(GPU_RUNNER_NAME)/ollama_llama_server$(EXE_EXT): $(ROCBLAS_DIST_DEP_MANIFEST)
 $(ROCBLAS_DIST_DEP_MANIFEST):
 	@-mkdir -p $(dir $@)
 	@echo "Copying rocblas library..."
 	cd $(GPU_LIB_DIR)/rocblas/library/ && tar cf - . | (cd $(dir $@) && tar xf - )
 	@echo "rocblas library copy complete"
-$(ROCM_DIST_DEPS_LIBS): 
-	@-mkdir -p $(dir $@)
-	$(CP) $(dir $(filter %$(notdir $@),$(ROCM_LIBS) $(ROCM_TRANSITIVE_LIBS)))/$(notdir $@) $(dir $@)

llama/make/common-defs.make

@@ -57,12 +57,18 @@ ifeq ($(OS),windows)
 	EXE_EXT := .exe
 	SHARED_PREFIX := 
 	CPU_FLAG_PREFIX := /arch:
+ifneq ($(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 -af
 	OBJ_EXT := o
 	SHARED_EXT := so
 	SHARED_PREFIX := lib
 	CPU_FLAG_PREFIX := -m
+	HIP_PATH?=/opt/rocm
 else
 	OBJ_EXT := o
 	SHARED_EXT := so

llama/make/cuda.make

@@ -19,6 +19,9 @@ GPU_COMPILER_CFLAGS_WIN = $(CFLAGS) -D_WIN32_WINNT=0x602
 GPU_COMPILER_CFLAGS_LINUX = $(CFLAGS) -Xcompiler -fPIC -D_GNU_SOURCE
 GPU_COMPILER_CXXFLAGS_WIN = $(CXXFLAGS) -D_WIN32_WINNT=0x602
 GPU_COMPILER_CXXFLAGS_LINUX = $(CXXFLAGS) -Xcompiler -fPIC -D_GNU_SOURCE
+GPU_LIBS = $(sort $(wildcard $(addsuffix *.$(SHARED_EXT)*,$(addprefix $(GPU_LIB_DIR)/$(SHARED_PREFIX),$(GPU_RUNNER_LIBS_SHORT)))))
+GPU_DIST_DEPS_LIBS= $(sort $(addprefix $(DIST_LIB_DIR)/,$(notdir $(GPU_LIBS))))
+
 ifeq ($(OS),linux)
 	CUDA_PATH?=/usr/local/cuda
 	GPU_COMPILER_FPIC = -fPIC -Wno-unused-function -std=c++11

llama/make/gpu.make

@@ -89,7 +89,7 @@ $(BUILD_DIR)/%.$(GPU_RUNNER_NAME).$(OBJ_EXT): %.cpp
 $(RUNNERS_BUILD_DIR)/$(GPU_RUNNER_NAME)/ollama_llama_server$(EXE_EXT): TARGET_CGO_LDFLAGS = -L"$(RUNNERS_BUILD_DIR)/$(GPU_RUNNER_NAME)/" $(CGO_EXTRA_LDFLAGS)
 $(RUNNERS_BUILD_DIR)/$(GPU_RUNNER_NAME)/ollama_llama_server$(EXE_EXT): $(RUNNERS_BUILD_DIR)/$(GPU_RUNNER_NAME)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).$(SHARED_EXT) *.go ./runner/*.go $(COMMON_SRCS) $(COMMON_HDRS)
 	@-mkdir -p $(dir $@)
-	GOARCH=$(ARCH) CGO_LDFLAGS="$(TARGET_CGO_LDFLAGS)" go build -buildmode=pie  $(GPU_GOFLAGS) -tags $(subst $(space),$(comma),$(GPU_RUNNER_CPU_FLAGS) $(GPU_RUNNER_GO_TAGS)) -o $@ ./runner
+	GOARCH=$(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 $@ ./runner
 $(RUNNERS_BUILD_DIR)/$(GPU_RUNNER_NAME)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).$(SHARED_EXT): $(GPU_RUNNER_OBJS) $(DIST_GPU_RUNNER_LIB_DEPS) $(COMMON_HDRS) $(GPU_RUNNER_HDRS)
 	@-mkdir -p $(dir $@)
 	$(CCACHE) $(GPU_COMPILER) --shared $(GPU_RUNNER_DRIVER_LIB_LINK) -L${DIST_GPU_RUNNER_DEPS_DIR} $(foreach lib, $(GPU_RUNNER_LIBS_SHORT), -l$(lib)) $(GPU_RUNNER_OBJS) -o $@
@@ -98,13 +98,16 @@ $(RUNNERS_BUILD_DIR)/$(GPU_RUNNER_NAME)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).
 $(RUNNERS_DIST_DIR)/%: $(RUNNERS_BUILD_DIR)/%
 	@-mkdir -p $(dir $@)
 	$(CP) $< $@
-$(RUNNERS_DIST_DIR)/$(GPU_RUNNER_NAME)/ollama_llama_server$(EXE_EXT): $(DIST_LIB_DIR)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).$(SHARED_EXT)
+$(RUNNERS_DIST_DIR)/$(GPU_RUNNER_NAME)/ollama_llama_server$(EXE_EXT): $(DIST_LIB_DIR)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).$(SHARED_EXT) $(GPU_DIST_DEPS_LIBS)
 $(DIST_LIB_DIR)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).$(SHARED_EXT): $(RUNNERS_BUILD_DIR)/$(GPU_RUNNER_NAME)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).$(SHARED_EXT)
 	@-mkdir -p $(dir $@)
 	$(CP) $< $@
 $(DIST_GPU_RUNNER_LIB_DEPS): 
 	@-mkdir -p $(dir $@)
 	$(CP) $(GPU_LIB_DIR)/$(notdir $@) $(dir $@)
+$(GPU_DIST_DEPS_LIBS): 
+	@-mkdir -p $(dir $@)
+	$(CP) $(dir $(filter %$(notdir $@),$(GPU_LIBS) $(GPU_TRANSITIVE_LIBS)))/$(notdir $@) $(dir $@)
 
 # Payload targets
 $(RUNNERS_PAYLOAD_DIR)/%/ollama_llama_server.gz: $(RUNNERS_BUILD_DIR)/%/ollama_llama_server