tcsenpai/ollama
1
0
Fork 0
mirror of https://github.com/tcsenpai/ollama.git synced 2025-06-16 07:47:33 +00:00
Code Issues Packages Projects Releases Wiki Activity

Prefix all build artifacts with an OS/ARCH dir

Browse Source 
This will help keep incremental builds from stomping on each other and make it
easier to stitch together the final runner payloads
This commit is contained in:
Daniel Hiltgen 2024-06-24 09:23:34 -07:00
parent 6456a7fd73
commit 5152a430f5
3 changed files with 47 additions and 35 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
llama/.gitignore vendored
View File

@ -1,9 +1,3 @@
*.bin *.bin
*.gguf *.gguf
*.lib build/
*.exp
*.dll
*.so
*.o
*.obj
ollama_runner*

63
llama/Makefile
View File

@ -1,15 +1,20 @@
OS := $(shell uname -s) OS := $(shell uname -s)
ARCH := $(or $(ARCH), $(shell uname -m)) ARCH := $(or $(ARCH), $(shell uname -m))
ifeq ($(ARCH),x86_64)
ARCH := amd64
endif
ifneq (,$(findstring MINGW,$(OS))$(findstring MSYS,$(OS)))
OS := Windows
endif
export CGO_CFLAGS_ALLOW = -mfma|-mf16c export CGO_CFLAGS_ALLOW = -mfma|-mf16c
export CGO_CXXFLAGS_ALLOW = -mfma|-mf16c export CGO_CXXFLAGS_ALLOW = -mfma|-mf16c
export HIP_PLATFORM = amd export HIP_PLATFORM = amd
ifeq ($(ARCH),x86_64) SRC_DIR := $(dir $(abspath $(lastword $(MAKEFILE_LIST))))
ARCH := amd64 BUILD_DIR := $(SRC_DIR)/build/$(OS)/$(ARCH)/
endif
ifneq (,$(findstring MINGW,$(OS))$(findstring MSYS,$(OS))) ifeq ($(OS),Windows)
OBJ_EXT := obj OBJ_EXT := obj
SHARED_EXT := dll SHARED_EXT := dll
EXE_EXT := .exe EXE_EXT := .exe
@ -48,11 +53,11 @@ CUDA_SRCS := \
CUDA_OBJS := $(CUDA_SRCS:.cu=.cuda.$(OBJ_EXT)) CUDA_OBJS := $(CUDA_SRCS:.cu=.cuda.$(OBJ_EXT))
CUDA_OBJS := $(CUDA_OBJS:.c=.cuda.$(OBJ_EXT)) CUDA_OBJS := $(CUDA_OBJS:.c=.cuda.$(OBJ_EXT))
CUDA_OBJS := $(CUDA_OBJS:.cpp=.cuda.$(OBJ_EXT)) CUDA_OBJS := $(addprefix $(BUILD_DIR),$(CUDA_OBJS:.cpp=.cuda.$(OBJ_EXT)))
HIP_OBJS := $(CUDA_SRCS:.cu=.hip.$(OBJ_EXT)) HIP_OBJS := $(CUDA_SRCS:.cu=.hip.$(OBJ_EXT))
HIP_OBJS := $(HIP_OBJS:.c=.hip.$(OBJ_EXT)) HIP_OBJS := $(HIP_OBJS:.c=.hip.$(OBJ_EXT))
HIP_OBJS := $(HIP_OBJS:.cpp=.hip.$(OBJ_EXT)) HIP_OBJS := $(addprefix $(BUILD_DIR),$(HIP_OBJS:.cpp=.hip.$(OBJ_EXT)))
CUDA_FLAGS := \ CUDA_FLAGS := \
--generate-code=arch=compute_50,code=[compute_50,sm_50] \ --generate-code=arch=compute_50,code=[compute_50,sm_50] \
@ -111,64 +116,72 @@ ifeq ($(OS), Linux)
CUDA_FLAGS += -fPIC CUDA_FLAGS += -fPIC
CFLAGS += -Xcompiler -fPIC -D_GNU_SOURCE CFLAGS += -Xcompiler -fPIC -D_GNU_SOURCE
CXXFLAGS += -Xcompiler -fPIC -D_GNU_SOURCE CXXFLAGS += -Xcompiler -fPIC -D_GNU_SOURCE
else ifneq (,$(findstring MINGW,$(OS))) else ifeq ($(OS),Windows)
HIP_FLAGS += -Xclang --dependent-lib=msvcrt HIP_FLAGS += -Xclang --dependent-lib=msvcrt
endif endif
RUNNERS := ollama_runner$(EXE_EXT) RUNNERS := $(BUILD_DIR)ollama_runner$(EXE_EXT)
ifeq ($(ARCH),amd64) ifeq ($(ARCH),amd64)
RUNNERS += ollama_runner_avx$(EXE_EXT) ollama_runner_avx2$(EXE_EXT) RUNNERS += $(BUILD_DIR)ollama_runner_avx$(EXE_EXT) $(BUILD_DIR)ollama_runner_avx2$(EXE_EXT)
endif endif
ifneq ($(NVCC),) ifneq ($(NVCC),)
RUNNERS += ollama_runner_cuda$(EXE_EXT) RUNNERS += $(BUILD_DIR)ollama_runner_cuda$(EXE_EXT)
endif endif
ifneq ($(HIPCC),) ifneq ($(HIPCC),)
RUNNERS += ollama_runner_rocm$(EXE_EXT) RUNNERS += $(BUILD_DIR)ollama_runner_rocm$(EXE_EXT)
endif endif
runners: $(RUNNERS) runners: $(RUNNERS)
%.cuda.$(OBJ_EXT): %.cu $(BUILD_DIR)%.cuda.$(OBJ_EXT): %.cu
@-mkdir -p $(dir $@)
$(NVCC) -c $(CUDA_FLAGS) -o $@ $< $(NVCC) -c $(CUDA_FLAGS) -o $@ $<
%.cuda.$(OBJ_EXT): %.c $(BUILD_DIR)%.cuda.$(OBJ_EXT): %.c
@-mkdir -p $(dir $@)
$(NVCC) -c $(CFLAGS) -o $@ $< $(NVCC) -c $(CFLAGS) -o $@ $<
%.cuda.$(OBJ_EXT): %.cpp $(BUILD_DIR)%.cuda.$(OBJ_EXT): %.cpp
@-mkdir -p $(dir $@)
$(NVCC) -c $(CXXFLAGS) -o $@ $< $(NVCC) -c $(CXXFLAGS) -o $@ $<
$(SHARED_PREFIX)ggml_cuda.$(SHARED_EXT): $(CUDA_OBJS) $(BUILD_DIR)$(SHARED_PREFIX)ggml_cuda.$(SHARED_EXT): $(CUDA_OBJS)
@-mkdir -p $(dir $@)
$(NVCC) --shared -lcuda -lcublas -lcudart -lcublasLt $(CUDA_FLAGS) $(CUDA_OBJS) -o $@ $(NVCC) --shared -lcuda -lcublas -lcudart -lcublasLt $(CUDA_FLAGS) $(CUDA_OBJS) -o $@
%.hip.$(OBJ_EXT): %.cu $(BUILD_DIR)%.hip.$(OBJ_EXT): %.cu
@-mkdir -p $(dir $@)
$(HIPCC) -c $(HIP_FLAGS) -o $@ $< $(HIPCC) -c $(HIP_FLAGS) -o $@ $<
%.hip.$(OBJ_EXT): %.c $(BUILD_DIR)%.hip.$(OBJ_EXT): %.c
@-mkdir -p $(dir $@)
$(HIPCC) -c $(CFLAGS) -o $@ $< $(HIPCC) -c $(CFLAGS) -o $@ $<
%.hip.$(OBJ_EXT): %.cpp $(BUILD_DIR)%.hip.$(OBJ_EXT): %.cpp
@-mkdir -p $(dir $@)
$(HIPCC) -c $(CXXFLAGS) -o $@ $< $(HIPCC) -c $(CXXFLAGS) -o $@ $<
$(SHARED_PREFIX)ggml_hipblas.$(SHARED_EXT): $(HIP_OBJS) $(BUILD_DIR)$(SHARED_PREFIX)ggml_hipblas.$(SHARED_EXT): $(HIP_OBJS)
@-mkdir -p $(dir $@)
$(HIPCC) --shared -lhipblas -lamdhip64 -lrocblas $(HIP_OBJS) -o $@ $(HIPCC) --shared -lhipblas -lamdhip64 -lrocblas $(HIP_OBJS) -o $@
ollama_runner$(EXE_EXT): $(BUILD_DIR)ollama_runner$(EXE_EXT):
CGO_ENABLED=1 GOARCH=$(ARCH) go build -ldflags "-s -w" -o $@ ./runner CGO_ENABLED=1 GOARCH=$(ARCH) go build -ldflags "-s -w" -o $@ ./runner
ollama_runner_avx$(EXE_EXT): $(BUILD_DIR)ollama_runner_avx$(EXE_EXT):
CGO_ENABLED=1 GOARCH=$(ARCH) go build -ldflags "-s -w" -tags avx -o $@ ./runner CGO_ENABLED=1 GOARCH=$(ARCH) go build -ldflags "-s -w" -tags avx -o $@ ./runner
ollama_runner_avx2$(EXE_EXT): $(BUILD_DIR)ollama_runner_avx2$(EXE_EXT):
CGO_ENABLED=1 GOARCH=$(ARCH) go build -ldflags "-s -w" -tags avx,avx2 -o $@ ./runner CGO_ENABLED=1 GOARCH=$(ARCH) go build -ldflags "-s -w" -tags avx,avx2 -o $@ ./runner
ollama_runner_cuda$(EXE_EXT): $(SHARED_PREFIX)ggml_cuda.$(SHARED_EXT) $(BUILD_DIR)ollama_runner_cuda$(EXE_EXT): $(BUILD_DIR)$(SHARED_PREFIX)ggml_cuda.$(SHARED_EXT)
CGO_ENABLED=1 GOARCH=$(ARCH) CGO_LDFLAGS=-L"$(CUDA_LIB)" go build -ldflags "-s -w" -tags avx,cuda -o $@ ./runner CGO_ENABLED=1 GOARCH=$(ARCH) CGO_LDFLAGS=-L"$(CUDA_LIB)" go build -ldflags "-s -w" -tags avx,cuda -o $@ ./runner
ollama_runner_rocm$(EXE_EXT): $(SHARED_PREFIX)ggml_hipblas.$(SHARED_EXT) $(BUILD_DIR)ollama_runner_rocm$(EXE_EXT): $(BUILD_DIR)$(SHARED_PREFIX)ggml_hipblas.$(SHARED_EXT)
CGO_ENABLED=1 GOARCH=$(ARCH) CGO_LDFLAGS=-L"$(HIP_LIB)" go build -ldflags "-s -w" -tags avx,rocm -o $@ ./runner CGO_ENABLED=1 GOARCH=$(ARCH) CGO_LDFLAGS=-L"$(HIP_LIB)" go build -ldflags "-s -w" -tags avx,rocm -o $@ ./runner
clean: clean:
rm -f $(CUDA_OBJS) $(HIP_OBJS) $(SHARED_PREFIX)ggml_cuda.$(SHARED_EXT) ggml_cuda.* $(SHARED_PREFIX)ggml_hipblas.* ollama_runner* rm -rf $(BUILD_DIR)
.PHONY: runners clean ollama_runner$(EXE_EXT) ollama_runner_avx$(EXE_EXT) ollama_runner_avx2$(EXE_EXT) ollama_runner_cuda$(EXE_EXT) ollama_runner_rocm$(EXE_EXT) .PHONY: runners clean ollama_runner$(EXE_EXT) ollama_runner_avx$(EXE_EXT) ollama_runner_avx2$(EXE_EXT) ollama_runner_cuda$(EXE_EXT) ollama_runner_rocm$(EXE_EXT)

11
llama/llama.go
View File

@ -9,13 +9,18 @@ package llama
#cgo darwin,amd64 CFLAGS: -Wno-incompatible-pointer-types-discards-qualifiers #cgo darwin,amd64 CFLAGS: -Wno-incompatible-pointer-types-discards-qualifiers
#cgo darwin,amd64 CXXFLAGS: -Wno-incompatible-pointer-types-discards-qualifiers #cgo darwin,amd64 CXXFLAGS: -Wno-incompatible-pointer-types-discards-qualifiers
#cgo darwin,amd64 LDFLAGS: -framework Foundation #cgo darwin,amd64 LDFLAGS: -framework Foundation
#cgo darwin,amd64 LDFLAGS: -L${SRCDIR}/build/Darwin/amd64
#cgo darwin,amd64,avx2 CFLAGS: -DGGML_USE_ACCELERATE -DACCELERATE_NEW_LAPACK -DACCELERATE_LAPACK_ILP64 #cgo darwin,amd64,avx2 CFLAGS: -DGGML_USE_ACCELERATE -DACCELERATE_NEW_LAPACK -DACCELERATE_LAPACK_ILP64
#cgo darwin,amd64,avx2 CXXFLAGS: -DGGML_USE_ACCELERATE -DACCELERATE_NEW_LAPACK -DACCELERATE_LAPACK_ILP64 #cgo darwin,amd64,avx2 CXXFLAGS: -DGGML_USE_ACCELERATE -DACCELERATE_NEW_LAPACK -DACCELERATE_LAPACK_ILP64
#cgo darwin,amd64,avx2 LDFLAGS: -framework Accelerate #cgo darwin,amd64,avx2 LDFLAGS: -framework Accelerate
#cgo linux CFLAGS: -D_GNU_SOURCE #cgo linux CFLAGS: -D_GNU_SOURCE
#cgo linux CXXFLAGS: -D_GNU_SOURCE #cgo linux CXXFLAGS: -D_GNU_SOURCE
#cgo linux,arm64 LDFLAGS: -L${SRCDIR}/build/Linux/arm64
#cgo linux,amd64 LDFLAGS: -L${SRCDIR}/build/Linux/amd64
#cgo windows CFLAGS: -Wno-discarded-qualifiers #cgo windows CFLAGS: -Wno-discarded-qualifiers
#cgo windows LDFLAGS: -lmsvcrt #cgo windows LDFLAGS: -lmsvcrt
#cgo windows,arm64 LDFLAGS: -L${SRCDIR}/build/Windows/arm64
#cgo windows,amd64 LDFLAGS: -L${SRCDIR}/build/Windows/amd64
#cgo avx CFLAGS: -mavx #cgo avx CFLAGS: -mavx
#cgo avx CXXFLAGS: -mavx #cgo avx CXXFLAGS: -mavx
#cgo avx2 CFLAGS: -mavx2 -mfma #cgo avx2 CFLAGS: -mavx2 -mfma
@ -25,9 +30,9 @@ package llama
#cgo rocm CFLAGS: -DGGML_USE_CUDA -DGGML_USE_HIPBLAS -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_CUDA_MMV_Y=1 -DGGML_BUILD=1 #cgo rocm CFLAGS: -DGGML_USE_CUDA -DGGML_USE_HIPBLAS -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_CUDA_MMV_Y=1 -DGGML_BUILD=1
#cgo rocm CXXFLAGS: -DGGML_USE_CUDA -DGGML_USE_HIPBLAS -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_CUDA_MMV_Y=1 -DGGML_BUILD=1 #cgo rocm CXXFLAGS: -DGGML_USE_CUDA -DGGML_USE_HIPBLAS -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_CUDA_MMV_Y=1 -DGGML_BUILD=1
#cgo rocm LDFLAGS: -L${SRCDIR} -lggml_hipblas -lhipblas -lamdhip64 -lrocblas #cgo rocm LDFLAGS: -L${SRCDIR} -lggml_hipblas -lhipblas -lamdhip64 -lrocblas
#cgo windows,cuda LDFLAGS: -L${SRCDIR} -lggml_cuda -lcuda -lcudart -lcublas -lcublasLt #cgo windows,cuda LDFLAGS: -lggml_cuda -lcuda -lcudart -lcublas -lcublasLt
#cgo windows,rocm LDFLAGS: -L${SRCDIR} -lggml_hipblas -lhipblas -lamdhip64 -lrocblas #cgo windows,rocm LDFLAGS: -lggml_hipblas -lhipblas -lamdhip64 -lrocblas
#cgo linux,cuda LDFLAGS: -L${SRCDIR} -L/usr/local/cuda/lib64 -lggml_cuda -lcuda -lcudart -lcublas -lcublasLt -lpthread -ldl -lrt #cgo linux,cuda LDFLAGS: -L/usr/local/cuda/lib64 -lggml_cuda -lcuda -lcudart -lcublas -lcublasLt -lpthread -ldl -lrt
#cgo linux,rocm LDFLAGS: -L/opt/rocm/lib #cgo linux,rocm LDFLAGS: -L/opt/rocm/lib
#include <stdlib.h> #include <stdlib.h>