Refine build to support CPU only

If someone checks out the ollama repo and doesn't install the CUDA
library, this will ensure they can build a CPU only version

Dockerfile.cpu

@@ -0,0 +1,35 @@
+# Dockerfile variant to ensure we can build CPU only on linux
+FROM --platform=linux/amd64 ubuntu:20.04 AS base-cpu-amd64
+ENV CMAKE_ARCH "x86_64"
+
+FROM --platform=linux/arm64 ubuntu:20.04 AS base-cpu-arm64
+ENV CMAKE_ARCH "aarch64"
+
+FROM base-cpu-${TARGETARCH} AS cpu-builder
+ARG TARGETARCH
+ARG GOFLAGS
+ARG CGO_CFLAGS
+
+# Common toolchain
+RUN apt-get update && \
+    DEBIAN_FRONTEND=noninteractive apt-get install -y wget make gcc-10 g++-10 cpp-10 git ocl-icd-opencl-dev && \
+    update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 100 --slave /usr/bin/g++ g++ /usr/bin/g++-10 --slave /usr/bin/gcov gcov /usr/bin/gcov-10
+RUN wget "https://github.com/Kitware/CMake/releases/download/v3.27.6/cmake-3.27.6-linux-${CMAKE_ARCH}.sh" -O /tmp/cmake-installer.sh && \
+    chmod +x /tmp/cmake-installer.sh && /tmp/cmake-installer.sh --skip-license --prefix=/usr
+
+# install go
+ADD https://dl.google.com/go/go1.21.3.linux-$TARGETARCH.tar.gz /tmp/go1.21.3.tar.gz
+RUN mkdir -p /usr/local && tar xz -C /usr/local </tmp/go1.21.3.tar.gz
+
+# build the final binary
+FROM cpu-builder AS cpu-build
+WORKDIR /go/src/github.com/jmorganca/ollama
+COPY . .
+
+ENV GOOS=linux
+ENV GOARCH=$TARGETARCH
+ENV GOFLAGS=$GOFLAGS
+ENV CGO_CFLAGS=${CGO_CFLAGS}
+
+RUN /usr/local/go/bin/go generate ./... && \
+    /usr/local/go/bin/go build .

README.md

@@ -200,35 +200,8 @@ Then build the binary:
 go build .
 ```
 
-### Linux/Windows CUDA (NVIDIA)
-*Your operating system distribution may already have packages for NVIDIA CUDA. Distro packages are often preferable, but instructions are distro-specific. Please consult distro-specific docs for dependencies if available!*
+More detailed instructions can be found in the [developer guide](https://github.com/jmorganca/ollama/blob/main/docs/development.md)
 
-Note: at present, Ollama is optimized for GPU usage on linux, and requires the CUDA libraries at a minimum to compile even if you do not have an NVIDIA GPU.
-
-Install `cmake` and `golang` as well as [NVIDIA CUDA](https://developer.nvidia.com/cuda-downloads) development and runtime packages.
-Then generate dependencies:
-```
-go generate ./...
-```
-Then build the binary:
-```
-go build .
-```
-
-### Linux ROCm (AMD)
-*Your operating system distribution may already have packages for AMD ROCm and CLBlast. Distro packages are often preferable, but instructions are distro-specific. Please consult distro-specific docs for dependencies if available!*
-
-Install [CLBlast](https://github.com/CNugteren/CLBlast/blob/master/doc/installation.md) and [ROCm](https://rocm.docs.amd.com/en/latest/deploy/linux/quick_start.html) developement packages first, as well as `cmake` and `golang`.
-Adjust the paths below (correct for Arch) as appropriate for your distributions install locations and generate dependencies:
-```
-CLBlast_DIR=/usr/lib/cmake/CLBlast ROCM_PATH=/opt/rocm go generate ./...
-```
-Then build the binary:
-```
-go build .
-```
-
-ROCm requires elevated privileges to access the GPU at runtime.  On most distros you can add your user account to the `render` group, or run as root.
 
 ### Running local builds
 Next, start the server:

docs/development.md

@@ -34,6 +34,35 @@ Now you can run `ollama`:
 
 ## Building on Linux with GPU support
 
-- Install cmake and nvidia-cuda-toolkit
-- run `go generate ./...`
-- run `go build .`
+
+### Linux/Windows CUDA (NVIDIA)
+*Your operating system distribution may already have packages for NVIDIA CUDA. Distro packages are often preferable, but instructions are distro-specific. Please consult distro-specific docs for dependencies if available!*
+
+Install `cmake` and `golang` as well as [NVIDIA CUDA](https://developer.nvidia.com/cuda-downloads) development and runtime packages.
+Then generate dependencies:
+```
+go generate ./...
+```
+Then build the binary:
+```
+go build .
+```
+
+### Linux ROCm (AMD)
+*Your operating system distribution may already have packages for AMD ROCm and CLBlast. Distro packages are often preferable, but instructions are distro-specific. Please consult distro-specific docs for dependencies if available!*
+
+Install [CLBlast](https://github.com/CNugteren/CLBlast/blob/master/doc/installation.md) and [ROCm](https://rocm.docs.amd.com/en/latest/deploy/linux/quick_start.html) developement packages first, as well as `cmake` and `golang`.
+Adjust the paths below (correct for Arch) as appropriate for your distributions install locations and generate dependencies:
+```
+CLBlast_DIR=/usr/lib/cmake/CLBlast ROCM_PATH=/opt/rocm go generate ./...
+```
+Then build the binary:
+```
+go build .
+```
+
+ROCm requires elevated privileges to access the GPU at runtime.  On most distros you can add your user account to the `render` group, or run as root.
+
+## Containerized Build
+
+If you have Docker available, you can build linux binaries with `./scripts/build_linux.sh` which has the CUDA and ROCm dependencies included.

gpu/gpu.go

@@ -3,6 +3,9 @@
 package gpu
 
 /*
+#cgo linux LDFLAGS: -lrt -lpthread -ldl -lstdc++ -lm
+#cgo windows LDFLAGS: -lpthread
+
 #include "gpu_info.h"
 
 */
@@ -26,6 +29,7 @@ var gpuHandles *handles = nil
 
 // Note: gpuMutex must already be held
 func initGPUHandles() {
+	// TODO - if the ollama build is CPU only, don't do these checks as they're irrelevant and confusing
 	log.Printf("Detecting GPU type")
 	gpuHandles = &handles{nil, nil}
 	var resp C.cuda_init_resp_t
@@ -61,20 +65,32 @@ func GetGPUInfo() GpuInfo {
 	}
 
 	var memInfo C.mem_info_t
-	var resp GpuInfo
+	resp := GpuInfo{"", 0, 0}
 	if gpuHandles.cuda != nil {
 		C.cuda_check_vram(*gpuHandles.cuda, &memInfo)
-		resp.Driver = "CUDA"
+		if memInfo.err != nil {
+			log.Printf("error looking up CUDA GPU memory: %s", C.GoString(memInfo.err))
+			C.free(unsafe.Pointer(memInfo.err))
+		} else {
+			resp.Driver = "CUDA"
+		}
 	} else if gpuHandles.rocm != nil {
 		C.rocm_check_vram(*gpuHandles.rocm, &memInfo)
-		resp.Driver = "ROCM"
-	} else {
+		if memInfo.err != nil {
+			log.Printf("error looking up ROCm GPU memory: %s", C.GoString(memInfo.err))
+			C.free(unsafe.Pointer(memInfo.err))
+		} else {
+			resp.Driver = "ROCM"
+		}
+	}
+	if resp.Driver == "" {
 		C.cpu_check_ram(&memInfo)
 		resp.Driver = "CPU"
 	}
 	if memInfo.err != nil {
-		log.Printf("error looking up GPU memory: %s", C.GoString(memInfo.err))
+		log.Printf("error looking up CPU memory: %s", C.GoString(memInfo.err))
 		C.free(unsafe.Pointer(memInfo.err))
+		return resp
 	}
 	resp.FreeMemory = uint64(memInfo.free)
 	resp.TotalMemory = uint64(memInfo.total)
@@ -108,12 +124,7 @@ func NumGPU(numLayer, fileSizeBytes int64, opts api.Options) int {
 	// 75% of the absolute max number of layers we can fit in available VRAM, off-loading too many layers to the GPU can cause OOM errors
 	layers := int(info.FreeMemory/bytesPerLayer) * 3 / 4
 
-	// TODO - not sure on this part... if we can't fit all the layers, just fallback to CPU
-	// if int64(layers) < numLayer {
-	// 	log.Printf("%d MB VRAM available, insufficient to load current model (reuires %d MB) - falling back to CPU %d", freeBytes/(1024*1024), fileSizeBytes/(1024*1024))
-	// 	return 0
-	// }
-	log.Printf("%d MB VRAM available, loading up to %d GPU layers out of %d", info.FreeMemory/(1024*1024), layers, numLayer)
+	log.Printf("%d MB VRAM available, loading up to %d %s GPU layers out of %d", info.FreeMemory/(1024*1024), layers, info.Driver, numLayer)
 
 	return layers
 }

gpu/gpu_info_cuda.c

@@ -19,6 +19,7 @@ const char *cuda_lib_paths[] = {
 #endif
 
 void cuda_init(cuda_init_resp_t *resp) {
+  nvmlReturn_t ret;
   resp->err = NULL;
   const int buflen = 256;
   char buf[buflen + 1];
@@ -56,6 +57,13 @@ void cuda_init(cuda_init_resp_t *resp) {
       return;
     }
   }
+
+  ret = (*resp->ch.initFn)();
+  if (ret != NVML_SUCCESS) {
+    snprintf(buf, buflen, "nvml vram init failure: %d", ret);
+    resp->err = strdup(buf);
+  }
+
   return;
 }
 
@@ -73,17 +81,9 @@ void cuda_check_vram(cuda_handle_t h, mem_info_t *resp) {
     return;
   }
 
-  ret = (*h.initFn)();
-  if (ret != NVML_SUCCESS) {
-    snprintf(buf, buflen, "nvml vram init failure: %d", ret);
-    resp->err = strdup(buf);
-    return;
-  }
-
   // TODO - handle multiple GPUs
   ret = (*h.getHandle)(0, &device);
   if (ret != NVML_SUCCESS) {
-    (*h.shutdownFn)();
     snprintf(buf, buflen, "unable to get device handle: %d", ret);
     resp->err = strdup(buf);
     return;
@@ -91,20 +91,12 @@ void cuda_check_vram(cuda_handle_t h, mem_info_t *resp) {
 
   ret = (*h.getMemInfo)(device, &memInfo);
   if (ret != NVML_SUCCESS) {
-    (*h.shutdownFn)();
     snprintf(buf, buflen, "device memory info lookup failure: %d", ret);
     resp->err = strdup(buf);
     return;
   }
   resp->total = memInfo.total;
   resp->free = memInfo.free;
-
-  ret = (*h.shutdownFn)();
-  if (ret != NVML_SUCCESS) {
-    snprintf(buf, buflen, "nvml vram shutdown failure: %d", ret);
-    resp->err = strdup(buf);
-  }
-
   return;
 }
 #endif  // __APPLE__

gpu/gpu_info_rocm.c

@@ -20,6 +20,7 @@ const char *rocm_lib_paths[] = {
 #endif
 
 void rocm_init(rocm_init_resp_t *resp) {
+  rsmi_status_t ret;
   resp->err = NULL;
   const int buflen = 256;
   char buf[buflen + 1];
@@ -56,6 +57,13 @@ void rocm_init(rocm_init_resp_t *resp) {
       return;
     }
   }
+
+  ret = (*resp->rh.initFn)(0);
+  if (ret != RSMI_STATUS_SUCCESS) {
+    snprintf(buf, buflen, "rocm vram init failure: %d", ret);
+    resp->err = strdup(buf);
+  }
+
   return;
 }
 
@@ -70,10 +78,8 @@ void rocm_check_vram(rocm_handle_t h, mem_info_t *resp) {
   char buf[buflen + 1];
   int i;
 
-  ret = (*h.initFn)(0);
-  if (ret != RSMI_STATUS_SUCCESS) {
-    snprintf(buf, buflen, "rocm vram init failure: %d", ret);
-    resp->err = strdup(buf);
+  if (h.handle == NULL) {
+    resp->err = strdup("nvml handle sn't initialized");
     return;
   }
 
@@ -89,20 +95,17 @@ void rocm_check_vram(rocm_handle_t h, mem_info_t *resp) {
   // Get total memory - used memory for available memory
   ret = (*h.totalMemFn)(0, RSMI_MEM_TYPE_VRAM, &totalMem);
   if (ret != RSMI_STATUS_SUCCESS) {
-    (*h.shutdownFn)();
     snprintf(buf, buflen, "rocm total mem lookup failure: %d", ret);
     resp->err = strdup(buf);
     return;
   }
   ret = (*h.usageMemFn)(0, RSMI_MEM_TYPE_VRAM, &usedMem);
   if (ret != RSMI_STATUS_SUCCESS) {
-    (*h.shutdownFn)();
     snprintf(buf, buflen, "rocm usage mem lookup failure: %d", ret);
     resp->err = strdup(buf);
     return;
   }
 
-  (*h.shutdownFn)();
   resp->total = totalMem;
   resp->free = totalMem - usedMem;
   return;

llm/ext_server.go

@@ -21,17 +21,7 @@ package llm
 #cgo linux CFLAGS: -D_GNU_SOURCE
 #cgo linux windows CFLAGS: -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_MMV_Y=1 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_USE_CUBLAS
 #cgo linux LDFLAGS: -L/usr/local/cuda/targets/x86_64-linux/lib -L/usr/local/cuda/lib64 -L/usr/local/cuda/targets/x86_64-linux/lib/stubs
-#cgo linux LDFLAGS: ${SRCDIR}/llama.cpp/gguf/build/cuda/examples/server/libext_server.a
-#cgo linux LDFLAGS: ${SRCDIR}/llama.cpp/gguf/build/cuda/common/libcommon.a
-#cgo linux LDFLAGS: ${SRCDIR}/llama.cpp/gguf/build/cuda/libllama.a
-#cgo linux LDFLAGS: ${SRCDIR}/llama.cpp/gguf/build/cuda/libggml_static.a
-
-// Note: the following requires cuda library presence on linux to build, even if you only have rocm or CPU only
-#cgo linux LDFLAGS: /usr/local/cuda/lib64/libcudart_static.a
-#cgo linux LDFLAGS: /usr/local/cuda/lib64/libcublas_static.a
-#cgo linux LDFLAGS: /usr/local/cuda/lib64/libcublasLt_static.a
-#cgo linux LDFLAGS: /usr/local/cuda/lib64/libcudadevrt.a
-#cgo linux LDFLAGS: /usr/local/cuda/lib64/libculibos.a
+#cgo linux LDFLAGS: ${SRCDIR}/llama.cpp/gguf/build/cuda/libollama.a
 #cgo linux LDFLAGS: -lrt -lpthread -ldl -lstdc++ -lm
 #cgo windows LDFLAGS: -L${SRCDIR}/llama.cpp/gguf/build/wincuda/dist/bin
 #cgo windows LDFLAGS: -lext_server_shared -lpthread

llm/llama.cpp/gen_linux.sh

@@ -13,28 +13,43 @@ source $(dirname $0)/gen_common.sh
 init_vars
 git_module_setup
 apply_patches
-CMAKE_DEFS="-DLLAMA_CUBLAS=on -DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+if [ -d /usr/local/cuda/lib64/ ] ; then
+    CMAKE_DEFS="-DLLAMA_CUBLAS=on -DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+else
+    CMAKE_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+fi
 BUILD_DIR="gguf/build/cuda"
 LIB_DIR="${BUILD_DIR}/lib"
 mkdir -p ../../dist/
 build
-# TODO - explore mechanism to soften the hard cuda dependency on linux
-#        by conditionally building some archive here that aggregates the cuda libs if present
-#        so that the cgo flags link this intermediate archive instead of the underlying cuda libs
-# 
-# gcc -fPIC -g -shared -o ${LIB_DIR}/libcuda_server.so \
-#     -Wl,--whole-archive \
-#     ${BUILD_DIR}/examples/server/CMakeFiles/ext_server.dir/server.cpp.o \
-#     ${BUILD_DIR}/common/libcommon.a \
-#     ${BUILD_DIR}/libllama.a \
-#     ${BUILD_DIR}/examples/llava/libllava_static.a \
-#     -Wl,--no-whole-archive \
-#     -lrt -lpthread -ldl -lstdc++ -lm \
-#     /usr/local/cuda/lib64/libcudart_static.a \
-#     /usr/local/cuda/lib64/libcublas_static.a \
-#     /usr/local/cuda/lib64/libcublasLt_static.a \
-#     /usr/local/cuda/lib64/libcudadevrt.a \
-#     /usr/local/cuda/lib64/libculibos.a
+
+if [ -d /usr/local/cuda/lib64/ ] ; then
+    pwd
+    ar -M <<EOF
+create ${BUILD_DIR}/libollama.a
+addlib ${BUILD_DIR}/examples/server/libext_server.a
+addlib ${BUILD_DIR}/common/libcommon.a
+addlib ${BUILD_DIR}/libllama.a
+addlib ${BUILD_DIR}/libggml_static.a
+addlib /usr/local/cuda/lib64/libcudart_static.a
+addlib /usr/local/cuda/lib64/libcublas_static.a
+addlib /usr/local/cuda/lib64/libcublasLt_static.a
+addlib /usr/local/cuda/lib64/libcudadevrt.a
+addlib /usr/local/cuda/lib64/libculibos.a
+save
+end
+EOF
+else
+    ar -M <<EOF
+create ${BUILD_DIR}/libollama.a
+addlib ${BUILD_DIR}/examples/server/libext_server.a
+addlib ${BUILD_DIR}/common/libcommon.a
+addlib ${BUILD_DIR}/libllama.a
+addlib ${BUILD_DIR}/libggml_static.a
+save
+end
+EOF
+fi
 
 if [ -z "${ROCM_PATH}" ] ; then
     # Try the default location in case it exists

scripts/build_linux.sh

@@ -8,8 +8,14 @@ export GOFLAGS="'-ldflags=-w -s \"-X=github.com/jmorganca/ollama/version.Version
 mkdir -p dist
 
 for TARGETARCH in amd64 arm64; do
-    docker buildx build --load --progress=plain --platform=linux/$TARGETARCH --build-arg=VERSION --build-arg=GOFLAGS -f Dockerfile.build -t builder:$TARGETARCH .
-    docker create --platform linux/$TARGETARCH --name builder-$TARGETARCH builder:$TARGETARCH
-    docker cp builder-$TARGETARCH:/go/src/github.com/jmorganca/ollama/ollama ./dist/ollama-linux-$TARGETARCH
-    docker rm builder-$TARGETARCH
+    docker buildx build --load --progress=plain --platform=linux/$TARGETARCH --build-arg=VERSION --build-arg=GOFLAGS -f Dockerfile.build -t gpubuilder:$TARGETARCH .
+    docker create --platform linux/$TARGETARCH --name gpubuilder-$TARGETARCH gpubuilder:$TARGETARCH
+    docker cp gpubuilder-$TARGETARCH:/go/src/github.com/jmorganca/ollama/ollama ./dist/ollama-linux-$TARGETARCH
+    docker rm gpubuilder-$TARGETARCH
+
+    docker buildx build --load --progress=plain --platform=linux/$TARGETARCH --build-arg=VERSION --build-arg=GOFLAGS -f Dockerfile.cpu -t cpubuilder:$TARGETARCH .
+    docker create --platform linux/$TARGETARCH --name cpubuilder-$TARGETARCH cpubuilder:$TARGETARCH
+    docker cp cpubuilder-$TARGETARCH:/go/src/github.com/jmorganca/ollama/ollama ./dist/ollama-linux-$TARGETARCH-cpu
+    docker rm cpubuilder-$TARGETARCH
+
 done