Adapted rocm support to cgo based llama.cpp

Dockerfile

@@ -11,8 +11,8 @@ RUN mkdir -p /usr/local && tar xz -C /usr/local </tmp/go1.21.3.tar.gz
 COPY . .
 ENV GOARCH=$TARGETARCH
 ENV GOFLAGS=$GOFLAGS
-RUN /usr/local/go/bin/go generate -tags cuda ./... \
-    && /usr/local/go/bin/go build -tags cuda .
+RUN /usr/local/go/bin/go generate ./... \
+    && /usr/local/go/bin/go build .
 
 FROM ubuntu:22.04
 RUN apt-get update && apt-get install -y ca-certificates
@@ -27,5 +27,3 @@ ENV NVIDIA_DRIVER_CAPABILITIES=compute,utility
 
 ENTRYPOINT ["/bin/ollama"]
 CMD ["serve"]
-
-

Dockerfile.build

@@ -1,19 +1,44 @@
-# centos7 amd64 dependencies
-FROM --platform=linux/amd64 nvidia/cuda:11.3.1-devel-centos7 AS base-amd64
-RUN yum install -y https://repo.ius.io/ius-release-el7.rpm centos-release-scl && \
-    yum update -y && \
-    yum install -y devtoolset-10-gcc devtoolset-10-gcc-c++ git236 wget
-RUN wget "https://github.com/Kitware/CMake/releases/download/v3.27.6/cmake-3.27.6-linux-x86_64.sh" -O cmake-installer.sh && chmod +x cmake-installer.sh && ./cmake-installer.sh --skip-license --prefix=/usr/local
-ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
+# Ubuntu 20.04 amd64 dependencies
+FROM --platform=linux/amd64 nvidia/cuda:11.7.1-devel-ubuntu22.04 AS base-amd64
+# ROCm only supports amd64
+ARG ROCM_VERSION=5.7
+# Note: https://rocm.docs.amd.com/en/latest/release/user_kernel_space_compat_matrix.html
+RUN apt-get update && \
+    apt-get install -y wget && \
+    wget "https://github.com/Kitware/CMake/releases/download/v3.22.1/cmake-3.22.1-linux-x86_64.sh" -O /tmp/cmake-installer.sh && \
+    chmod +x /tmp/cmake-installer.sh && /tmp/cmake-installer.sh --skip-license --prefix=/usr && \
+    mkdir --parents --mode=0755 /etc/apt/keyrings && \
+    wget https://repo.radeon.com/rocm/rocm.gpg.key -O - | gpg --dearmor > /etc/apt/keyrings/rocm.gpg && \
+    echo "deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/${ROCM_VERSION} focal main" > /etc/apt/sources.list.d/rocm.list && \
+    echo "Package: *" > /etc/apt/preferences.d/rocm-pin-600 && \
+    echo "Pin: release o=repo.radeon.com" >> /etc/apt/preferences.d/rocm-pin-600 && \
+    echo "Pin-Priority: 600" >> /etc/apt/preferences.d/rocm-pin-600 && \
+    apt-get update && \
+    DEBIAN_FRONTEND=noninteractive apt-get install -y rocm-hip-libraries rocm-device-libs rocm-libs rocm-ocl-icd rocm-hip-sdk rocm-hip-libraries rocm-cmake rocm-clang-ocl rocm-dev
 
-# centos8 arm64 dependencies
-FROM --platform=linux/arm64 nvidia/cuda-arm64:11.3.1-devel-centos8 AS base-arm64
-RUN sed -i -e 's/mirrorlist/#mirrorlist/g' -e 's|#baseurl=http://mirror.centos.org|baseurl=http://vault.centos.org|g' /etc/yum.repos.d/CentOS-*
-RUN yum install -y git cmake
+ENV ROCM_PATH=/opt/rocm
+
+# Ubuntu 22.04 arm64 dependencies
+FROM --platform=linux/arm64 nvidia/cuda:11.7.1-devel-ubuntu22.04 AS base-arm64
+RUN apt-get update && \
+    apt-get install -y wget && \
+    wget "https://github.com/Kitware/CMake/releases/download/v3.27.6/cmake-3.27.6-linux-aarch64.sh" -O /tmp/cmake-installer.sh && \
+    chmod +x /tmp/cmake-installer.sh && /tmp/cmake-installer.sh --skip-license --prefix=/usr
 
 FROM base-${TARGETARCH}
 ARG TARGETARCH
 ARG GOFLAGS="'-ldflags -w -s'"
+ARG CGO_CFLAGS
+ARG CLBLAST_VER=1.6.1
+
+# Common toolchain
+RUN apt-get update && \
+    DEBIAN_FRONTEND=noninteractive apt-get install -y gcc-11 g++-11 cpp-11 git ocl-icd-opencl-dev && \
+    update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-11 100 --slave /usr/bin/g++ g++ /usr/bin/g++-11 --slave /usr/bin/gcov gcov /usr/bin/gcov-11
+
+# CLBlast
+RUN wget -qO- https://github.com/CNugteren/CLBlast/archive/refs/tags/${CLBLAST_VER}.tar.gz | tar zxv -C /tmp/ && \
+    cd /tmp/CLBlast-${CLBLAST_VER} && mkdir build && cd build && cmake .. && make && make install
 
 # install go
 ADD https://dl.google.com/go/go1.21.3.linux-$TARGETARCH.tar.gz /tmp/go1.21.3.tar.gz
@@ -26,6 +51,7 @@ COPY . .
 ENV GOOS=linux
 ENV GOARCH=$TARGETARCH
 ENV GOFLAGS=$GOFLAGS
+ENV CGO_CFLAGS=${CGO_CFLAGS}
 
-RUN /usr/local/go/bin/go generate -tags cuda ./... && \
-    /usr/local/go/bin/go build -tags cuda .
+RUN /usr/local/go/bin/go generate ./... && \
+    /usr/local/go/bin/go build .

README.md

@@ -185,8 +185,6 @@ ollama list
 
 ## Building
 
-### Generic (CPU)
-
 Install `cmake` and `go`:
 
 ```
@@ -202,32 +200,36 @@ Then build the binary:
 go build .
 ```
 
-### CUDA (NVIDIA)
+### 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://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html) development and runtime packages.
+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 -tags cuda ./...
+go generate ./...
 ```
 Then build the binary:
 ```
-go build -tags cuda .
+go build .
 ```
 
-### ROCm (AMD)
+### 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 -tags rocm ./...
+CLBlast_DIR=/usr/lib/cmake/CLBlast ROCM_PATH=/opt/rocm go generate ./...
 ```
 Then build the binary:
 ```
-go build -tags rocm
+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:
 

gpu/gpu.go

@@ -0,0 +1,119 @@
+//go:build linux || windows
+
+package gpu
+
+/*
+#include "gpu_info.h"
+
+*/
+import "C"
+import (
+	"fmt"
+	"log"
+	"sync"
+	"unsafe"
+
+	"github.com/jmorganca/ollama/api"
+)
+
+type handles struct {
+	cuda *C.cuda_handle_t
+	rocm *C.rocm_handle_t
+}
+
+var gpuMutex sync.Mutex
+var gpuHandles *handles = nil
+
+// Note: gpuMutex must already be held
+func initGPUHandles() {
+	log.Printf("Detecting GPU type")
+	gpuHandles = &handles{nil, nil}
+	var resp C.cuda_init_resp_t
+	C.cuda_init(&resp)
+	if resp.err != nil {
+		log.Printf("CUDA not detected: %s", C.GoString(resp.err))
+		C.free(unsafe.Pointer(resp.err))
+
+		var resp C.rocm_init_resp_t
+		C.rocm_init(&resp)
+		if resp.err != nil {
+			log.Printf("ROCm not detected: %s", C.GoString(resp.err))
+			C.free(unsafe.Pointer(resp.err))
+		} else {
+			log.Printf("Radeon GPU detected")
+			rocm := resp.rh
+			gpuHandles.rocm = &rocm
+		}
+	} else {
+		log.Printf("Nvidia GPU detected")
+		cuda := resp.ch
+		gpuHandles.cuda = &cuda
+	}
+}
+
+func GetGPUInfo() GpuInfo {
+	// TODO - consider exploring lspci (and equivalent on windows) to check for
+	// GPUs so we can report warnings if we see Nvidia/AMD but fail to load the libraries
+	gpuMutex.Lock()
+	defer gpuMutex.Unlock()
+	if gpuHandles == nil {
+		initGPUHandles()
+	}
+
+	var memInfo C.mem_info_t
+	var resp GpuInfo
+	if gpuHandles.cuda != nil {
+		C.cuda_check_vram(*gpuHandles.cuda, &memInfo)
+		resp.Driver = "CUDA"
+	} else if gpuHandles.rocm != nil {
+		C.rocm_check_vram(*gpuHandles.rocm, &memInfo)
+		resp.Driver = "ROCM"
+	} else {
+		C.cpu_check_ram(&memInfo)
+		resp.Driver = "CPU"
+	}
+	if memInfo.err != nil {
+		log.Printf("error looking up GPU memory: %s", C.GoString(memInfo.err))
+		C.free(unsafe.Pointer(memInfo.err))
+	}
+	resp.FreeMemory = uint64(memInfo.free)
+	resp.TotalMemory = uint64(memInfo.total)
+	return resp
+}
+
+func CheckVRAM() (int64, error) {
+	gpuInfo := GetGPUInfo()
+	if gpuInfo.FreeMemory > 0 && gpuInfo.Driver != "CPU" {
+		return int64(gpuInfo.FreeMemory), nil
+	}
+	return 0, fmt.Errorf("no GPU detected") // TODO - better handling of CPU based memory determiniation
+}
+
+func NumGPU(numLayer, fileSizeBytes int64, opts api.Options) int {
+	if opts.NumGPU != -1 {
+		return opts.NumGPU
+	}
+	info := GetGPUInfo()
+	if info.Driver == "CPU" {
+		return 0
+	}
+
+	/*
+		Calculate bytes per layer, this will roughly be the size of the model file divided by the number of layers.
+		We can store the model weights and the kv cache in vram,
+		to enable kv chache vram storage add two additional layers to the number of layers retrieved from the model file.
+	*/
+	bytesPerLayer := uint64(fileSizeBytes / numLayer)
+
+	// 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)
+
+	return layers
+}

gpu/gpu_darwin.go

@@ -1,7 +1,8 @@
 //go:build darwin
 
-package llm
+package gpu
 
+import "C"
 import (
 	"github.com/jmorganca/ollama/api"
 )
@@ -9,11 +10,25 @@ import (
 // CheckVRAM returns the free VRAM in bytes on Linux machines with NVIDIA GPUs
 func CheckVRAM() (int64, error) {
 	// TODO - assume metal, and return free memory?
-	return 0, errNvidiaSMI
+	return 0, nil
 
 }
 
+func GetGPUInfo() GpuInfo {
+	// TODO - Metal vs. x86 macs...
+
+	return GpuInfo{
+		Driver:      "METAL",
+		TotalMemory: 0,
+		FreeMemory:  0,
+	}
+}
+
 func NumGPU(numLayer, fileSizeBytes int64, opts api.Options) int {
 	// default to enable metal on macOS
 	return 1
 }
+
+func nativeInit() error {
+	return nil
+}

gpu/gpu_info.h

@@ -0,0 +1,49 @@
+#ifndef __APPLE__
+#ifndef __GPU_INFO_H__
+#define __GPU_INFO_H__
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+#ifndef _WIN32
+#include <dlfcn.h>
+#define LOAD_LIBRARY(lib, flags) dlopen(lib, flags)
+#define LOAD_SYMBOL(handle, sym) dlsym(handle, sym)
+#define LOAD_ERR() dlerror()
+#define UNLOAD_LIBRARY(handle) dlclose(handle)
+#else
+#include <windows.h>
+#define LOAD_LIBRARY(lib, flags) LoadLibrary(lib)
+#define LOAD_SYMBOL(handle, sym) GetProcAddress(handle, sym)
+#define UNLOAD_LIBRARY(handle) FreeLibrary(handle)
+
+// TODO - refactor this with proper error message handling on windows
+inline static char *LOAD_ERR() {
+  static char errbuf[8];
+  snprintf(errbuf, 8, "0x%lx", GetLastError());
+  return errbuf;
+}
+
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct mem_info {
+  uint64_t total;
+  uint64_t free;
+  char *err;  // If non-nill, caller responsible for freeing
+} mem_info_t;
+
+void cpu_check_ram(mem_info_t *resp);
+
+#ifdef __cplusplus
+}
+#endif
+
+#include "gpu_info_cuda.h"
+#include "gpu_info_rocm.h"
+
+#endif  // __GPU_INFO_H__
+#endif  // __APPLE__

gpu/gpu_info_cpu.c

@@ -0,0 +1,42 @@
+#include "gpu_info.h"
+// Fallbacks for CPU mode
+
+#ifdef _WIN32
+#include <sysinfoapi.h>
+void cpu_check_ram(mem_info_t *resp) {
+  resp->err = NULL;
+  MEMORYSTATUSEX info;
+  if (GlobalMemoryStatusEx(&info) != 0) {
+    resp->total = info.ullTotalPhys;
+    resp->free = info.ullAvailPhys;
+  } else {
+    resp->err = strdup(LOAD_ERR());
+  }
+  return;
+}
+
+#elif __linux__
+#include <errno.h>
+#include <string.h>
+#include <sys/sysinfo.h>
+void cpu_check_ram(mem_info_t *resp) {
+  struct sysinfo info;
+  resp->err = NULL;
+  if (sysinfo(&info) != 0) {
+    resp->err = strdup(strerror(errno));
+  } else {
+    resp->total = info.totalram * info.mem_unit;
+    resp->free = info.freeram * info.mem_unit;
+  }
+  return;
+}
+
+#elif __APPLE__
+// TODO consider an Apple implementation that does something useful
+// mem_info_t cpu_check_ram() {
+//   mem_info_t resp = {0, 0, NULL};
+//   return resp;
+// }
+#else
+#error "Unsupported platform"
+#endif

gpu/gpu_info_cuda.c

@@ -0,0 +1,110 @@
+#ifndef __APPLE__  // TODO - maybe consider nvidia support on intel macs?
+
+#include "gpu_info_cuda.h"
+
+#include <string.h>
+
+#ifndef _WIN32
+const char *cuda_lib_paths[] = {
+    "libnvidia-ml.so",
+    "/usr/local/cuda/lib64/libnvidia-ml.so",
+    NULL,
+};
+#else
+const char *cuda_lib_paths[] = {
+    "nvml.dll",
+    "",
+    NULL,
+};
+#endif
+
+void cuda_init(cuda_init_resp_t *resp) {
+  resp->err = NULL;
+  const int buflen = 256;
+  char buf[buflen + 1];
+  int i;
+
+  struct lookup {
+    char *s;
+    void **p;
+  } l[4] = {
+      {"nvmlInit_v2", (void *)&resp->ch.initFn},
+      {"nvmlShutdown", (void *)&resp->ch.shutdownFn},
+      {"nvmlDeviceGetHandleByIndex", (void *)&resp->ch.getHandle},
+      {"nvmlDeviceGetMemoryInfo", (void *)&resp->ch.getMemInfo},
+  };
+
+  for (i = 0; cuda_lib_paths[i] != NULL && resp->ch.handle == NULL; i++) {
+    resp->ch.handle = LOAD_LIBRARY(cuda_lib_paths[i], RTLD_LAZY);
+  }
+  if (!resp->ch.handle) {
+    snprintf(buf, buflen,
+             "Unable to load %s library to query for Nvidia GPUs: %s",
+             cuda_lib_paths[0], LOAD_ERR());
+    resp->err = strdup(buf);
+    return;
+  }
+
+  for (i = 0; i < 4; i++) {  // TODO - fix this to use a null terminated list
+    *l[i].p = LOAD_SYMBOL(resp->ch.handle, l[i].s);
+    if (!l[i].p) {
+      UNLOAD_LIBRARY(resp->ch.handle);
+      resp->ch.handle = NULL;
+      snprintf(buf, buflen, "symbol lookup for %s failed: %s", l[i].s,
+               LOAD_ERR());
+      resp->err = strdup(buf);
+      return;
+    }
+  }
+  return;
+}
+
+void cuda_check_vram(cuda_handle_t h, mem_info_t *resp) {
+  resp->err = NULL;
+  nvmlDevice_t device;
+  nvmlMemory_t memInfo = {0};
+  nvmlReturn_t ret;
+  const int buflen = 256;
+  char buf[buflen + 1];
+  int i;
+
+  if (h.handle == NULL) {
+    resp->err = strdup("nvml handle sn't initialized");
+    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;
+  }
+
+  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_cuda.h

@@ -0,0 +1,35 @@
+#ifndef __APPLE__
+#ifndef __GPU_INFO_CUDA_H__
+#define __GPU_INFO_CUDA_H__
+#include "gpu_info.h"
+
+// Just enough typedef's to dlopen/dlsym for memory information
+typedef enum nvmlReturn_enum {
+  NVML_SUCCESS = 0,
+  // Other values omitted for now...
+} nvmlReturn_t;
+typedef void *nvmlDevice_t;  // Opaque is sufficient
+typedef struct nvmlMemory_st {
+  unsigned long long total;
+  unsigned long long free;
+  unsigned long long used;
+} nvmlMemory_t;
+
+typedef struct cuda_handle {
+  void *handle;
+  nvmlReturn_t (*initFn)(void);
+  nvmlReturn_t (*shutdownFn)(void);
+  nvmlReturn_t (*getHandle)(unsigned int, nvmlDevice_t *);
+  nvmlReturn_t (*getMemInfo)(nvmlDevice_t, nvmlMemory_t *);
+} cuda_handle_t;
+
+typedef struct cuda_init_resp {
+  char *err;  // If err is non-null handle is invalid
+  cuda_handle_t ch;
+} cuda_init_resp_t;
+
+void cuda_init(cuda_init_resp_t *resp);
+void cuda_check_vram(cuda_handle_t ch, mem_info_t *resp);
+
+#endif  // __GPU_INFO_CUDA_H__
+#endif  // __APPLE__

gpu/gpu_info_rocm.c

@@ -0,0 +1,111 @@
+#ifndef __APPLE__
+
+#include "gpu_info_rocm.h"
+
+#include <string.h>
+
+#ifndef _WIN32
+const char *rocm_lib_paths[] = {
+    "librocm_smi64.so",
+    "/opt/rocm/lib/librocm_smi64.so",
+    NULL,
+};
+#else
+// TODO untested
+const char *rocm_lib_paths[] = {
+    "rocm_smi64.dll",
+    "/opt/rocm/lib/rocm_smi64.dll",
+    NULL,
+};
+#endif
+
+void rocm_init(rocm_init_resp_t *resp) {
+  resp->err = NULL;
+  const int buflen = 256;
+  char buf[buflen + 1];
+  int i;
+  struct lookup {
+    char *s;
+    void **p;
+  } l[4] = {
+      {"rsmi_init", (void *)&resp->rh.initFn},
+      {"rsmi_shut_down", (void *)&resp->rh.shutdownFn},
+      {"rsmi_dev_memory_total_get", (void *)&resp->rh.totalMemFn},
+      {"rsmi_dev_memory_usage_get", (void *)&resp->rh.usageMemFn},
+      // { "rsmi_dev_id_get", (void*)&resp->rh.getHandle },
+  };
+
+  for (i = 0; rocm_lib_paths[i] != NULL && resp->rh.handle == NULL; i++) {
+    resp->rh.handle = LOAD_LIBRARY(rocm_lib_paths[i], RTLD_LAZY);
+  }
+  if (!resp->rh.handle) {
+    snprintf(buf, buflen,
+             "Unable to load %s library to query for Radeon GPUs: %s\n",
+             rocm_lib_paths[0], LOAD_ERR());
+    resp->err = strdup(buf);
+    return;
+  }
+
+  for (i = 0; i < 4; i++) {
+    *l[i].p = LOAD_SYMBOL(resp->rh.handle, l[i].s);
+    if (!l[i].p) {
+      UNLOAD_LIBRARY(resp->rh.handle);
+      snprintf(buf, buflen, "symbol lookup for %s failed: %s", l[i].s,
+               LOAD_ERR());
+      resp->err = strdup(buf);
+      return;
+    }
+  }
+  return;
+}
+
+void rocm_check_vram(rocm_handle_t h, mem_info_t *resp) {
+  resp->err = NULL;
+  // uint32_t num_devices;
+  // uint16_t device;
+  uint64_t totalMem = 0;
+  uint64_t usedMem = 0;
+  rsmi_status_t ret;
+  const int buflen = 256;
+  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);
+    return;
+  }
+
+  // TODO - iterate through devices...  ret =
+  // rsmi_num_monitor_devices(&num_devices);
+
+  // ret = (*h.getHandle)(0, &device);
+  // if (ret != RSMI_STATUS_SUCCESS) {
+  //     printf("rocm vram device lookup failure: %d\n", ret);
+  //     return -1;
+  // }
+
+  // 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;
+}
+
+#endif  // __APPLE__

gpu/gpu_info_rocm.h

@@ -0,0 +1,36 @@
+#ifndef __APPLE__
+#ifndef __GPU_INFO_ROCM_H__
+#define __GPU_INFO_ROCM_H__
+#include "gpu_info.h"
+
+// Just enough typedef's to dlopen/dlsym for memory information
+typedef enum rsmi_status_return {
+  RSMI_STATUS_SUCCESS = 0,
+  // Other values omitted for now...
+} rsmi_status_t;
+
+typedef enum rsmi_memory_type {
+  RSMI_MEM_TYPE_VRAM = 0,
+  RSMI_MEM_TYPE_VIS_VRAM,
+  RSMI_MEM_TYPE_GTT,
+} rsmi_memory_type_t;
+
+typedef struct rocm_handle {
+  void *handle;
+  rsmi_status_t (*initFn)(uint64_t);
+  rsmi_status_t (*shutdownFn)(void);
+  rsmi_status_t (*totalMemFn)(uint32_t, rsmi_memory_type_t, uint64_t *);
+  rsmi_status_t (*usageMemFn)(uint32_t, rsmi_memory_type_t, uint64_t *);
+  // rsmi_status_t (*getHandle)(uint32_t, uint16_t *);
+} rocm_handle_t;
+
+typedef struct rocm_init_resp {
+  char *err;  // If err is non-null handle is invalid
+  rocm_handle_t rh;
+} rocm_init_resp_t;
+
+void rocm_init(rocm_init_resp_t *resp);
+void rocm_check_vram(rocm_handle_t rh, mem_info_t *resp);
+
+#endif  // __GPU_INFO_ROCM_H__
+#endif  // __APPLE__

gpu/gpu_test.go

@@ -0,0 +1,26 @@
+package gpu
+
+import (
+	"runtime"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+)
+
+func TestBasicGetGPUInfo(t *testing.T) {
+	info := GetGPUInfo()
+	assert.Contains(t, "CUDA ROCM CPU METAL", info.Driver)
+
+	switch runtime.GOOS {
+	case "darwin":
+		// TODO - remove this once MacOS returns some size for CPU
+		return
+	case "linux", "windows":
+		assert.Greater(t, info.TotalMemory, uint64(0))
+		assert.Greater(t, info.FreeMemory, uint64(0))
+	default:
+		return
+	}
+}
+
+// TODO - add some logic to figure out card type through other means and actually verify we got back what we expected

gpu/types.go

@@ -0,0 +1,10 @@
+package gpu
+
+// Beginning of an `ollama info` command
+type GpuInfo struct {
+	Driver      string `json:"driver,omitempty"`
+	TotalMemory uint64 `json:"total_memory,omitempty"`
+	FreeMemory  uint64 `json:"free_memory,omitempty"`
+
+	// TODO add other useful attributes about the card here for discovery information
+}

llm/accelerator_cuda.go

@@ -1,67 +0,0 @@
-//go:build cuda
-
-package llm
-
-import (
-	"bufio"
-	"bytes"
-	"errors"
-	"fmt"
-	"log"
-	"os/exec"
-	"path"
-	"strconv"
-	"strings"
-
-	"github.com/jmorganca/ollama/format"
-)
-
-var (
-	errNvidiaSMI     = errors.New("warning: gpu support may not be enabled, check that you have installed GPU drivers: nvidia-smi command failed")
-	errAvailableVRAM = errors.New("not enough VRAM available, falling back to CPU only")
-)
-
-// acceleratedRunner returns the runner for this accelerator given the provided buildPath string.
-func acceleratedRunner(buildPath string) []ModelRunner {
-	return []ModelRunner{
-		ModelRunner{
-			Path:        path.Join(buildPath, "cuda", "bin", "ollama-runner"),
-			Accelerated: true,
-		},
-	}
-}
-
-// CheckVRAM returns the free VRAM in bytes on Linux machines with NVIDIA GPUs
-func CheckVRAM() (int64, error) {
-	cmd := exec.Command("nvidia-smi", "--query-gpu=memory.free", "--format=csv,noheader,nounits")
-	var stdout bytes.Buffer
-	cmd.Stdout = &stdout
-	err := cmd.Run()
-	if err != nil {
-		return 0, errNoAccel
-	}
-
-	var freeMiB int64
-	scanner := bufio.NewScanner(&stdout)
-	for scanner.Scan() {
-		line := scanner.Text()
-		if strings.Contains(line, "[Insufficient Permissions]") {
-			return 0, fmt.Errorf("GPU support may not enabled, check you have installed GPU drivers and have the necessary permissions to run nvidia-smi")
-		}
-
-		vram, err := strconv.ParseInt(strings.TrimSpace(line), 10, 64)
-		if err != nil {
-			return 0, fmt.Errorf("failed to parse available VRAM: %v", err)
-		}
-
-		freeMiB += vram
-	}
-
-	freeBytes := freeMiB * 1024 * 1024
-	if freeBytes < 2*format.GigaByte {
-		log.Printf("less than 2 GB VRAM available")
-		return 0, errAvailableVRAM
-	}
-
-	return freeBytes, nil
-}

llm/accelerator_none.go

@@ -1,21 +0,0 @@
-//go:build !rocm && !cuda
-
-package llm
-
-import (
-	"errors"
-)
-
-var (
-	errNoAccel = errors.New("no accelerator support in this binary")
-)
-
-// acceleratedRunner returns the runner for this accelerator given the provided buildPath string.
-func acceleratedRunner(buildPath string) []ModelRunner {
-	return make([]ModelRunner, 0, 1)
-}
-
-// CheckVRAM is a stub with no accelerator.
-func CheckVRAM() (int64, error) {
-	return 0, errNoGPU
-}

llm/accelerator_rocm.go

@@ -1,85 +0,0 @@
-//go:build rocm
-
-package llm
-
-import (
-	"bytes"
-	"encoding/csv"
-	"errors"
-	"fmt"
-	"io"
-	"log"
-	"os"
-	"os/exec"
-	"path"
-	"path/filepath"
-	"strconv"
-	"strings"
-)
-
-var errNoAccel = errors.New("rocm-smi command failed")
-
-// acceleratedRunner returns the runner for this accelerator given the provided buildPath string.
-func acceleratedRunner(buildPath string) []ModelRunner {
-	return []ModelRunner{
-		ModelRunner{
-			Path:        path.Join(buildPath, "rocm", "bin", "ollama-runner"),
-			Accelerated: true,
-		},
-	}
-}
-
-// CheckVRAM returns the available VRAM in MiB on Linux machines with AMD GPUs
-func CheckVRAM() (int64, error) {
-	rocmHome := os.Getenv("ROCM_PATH")
-	if rocmHome == "" {
-		rocmHome = os.Getenv("ROCM_HOME")
-	}
-	if rocmHome == "" {
-		log.Println("warning: ROCM_PATH is not set. Trying a likely fallback path, but it is recommended to set this variable in the environment.")
-		rocmHome = "/opt/rocm"
-	}
-	cmd := exec.Command(filepath.Join(rocmHome, "bin/rocm-smi"), "--showmeminfo", "VRAM", "--csv")
-	var stdout bytes.Buffer
-	cmd.Stdout = &stdout
-	err := cmd.Run()
-	if err != nil {
-		return 0, errNoAccel
-	}
-	csvData := csv.NewReader(&stdout)
-	// llama.cpp or ROCm don't seem to understand splitting the VRAM allocations across them properly, so try to find the biggest card instead :(. FIXME.
-	totalBiggestCard := int64(0)
-	bigCardName := ""
-	for {
-		record, err := csvData.Read()
-		if err == io.EOF {
-			break
-		}
-		if err != nil {
-			return 0, fmt.Errorf("failed to parse available VRAM: %v", err)
-		}
-		if !strings.HasPrefix(record[0], "card") {
-			continue
-		}
-		cardTotal, err := strconv.ParseInt(record[1], 10, 64)
-		if err != nil {
-			return 0, err
-		}
-		cardUsed, err := strconv.ParseInt(record[2], 10, 64)
-		if err != nil {
-			return 0, err
-		}
-		possible := (cardTotal - cardUsed)
-		log.Printf("ROCm found %d MiB of available VRAM on device %q", possible/1024/1024, record[0])
-		if possible > totalBiggestCard {
-			totalBiggestCard = possible
-			bigCardName = record[0]
-		}
-	}
-	if totalBiggestCard == 0 {
-		log.Printf("found ROCm GPU but failed to parse free VRAM!")
-		return 0, errNoAccel
-	}
-	log.Printf("ROCm selecting device %q", bigCardName)
-	return totalBiggestCard, nil
-}

llm/ext_server.go

@@ -1,7 +1,7 @@
 package llm
 
 /*
-#cgo CFLAGS: -I${SRCDIR}/llama.cpp/gguf -I${SRCDIR}/llama.cpp/gguf/common
+#cgo CFLAGS: -I${SRCDIR}/llama.cpp/gguf -I${SRCDIR}/llama.cpp/gguf/common -I${SRCDIR}/llama.cpp/gguf/examples/server
 #cgo CFLAGS: -DNDEBUG -DLLAMA_SERVER_LIBRARY=1 -D_XOPEN_SOURCE=600 -DACCELERATE_NEW_LAPACK -DACCELERATE_LAPACK_ILP64
 #cgo CFLAGS: -Wmissing-noreturn -Wall -Wextra -Wcast-qual -Wno-unused-function -Wno-array-bounds
 #cgo CPPFLAGS: -Ofast -Wall -Wextra -Wno-unused-function -Wno-unused-variable -Wno-deprecated-declarations -Wno-unused-but-set-variable
@@ -25,6 +25,8 @@ package llm
 #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
@@ -35,7 +37,7 @@ package llm
 #cgo windows LDFLAGS: -lext_server_shared -lpthread
 
 #include <stdlib.h>
-#include "examples/server/server.h"
+#include "server.h"
 
 */
 import "C"
@@ -43,25 +45,51 @@ import (
 	"bytes"
 	"context"
 	"encoding/json"
-	"errors"
 	"fmt"
 	"log"
 	"os"
 	"runtime"
+	"strings"
 	"sync"
 	"time"
 	"unsafe"
 
 	"github.com/jmorganca/ollama/api"
+	"github.com/jmorganca/ollama/gpu"
 )
 
-func errWrap(resp C.ext_server_err) error {
-	if resp.code == 0 {
-		return nil
+func newExtServerResp(len C.size_t) C.ext_server_resp_t {
+	var resp C.ext_server_resp_t
+	resp.msg_len = len
+	bytes := make([]byte, len)
+	resp.msg = (*C.char)(C.CBytes(bytes))
+	return resp
+}
+
+func freeExtServerResp(resp C.ext_server_resp_t) {
+	if resp.msg_len == 0 {
+		return
 	}
-	err := fmt.Errorf(C.GoString(resp.err))
-	C.free(unsafe.Pointer(resp.err))
-	return err
+	C.free(unsafe.Pointer(resp.msg))
+}
+
+func extServerResponseToErr(resp C.ext_server_resp_t) error {
+	return fmt.Errorf(C.GoString(resp.msg))
+}
+
+type extServer interface {
+	LLM
+	llama_server_init(sparams *C.ext_server_params_t, err *C.ext_server_resp_t)
+	llama_server_start()
+	llama_server_stop()
+	llama_server_completion(json_req *C.char, resp *C.ext_server_resp_t)
+	llama_server_completion_next_result(task_id C.int, resp *C.ext_server_task_result_t)
+	llama_server_completion_cancel(task_id C.int, err *C.ext_server_resp_t)
+	llama_server_release_task_result(result *C.ext_server_task_result_t)
+	llama_server_tokenize(json_req *C.char, json_resp **C.char, err *C.ext_server_resp_t)
+	llama_server_detokenize(json_req *C.char, json_resp **C.char, err *C.ext_server_resp_t)
+	llama_server_embedding(json_req *C.char, json_resp **C.char, err *C.ext_server_resp_t)
+	llama_server_release_json_resp(json_resp **C.char)
 }
 
 type llamaExtServer struct {
@@ -71,21 +99,61 @@ type llamaExtServer struct {
 // Note: current implementation does not support concurrent instantiations
 var mutex sync.Mutex
 
-func newLlamaExtServer(model string, adapters, projectors []string, numLayers int64, opts api.Options) (*llamaExtServer, error) {
+func (llm *llamaExtServer) llama_server_init(sparams *C.ext_server_params_t, err *C.ext_server_resp_t) {
+	C.llama_server_init(sparams, err)
+}
+func (llm *llamaExtServer) llama_server_start() {
+	C.llama_server_start()
+}
+func (llm *llamaExtServer) llama_server_stop() {
+	C.llama_server_stop()
+}
+
+func (llm *llamaExtServer) llama_server_completion(json_req *C.char, resp *C.ext_server_resp_t) {
+	C.llama_server_completion(json_req, resp)
+}
+func (llm *llamaExtServer) llama_server_completion_next_result(task_id C.int, resp *C.ext_server_task_result_t) {
+	C.llama_server_completion_next_result(task_id, resp)
+}
+func (llm *llamaExtServer) llama_server_completion_cancel(task_id C.int, err *C.ext_server_resp_t) {
+	C.llama_server_completion_cancel(task_id, err)
+}
+func (llm *llamaExtServer) llama_server_release_task_result(result *C.ext_server_task_result_t) {
+	C.llama_server_release_task_result(result)
+}
+
+func (llm *llamaExtServer) llama_server_tokenize(json_req *C.char, json_resp **C.char, err *C.ext_server_resp_t) {
+	C.llama_server_tokenize(json_req, json_resp, err)
+}
+func (llm *llamaExtServer) llama_server_detokenize(json_req *C.char, json_resp **C.char, err *C.ext_server_resp_t) {
+	C.llama_server_detokenize(json_req, json_resp, err)
+}
+func (llm *llamaExtServer) llama_server_embedding(json_req *C.char, json_resp **C.char, err *C.ext_server_resp_t) {
+	C.llama_server_embedding(json_req, json_resp, err)
+}
+func (llm *llamaExtServer) llama_server_release_json_resp(json_resp **C.char) {
+	C.llama_server_release_json_resp(json_resp)
+}
+
+func newLlamaExtServer(model string, adapters, projectors []string, numLayers int64, opts api.Options) (extServer, error) {
+	server := &llamaExtServer{opts}
+	return newExtServer(server, model, adapters, projectors, numLayers, opts)
+}
+
+func newExtServer(server extServer, model string, adapters, projectors []string, numLayers int64, opts api.Options) (extServer, error) {
 	if !mutex.TryLock() {
 		log.Printf("concurrent llm servers not yet supported, waiting for prior server to complete")
 		mutex.Lock()
 	}
-	server := &llamaExtServer{opts}
 	fileInfo, err := os.Stat(model)
 	if err != nil {
 		return nil, err
 	}
-	var sparams C.ext_server_params
+	var sparams C.ext_server_params_t
 	sparams.model = C.CString(model)
 	defer C.free(unsafe.Pointer(sparams.model))
 
-	numGPU := NumGPU(numLayers, fileInfo.Size(), opts)
+	numGPU := gpu.NumGPU(numLayers, fileInfo.Size(), opts)
 
 	sparams.embedding = true
 	sparams.n_ctx = C.uint(opts.NumCtx)
@@ -97,10 +165,14 @@ func newLlamaExtServer(model string, adapters, projectors []string, numLayers in
 	// Always use the value encoded in the model
 	sparams.rope_freq_base = 0.0
 	sparams.rope_freq_scale = 0.0
+	sparams.memory_f16 = C.bool(opts.F16KV)
+	sparams.use_mlock = C.bool(opts.UseMLock)
+	sparams.use_mmap = C.bool(opts.UseMMap)
+	sparams.numa = C.bool(opts.UseNUMA)
 
 	sparams.lora_adapters = nil
 	for i := 0; i < len(adapters); i++ {
-		la := (*C.ext_server_lora_adapter)(C.malloc(C.sizeof_struct_ext_server_lora_adapter))
+		la := (*C.ext_server_lora_adapter_t)(C.malloc(C.sizeof_ext_server_lora_adapter_t))
 		defer C.free(unsafe.Pointer(la))
 		la.adapter = C.CString(adapters[i])
 		defer C.free(unsafe.Pointer(la.adapter))
@@ -116,11 +188,13 @@ func newLlamaExtServer(model string, adapters, projectors []string, numLayers in
 		}
 	}
 
-	// TODO - implement ME
-	// if len(projectors) > 0 {
-	// 	// TODO: applying multiple projectors is not supported by the llama.cpp server yet
-	// 	params = append(params, "--mmproj", projectors[0])
-	// }
+	if len(projectors) > 0 {
+		// TODO: applying multiple projectors is not supported by the llama.cpp server yet
+		sparams.mmproj = C.CString(projectors[0])
+		defer C.free(unsafe.Pointer(sparams.mmproj))
+	} else {
+		sparams.mmproj = nil
+	}
 
 	if opts.NumThread > 0 {
 		sparams.n_threads = C.uint(opts.NumThread)
@@ -128,136 +202,167 @@ func newLlamaExtServer(model string, adapters, projectors []string, numLayers in
 		sparams.n_threads = C.uint(runtime.NumCPU())
 	}
 
-	sparams.memory_f16 = false
-	if opts.F16KV {
-		sparams.memory_f16 = true
-	}
-	sparams.use_mlock = false
-	if opts.UseMLock {
-		sparams.use_mlock = true
-	}
-	sparams.use_mmap = true
-	if !opts.UseMMap {
-		sparams.use_mmap = false
-	}
-	sparams.numa = false
-	if opts.UseNUMA {
-		sparams.numa = true
-	}
-
 	log.Printf("Initializing internal llama server")
-	err = errWrap(C.llama_server_init(&sparams))
-	if err != nil {
-		return nil, err
+	resp := newExtServerResp(128)
+	defer freeExtServerResp(resp)
+	server.llama_server_init(&sparams, &resp)
+	if resp.id < 0 {
+		return nil, extServerResponseToErr(resp)
 	}
 
 	log.Printf("Starting internal llama main loop")
-	C.llama_server_start()
+	server.llama_server_start()
 	return server, nil
 }
 
-func (llm *llamaExtServer) Predict(ctx context.Context, predict PredictOpts, fn func(PredictResult)) error {
+func (llm *llamaExtServer) Predict(ctx context.Context, pred PredictOpts, fn func(PredictResult)) error {
+	return predict(llm, llm.Options, ctx, pred, fn)
+}
+
+func predict(llm extServer, opts api.Options, ctx context.Context, predict PredictOpts, fn func(PredictResult)) error {
+	resp := newExtServerResp(128)
+	defer freeExtServerResp(resp)
+	var imageData []ImageData
+	if len(predict.Images) > 0 {
+		for cnt, i := range predict.Images {
+			imageData = append(imageData, ImageData{Data: i, ID: cnt})
+		}
+	}
+	log.Printf("loaded %d images", len(imageData))
 
 	request := map[string]any{
 		"prompt":            predict.Prompt,
 		"stream":            true,
-		"n_predict":         llm.NumPredict,
-		"n_keep":            llm.NumKeep,
-		"temperature":       llm.Temperature,
-		"top_k":             llm.TopK,
-		"top_p":             llm.TopP,
-		"tfs_z":             llm.TFSZ,
-		"typical_p":         llm.TypicalP,
-		"repeat_last_n":     llm.RepeatLastN,
-		"repeat_penalty":    llm.RepeatPenalty,
-		"presence_penalty":  llm.PresencePenalty,
-		"frequency_penalty": llm.FrequencyPenalty,
-		"mirostat":          llm.Mirostat,
-		"mirostat_tau":      llm.MirostatTau,
-		"mirostat_eta":      llm.MirostatEta,
-		"penalize_nl":       llm.PenalizeNewline,
-		"seed":              llm.Seed,
-		"stop":              llm.Stop,
+		"n_predict":         opts.NumPredict,
+		"n_keep":            opts.NumKeep,
+		"temperature":       opts.Temperature,
+		"top_k":             opts.TopK,
+		"top_p":             opts.TopP,
+		"tfs_z":             opts.TFSZ,
+		"typical_p":         opts.TypicalP,
+		"repeat_last_n":     opts.RepeatLastN,
+		"repeat_penalty":    opts.RepeatPenalty,
+		"presence_penalty":  opts.PresencePenalty,
+		"frequency_penalty": opts.FrequencyPenalty,
+		"mirostat":          opts.Mirostat,
+		"mirostat_tau":      opts.MirostatTau,
+		"mirostat_eta":      opts.MirostatEta,
+		"penalize_nl":       opts.PenalizeNewline,
+		"seed":              opts.Seed,
+		"stop":              opts.Stop,
+		"image_data":        imageData,
 	}
 
 	if predict.Format == "json" {
 		request["grammar"] = jsonGrammar
 	}
 
-	// Handling JSON marshaling with special characters unescaped.
-	buffer := &bytes.Buffer{}
-	enc := json.NewEncoder(buffer)
-	enc.SetEscapeHTML(false)
+	retryDelay := 100 * time.Microsecond
+	for retries := 0; retries < maxRetries; retries++ {
+		if retries > 0 {
+			time.Sleep(retryDelay) // wait before retrying
+			retryDelay *= 2        // exponential backoff
+		}
 
-	if err := enc.Encode(request); err != nil {
-		return fmt.Errorf("failed to marshal data: %w", err)
-	}
+		// Handling JSON marshaling with special characters unescaped.
+		buffer := &bytes.Buffer{}
+		enc := json.NewEncoder(buffer)
+		enc.SetEscapeHTML(false)
 
-	req := C.CString(buffer.String())
-	defer C.free(unsafe.Pointer(req))
+		if err := enc.Encode(request); err != nil {
+			return fmt.Errorf("failed to marshal data: %w", err)
+		}
 
-	cmpCtx := C.llama_server_completion(req)
-	if cmpCtx.task_id < 0 {
-		defer C.free(unsafe.Pointer(cmpCtx.err))
-		return fmt.Errorf(C.GoString(cmpCtx.err))
-	}
+		req := C.CString(buffer.String())
+		defer C.free(unsafe.Pointer(req))
 
-	for {
-		select {
-		case <-ctx.Done():
-			// This handles the request cancellation
-			return errWrap(C.llama_server_completion_cancel(cmpCtx.task_id))
-		default:
-			result := C.llama_server_completion_next_result(cmpCtx.task_id)
-			if result.result_json != nil {
-				defer C.free(unsafe.Pointer(result.result_json))
-			}
-			var p prediction
-			if err := json.Unmarshal([]byte(C.GoString(result.result_json)), &p); err != nil {
-				err2 := errWrap(C.llama_server_completion_cancel(cmpCtx.task_id))
-				return errors.Join(fmt.Errorf("error unmarshaling llm prediction response: %w", err), err2)
-			}
+		llm.llama_server_completion(req, &resp)
+		if resp.id < 0 {
+			return extServerResponseToErr(resp)
+		}
 
-			if p.Content != "" {
-				fn(PredictResult{
-					// Model:     predict.Model, // XXX remove or replace?
-					CreatedAt: time.Now().UTC(),
-					Content:   p.Content,
-				})
-			}
+		retryNeeded := false
+	out:
+		for {
+			select {
+			case <-ctx.Done():
+				// This handles the request cancellation
+				llm.llama_server_completion_cancel(resp.id, &resp)
+				if resp.id < 0 {
+					return extServerResponseToErr(resp)
+				} else {
+					return nil
+				}
+			default:
+				var result C.ext_server_task_result_t
+				llm.llama_server_completion_next_result(resp.id, &result)
+				json_resp := C.GoString(result.json_resp)
+				llm.llama_server_release_task_result(&result)
 
-			if p.Stop {
-				fn(PredictResult{
-					// Model:              predict.Model, // XXX remove or replace?
-					CreatedAt:          time.Now().UTC(),
-					TotalDuration:      time.Since(predict.CheckpointStart),
-					Done:               true,
-					PromptEvalCount:    p.Timings.PromptN,
-					PromptEvalDuration: parseDurationMs(p.Timings.PromptMS),
-					EvalCount:          p.Timings.PredictedN,
-					EvalDuration:       parseDurationMs(p.Timings.PredictedMS),
-				})
-				return nil
+				var p prediction
+				if err := json.Unmarshal([]byte(json_resp), &p); err != nil {
+					llm.llama_server_completion_cancel(resp.id, &resp)
+					if resp.id < 0 {
+						return fmt.Errorf("error unmarshaling llm prediction response: %w and cancel %s", err, C.GoString(resp.msg))
+					} else {
+						return fmt.Errorf("error unmarshaling llm prediction response: %w", err)
+					}
+				}
+
+				if bool(result.error) && strings.Contains(json_resp, "slot unavailable") {
+					retryNeeded = true
+					// task will already be canceled
+					break out
+				}
+
+				if p.Content != "" {
+					fn(PredictResult{
+						Content: p.Content,
+					})
+				}
+
+				if p.Stop {
+					fn(PredictResult{
+						Done:               true,
+						PromptEvalCount:    p.Timings.PromptN,
+						PromptEvalDuration: parseDurationMs(p.Timings.PromptMS),
+						EvalCount:          p.Timings.PredictedN,
+						EvalDuration:       parseDurationMs(p.Timings.PredictedMS),
+					})
+					return nil
+				}
 			}
 		}
+		if !retryNeeded {
+			return nil // success
+		}
 	}
+
+	// should never reach here ideally
+	return fmt.Errorf("max retries exceeded")
+}
+func (llm *llamaExtServer) Encode(ctx context.Context, prompt string) ([]int, error) {
+	return encode(llm, ctx, prompt)
 }
 
-func (llm *llamaExtServer) Encode(ctx context.Context, prompt string) ([]int, error) {
+func encode(llm extServer, ctx context.Context, prompt string) ([]int, error) {
 	data, err := json.Marshal(TokenizeRequest{Content: prompt})
 	if err != nil {
 		return nil, fmt.Errorf("marshaling encode data: %w", err)
 	}
 	req := C.CString(string(data))
 	defer C.free(unsafe.Pointer(req))
-	var resp C.ext_server_resp
-	err = errWrap(C.llama_server_tokenize(req, &resp))
-	if resp.json_resp != nil {
-		defer C.free(unsafe.Pointer(resp.json_resp))
+	var json_resp *C.char
+	resp := newExtServerResp(128)
+	defer freeExtServerResp(resp)
+	llm.llama_server_tokenize(req, &json_resp, &resp)
+	if resp.id < 0 {
+		return nil, extServerResponseToErr(resp)
 	}
+	defer llm.llama_server_release_json_resp(&json_resp)
 
 	var encoded TokenizeResponse
-	if err2 := json.Unmarshal([]byte(C.GoString(resp.json_resp)), &encoded); err2 != nil {
+	if err2 := json.Unmarshal([]byte(C.GoString(json_resp)), &encoded); err2 != nil {
 		return nil, fmt.Errorf("unmarshal encode response: %w", err2)
 	}
 
@@ -265,6 +370,10 @@ func (llm *llamaExtServer) Encode(ctx context.Context, prompt string) ([]int, er
 }
 
 func (llm *llamaExtServer) Decode(ctx context.Context, tokens []int) (string, error) {
+	return decode(llm, ctx, tokens)
+}
+
+func decode(llm extServer, ctx context.Context, tokens []int) (string, error) {
 	if len(tokens) == 0 {
 		return "", nil
 	}
@@ -275,14 +384,17 @@ func (llm *llamaExtServer) Decode(ctx context.Context, tokens []int) (string, er
 
 	req := C.CString(string(data))
 	defer C.free(unsafe.Pointer(req))
-	var resp C.ext_server_resp
-	err = errWrap(C.llama_server_detokenize(req, &resp))
-	if resp.json_resp != nil {
-		defer C.free(unsafe.Pointer(resp.json_resp))
+	var json_resp *C.char
+	resp := newExtServerResp(128)
+	defer freeExtServerResp(resp)
+	llm.llama_server_detokenize(req, &json_resp, &resp)
+	if resp.id < 0 {
+		return "", extServerResponseToErr(resp)
 	}
+	defer llm.llama_server_release_json_resp(&json_resp)
 
 	var decoded DetokenizeResponse
-	if err2 := json.Unmarshal([]byte(C.GoString(resp.json_resp)), &decoded); err2 != nil {
+	if err2 := json.Unmarshal([]byte(C.GoString(json_resp)), &decoded); err2 != nil {
 		return "", fmt.Errorf("unmarshal encode response: %w", err2)
 	}
 
@@ -290,6 +402,9 @@ func (llm *llamaExtServer) Decode(ctx context.Context, tokens []int) (string, er
 }
 
 func (llm *llamaExtServer) Embedding(ctx context.Context, input string) ([]float64, error) {
+	return embedding(llm, ctx, input)
+}
+func embedding(llm extServer, ctx context.Context, input string) ([]float64, error) {
 	data, err := json.Marshal(TokenizeRequest{Content: input})
 	if err != nil {
 		return nil, fmt.Errorf("error marshaling embed data: %w", err)
@@ -297,29 +412,28 @@ func (llm *llamaExtServer) Embedding(ctx context.Context, input string) ([]float
 
 	req := C.CString(string(data))
 	defer C.free(unsafe.Pointer(req))
-	var resp C.ext_server_resp
-	err = errWrap(C.llama_server_embedding(req, &resp))
-	if resp.json_resp != nil {
-		defer C.free(unsafe.Pointer(resp.json_resp))
-	}
-	if err != nil {
-		return nil, err
+	var json_resp *C.char
+	resp := newExtServerResp(128)
+	defer freeExtServerResp(resp)
+	llm.llama_server_embedding(req, &json_resp, &resp)
+	if resp.id < 0 {
+		return nil, extServerResponseToErr(resp)
 	}
+	defer llm.llama_server_release_json_resp(&json_resp)
 
 	var embedding EmbeddingResponse
-	if err := json.Unmarshal([]byte(C.GoString(resp.json_resp)), &embedding); err != nil {
+	if err := json.Unmarshal([]byte(C.GoString(json_resp)), &embedding); err != nil {
 		return nil, fmt.Errorf("unmarshal tokenize response: %w", err)
 	}
 
 	return embedding.Embedding, nil
 }
 
-func (llm *llamaExtServer) Ping(ctx context.Context) error {
-	// TODO - consider some mechanism to check if the main loop and llama.cpp are in a good state
-	return nil
+func (llm *llamaExtServer) Close() {
+	close(llm)
 }
 
-func (llm *llamaExtServer) Close() {
-	C.llama_server_stop()
+func close(llm extServer) {
+	llm.llama_server_stop()
 	mutex.Unlock()
 }

llm/gpu_cuda.go

@@ -1,57 +0,0 @@
-//go:build linux || windows
-
-package llm
-
-import (
-	"errors"
-	"log"
-
-	"github.com/jmorganca/ollama/api"
-)
-
-/*
-#cgo windows LDFLAGS: -L"/Program Files/NVIDIA Corporation/NVSMI/"
-#cgo linux LDFLAGS: -lnvidia-ml
-
-#include <stdlib.h>
-#include "examples/server/server.h"
-*/
-import "C"
-
-// CheckVRAM returns the free VRAM in bytes on Linux machines with NVIDIA GPUs
-func CheckVRAM() (int64, error) {
-	return int64(C.check_vram()), nil
-}
-
-func NumGPU(numLayer, fileSizeBytes int64, opts api.Options) int {
-	if opts.NumGPU != -1 {
-		return opts.NumGPU
-	}
-	freeBytes, err := CheckVRAM()
-	if err != nil {
-		if !errors.Is(err, errNvidiaSMI) {
-			log.Print(err.Error())
-		}
-		// nvidia driver not installed or no nvidia GPU found
-		return 0
-	}
-
-	/*
-		Calculate bytes per layer, this will roughly be the size of the model file divided by the number of layers.
-		We can store the model weights and the kv cache in vram,
-		to enable kv chache vram storage add two additional layers to the number of layers retrieved from the model file.
-	*/
-	bytesPerLayer := fileSizeBytes / numLayer
-
-	// 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(freeBytes/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", freeBytes/(1024*1024), layers, numLayer)
-
-	return layers
-}

llm/llama.cpp/gen_common.sh

@@ -1,10 +1,11 @@
 # common logic accross linux and darwin
 
 init_vars() {
+    LLAMACPP_DIR=gguf
     PATCHES="0001-Expose-callable-API-for-server.patch"
     CMAKE_DEFS="-DLLAMA_ACCELERATE=on"
     # TODO - LLAMA_K_QUANTS is stale and needs to be mapped to newer cmake settings
-    CMAKE_TARGETS="--target ggml --target ggml_static --target llama --target build_info --target common --target ext_server"
+    CMAKE_TARGETS="--target ggml --target ggml_static --target llama --target build_info --target common --target ext_server --target llava_static"
     if echo "${CGO_CFLAGS}" | grep -- '-g' > /dev/null ; then
         CMAKE_DEFS="-DCMAKE_BUILD_TYPE=RelWithDebInfo -DCMAKE_VERBOSE_MAKEFILE=on -DLLAMA_GPROF=on ${CMAKE_DEFS}"
     else
@@ -29,6 +30,6 @@ apply_patches() {
 }
 
 build() {
-    cmake -S gguf -B ${BUILD_DIR} ${CMAKE_DEFS}
-    cmake --build ${BUILD_DIR} ${CMAKE_TARGETS} -j8 
+    cmake -S ${LLAMACPP_DIR} -B ${BUILD_DIR} ${CMAKE_DEFS}
+    cmake --build ${BUILD_DIR} ${CMAKE_TARGETS} -j8
 }

llm/llama.cpp/gen_darwin.sh

@@ -1,4 +1,4 @@
-#!/bin/sh
+#!/bin/bash
 # This script is intended to run inside the go generate
 # working directory must be ../llm/llama.cpp
 
@@ -30,6 +30,7 @@ git_module_setup
 apply_patches
 build
 
+# TODO - improve this to handle test cases that need it to be in "." around the tree
 # Enable local debug/run usecase
 if [ -e "gguf/ggml-metal.metal" ]; then
     cp gguf/ggml-metal.metal ../../

llm/llama.cpp/gen_linux.sh

@@ -1,17 +1,73 @@
-#!/bin/sh
+#!/bin/bash
 # This script is intended to run inside the go generate
 # working directory must be ../llm/llama.cpp
 
 set -ex
 set -o pipefail
 
-# TODO - stopped here - map the variables from above over and refine the case statement below
-
 echo "Starting linux generate script"
+if [ -z "${CUDACXX}" -a -x /usr/local/cuda/bin/nvcc ] ; then
+    export CUDACXX=/usr/local/cuda/bin/nvcc
+fi
 source $(dirname $0)/gen_common.sh
 init_vars
-CMAKE_DEFS="-DLLAMA_CUBLAS=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
-BUILD_DIR="gguf/build/cuda"
 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 ${CMAKE_DEFS}"
+BUILD_DIR="gguf/build/cuda"
+LIB_DIR="${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 \
+#     ${BUILD_DIR}/examples/llava/libllava_static.a \
+#     -Wl,--no-whole-archive \
+#     -lrt -lpthread -ldl -lstdc++ -lm \
+#     /usr/local/cuda/lib64/libcudart_static.a \
+#     /usr/local/cuda/lib64/libcublas_static.a \
+#     /usr/local/cuda/lib64/libcublasLt_static.a \
+#     /usr/local/cuda/lib64/libcudadevrt.a \
+#     /usr/local/cuda/lib64/libculibos.a
+
+if [ -z "${ROCM_PATH}" ] ; then
+    # Try the default location in case it exists
+    ROCM_PATH=/opt/rocm
+fi
+
+if [ -z "${CLBlast_DIR}" ] ; then
+    # Try the default location in case it exists
+    if [ -d /usr/lib/cmake/CLBlast ]; then
+        export CLBlast_DIR=/usr/lib/cmake/CLBlast
+    fi
+fi
+
+BUILD_DIR="gguf/build/rocm"
+LIB_DIR="${BUILD_DIR}/lib"
+mkdir -p ${LIB_DIR}
+# Ensure we have at least one file present for the embed
+touch ${LIB_DIR}/.generated 
+
+if [ -d "${ROCM_PATH}" ] ; then
+    echo "Building ROCm"
+    init_vars
+    CMAKE_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DCMAKE_VERBOSE_MAKEFILE=on -DLLAMA_HIPBLAS=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS='gfx803;gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102' -DGPU_TARGETS='gfx803;gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102'"
+    CMAKE_DEFS="-DLLAMA_ACCELERATE=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+    build
+    gcc -fPIC -g -shared -o ${LIB_DIR}/librocm_server.so \
+        -Wl,--whole-archive \
+        ${BUILD_DIR}/examples/server/libext_server.a \
+        ${BUILD_DIR}/common/libcommon.a \
+        ${BUILD_DIR}/libllama.a \
+        -Wl,--no-whole-archive \
+        -lrt -lpthread -ldl -lstdc++ -lm \
+        -L/opt/rocm/lib -L/opt/amdgpu/lib/x86_64-linux-gnu/ \
+        -Wl,-rpath,/opt/rocm/lib,-rpath,/opt/amdgpu/lib/x86_64-linux-gnu/ \
+        -lhipblas -lrocblas -lamdhip64 -lrocsolver -lamd_comgr -lhsa-runtime64 -lrocsparse -ldrm -ldrm_amdgpu
+fi

llm/llama.cpp/gen_windows.ps1

@@ -48,4 +48,8 @@ init_vars
 git_module_setup
 apply_patches
 build
-install
+install
+
+# TODO - implement ROCm support on windows
+md gguf/build/winrocm/lib -ea 0
+echo $null >> gguf/build/winrocm/lib/.generated

llm/llama.cpp/generate_linux.go

@@ -1,3 +1,3 @@
 package llm
 
-//go:generate sh ./gen_linux.sh
+//go:generate bash ./gen_linux.sh

llm/llama.cpp/generate_linux_cuda.go

@@ -1,24 +0,0 @@
-//go:build cuda
-
-package llm
-
-//go:generate git submodule init
-
-//go:generate git submodule update --force ggml
-//go:generate git -C ggml apply ../patches/0001-add-detokenize-endpoint.patch
-//go:generate git -C ggml apply ../patches/0002-34B-model-support.patch
-//go:generate git -C ggml apply ../patches/0005-ggml-support-CUDA-s-half-type-for-aarch64-1455-2670.patch
-//go:generate git -C ggml apply ../patches/0001-copy-cuda-runtime-libraries.patch
-
-//go:generate rm -rf ggml/build/cuda
-//go:generate cmake -S ggml -B ggml/build/cuda -DLLAMA_CUBLAS=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on
-//go:generate cmake --build ggml/build/cuda --target server --config Release
-//go:generate mv ggml/build/cuda/bin/server ggml/build/cuda/bin/ollama-runner
-
-//go:generate git -C gguf apply ../patches/0001-copy-cuda-runtime-libraries.patch
-//go:generate git -C gguf apply ../patches/0001-update-default-log-target.patch
-
-//go:generate rm -rf gguf/build/cuda
-//go:generate cmake -S gguf -B gguf/build/cuda -DLLAMA_CUBLAS=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off
-//go:generate cmake --build gguf/build/cuda --target server --config Release
-//go:generate mv gguf/build/cuda/bin/server gguf/build/cuda/bin/ollama-runner

llm/llama.cpp/generate_linux_rocm.go

@@ -1,25 +0,0 @@
-//go:build rocm
-
-package llm
-
-//go:generate git submodule init
-
-//go:generate git submodule update --force ggml
-//go:generate git -C ggml apply ../patches/0001-add-detokenize-endpoint.patch
-//go:generate git -C ggml apply ../patches/0002-34B-model-support.patch
-//go:generate git -C ggml apply ../patches/0005-ggml-support-CUDA-s-half-type-for-aarch64-1455-2670.patch
-//go:generate git -C ggml apply ../patches/0001-copy-cuda-runtime-libraries.patch
-
-//go:generate git submodule update --force gguf
-//go:generate git -C gguf apply ../patches/0001-copy-cuda-runtime-libraries.patch
-//go:generate git -C gguf apply ../patches/0001-update-default-log-target.patch
-
-//go:generate rm -rf ggml/build/rocm
-//go:generate cmake -S ggml -B ggml/build/rocm -DLLAMA_CLBLAST=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on
-//go:generate cmake --build ggml/build/rocm --target server --config Release
-//go:generate mv ggml/build/rocm/bin/server ggml/build/rocm/bin/ollama-runner
-
-//go:generate rm -rf gguf/build/rocm
-//go:generate cmake -S gguf -B gguf/build/rocm -DLLAMA_HIPBLAS=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS='gfx803;gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102' -DGPU_TARGETS='gfx803;gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102'
-//go:generate cmake --build gguf/build/rocm --target server --config Release
-//go:generate mv gguf/build/rocm/bin/server gguf/build/rocm/bin/ollama-runner

llm/llama.cpp/patches/0001-Expose-callable-API-for-server.patch

@@ -1,15 +1,15 @@
-From 64b3fbb150d12b3ca63ac2fb4e57bc46f41d2ccd Mon Sep 17 00:00:00 2001
+From 087cf3300e973d7790db8f7cad01d2a790de38be Mon Sep 17 00:00:00 2001
 From: Daniel Hiltgen <daniel@ollama.com>
 Date: Mon, 13 Nov 2023 12:25:58 -0800
 Subject: [PATCH] Expose callable API for server
 
 This adds an extern "C" interface within the example server
 ---
- examples/server/CMakeLists.txt |  24 ++++
- examples/server/server.cpp     | 247 +++++++++++++++++++++++++++++++++
- examples/server/server.h       |  83 +++++++++++
+ examples/server/CMakeLists.txt |  24 +++
+ examples/server/server.cpp     | 274 +++++++++++++++++++++++++++++++++
+ examples/server/server.h       |  89 +++++++++++
  ggml-cuda.cu                   |   1 +
- 4 files changed, 355 insertions(+)
+ 4 files changed, 388 insertions(+)
  create mode 100644 examples/server/server.h
 
 diff --git a/examples/server/CMakeLists.txt b/examples/server/CMakeLists.txt
@@ -46,7 +46,7 @@ index 859cd12..4ea47a7 100644
 +endif()
 \ No newline at end of file
 diff --git a/examples/server/server.cpp b/examples/server/server.cpp
-index 895f751..f939590 100644
+index d0cd8e1..5f5d4c5 100644
 --- a/examples/server/server.cpp
 +++ b/examples/server/server.cpp
 @@ -5,6 +5,9 @@
@@ -59,7 +59,7 @@ index 895f751..f939590 100644
  
  #ifndef NDEBUG
  // crash the server in debug mode, otherwise send an http 500 error
-@@ -2631,6 +2634,7 @@ static void append_to_generated_text_from_generated_token_probs(llama_server_con
+@@ -2632,6 +2635,7 @@ static void append_to_generated_text_from_generated_token_probs(llama_server_con
      }
  }
  
@@ -67,84 +67,84 @@ index 895f751..f939590 100644
  int main(int argc, char **argv)
  {
      // own arguments required by this example
-@@ -3065,3 +3069,246 @@ int main(int argc, char **argv)
+@@ -3066,3 +3070,273 @@ int main(int argc, char **argv)
      llama_backend_free();
      return 0;
  }
 +
 +#else // LLAMA_SERVER_LIBRARY
 +// Expose the llama server as a callable extern "C" API
-+llama_server_context llama;
++llama_server_context *llama = NULL;
 +std::atomic<bool> ext_server_running(false);
 +std::thread ext_server_thread;
-+inline ext_server_err makeErr(uint32_t code, std::string msg) {
-+    if (code == 0) {
-+        return ext_server_err{0, NULL};
-+    }
-+    const std::string::size_type size = msg.size();
-+    ext_server_err ret = {
-+        code,
-+        new char[size + 1],
-+    };
-+    memcpy(ret.err, msg.c_str(), size + 1);
-+    return ret;
-+}
 +
-+ext_server_err llama_server_init(ext_server_params *sparams)
++void llama_server_init(ext_server_params *sparams, ext_server_resp_t *err)
 +{
-+    log_set_target(stdout);
-+    gpt_params params;
-+    params.n_ctx = sparams->n_ctx;
-+    params.n_batch = sparams->n_batch;
-+    params.n_threads = sparams->n_threads;
-+    params.n_parallel = sparams->n_parallel;
-+    params.rope_freq_base = sparams->rope_freq_base;
-+    params.rope_freq_scale = sparams->rope_freq_scale;
-+
-+    if (sparams->memory_f16)  {
-+        params.cache_type_k = "f16";
-+        params.cache_type_v = "f16";
-+    } else {
-+        params.cache_type_k = "f32";
-+        params.cache_type_v = "f32";
-+    }
-+    
-+    params.n_gpu_layers = sparams->n_gpu_layers;
-+    params.main_gpu = sparams->main_gpu;
-+    params.use_mlock = sparams->use_mlock;
-+    params.use_mmap = sparams->use_mmap;
-+    params.numa = sparams->numa;
-+    params.embedding = sparams->embedding;
-+    if (sparams->model != NULL) {
-+        params.model = sparams->model;
-+    }
-+
-+    for (ext_server_lora_adapter *la = sparams->lora_adapters; la != NULL; la = la->next) {
-+        params.lora_adapter.push_back(std::make_tuple(la->adapter, la->scale));
-+    }
-+    
++    assert(err != NULL && sparams != NULL);
++    err->id = 0;
++    err->msg[0] = '\0';
 +    try {
++        llama = new llama_server_context;
++        log_set_target(stdout);
++        gpt_params params;
++        params.n_ctx = sparams->n_ctx;
++        params.n_batch = sparams->n_batch;
++        params.n_threads = sparams->n_threads;
++        params.n_parallel = sparams->n_parallel;
++        params.rope_freq_base = sparams->rope_freq_base;
++        params.rope_freq_scale = sparams->rope_freq_scale;
++
++        if (sparams->memory_f16)  {
++            params.cache_type_k = "f16";
++            params.cache_type_v = "f16";
++        } else {
++            params.cache_type_k = "f32";
++            params.cache_type_v = "f32";
++        }
++
++        params.n_gpu_layers = sparams->n_gpu_layers;
++        params.main_gpu = sparams->main_gpu;
++        params.use_mlock = sparams->use_mlock;
++        params.use_mmap = sparams->use_mmap;
++        params.numa = sparams->numa;
++        params.embedding = sparams->embedding;
++        if (sparams->model != NULL) {
++            params.model = sparams->model;
++        }
++
++        for (ext_server_lora_adapter *la = sparams->lora_adapters; la != NULL; la = la->next) {
++            params.lora_adapter.push_back(std::make_tuple(la->adapter, la->scale));
++        }
++
++        if (sparams->mmproj != NULL) {
++            params.mmproj = std::string(sparams->mmproj);
++        }
++           
 +        llama_backend_init(params.numa);
 +
 +        // load the model
-+        if (!llama.load_model(params))
++        if (!llama->load_model(params))
 +        {
 +            // TODO - consider modifying the logging logic or patching load_model so we can capture more detailed error messages
 +            // and pass them back to the caller for better UX
-+            return makeErr(1, "error loading model " + params.model);
++            err->id = -1;
++            snprintf(err->msg, err->msg_len, "error loading model %s", params.model.c_str());
++            return;
 +        }
 +
-+        llama.initialize();
++        llama->initialize();
 +    } catch (std::exception &e) {
-+        return makeErr(1, e.what());
++        err->id = -1;
++        snprintf(err->msg, err->msg_len, "exception %s", e.what());
 +    } catch (...) {
-+        return makeErr(1, "Unknown Exception initializing llama server");
++        err->id = -1;
++        snprintf(err->msg, err->msg_len, "Unknown exception initializing llama server");
 +    }
-+    return makeErr(0, "");
 +}
 +
 +void llama_server_start()
 +{
++    assert(llama != NULL);
 +     // TODO mutex to protect thread creation
 +    ext_server_thread = std::thread([&]()
 +    {
@@ -154,7 +154,7 @@ index 895f751..f939590 100644
 +            ggml_time_init();
 +            while (ext_server_running.load())
 +            {
-+                if (!llama.update_slots()) {
++                if (!llama->update_slots()) {
 +                    LOG_TEE("unexpected error in llama server update_slots - exiting main loop\n");
 +                    break;
 +                }
@@ -170,124 +170,150 @@ index 895f751..f939590 100644
 +}
 +
 +void llama_server_stop() {
++    assert(llama != NULL);
 +    // TODO - too verbose, remove once things are solid
 +    LOG_TEE("requesting llama server shutdown\n");
 +    ext_server_running = false;
 +    ext_server_thread.join();
++    delete llama;
++    llama = NULL;
 +    LOG_TEE("llama server shutdown complete\n");
 +}
 +
-+ext_server_completion_resp llama_server_completion(const char *json_req) {
-+    std::string msg;
-+    ext_server_completion_resp resp = {
-+        0,
-+        NULL,
-+    };
++void llama_server_completion(const char *json_req, ext_server_resp_t *resp) {
++    assert(llama != NULL && json_req != NULL && resp != NULL);
++    resp->id = -1;
++    resp->msg[0] = '\0';
 +    try {
 +        json data = json::parse(json_req);
-+        resp.task_id = llama.request_completion(data, false, false, -1);
-+        return resp;
++        resp->id = llama->request_completion(data, false, false, -1);
 +    } catch (std::exception &e) {
-+        msg = e.what();
++        snprintf(resp->msg, resp->msg_len, "exception %s", e.what());
 +    } catch (...) {
-+        msg = "Unknown Exception during completion";
++        snprintf(resp->msg, resp->msg_len, "Unknown exception during completion");
 +    }
-+    const std::string::size_type size = msg.size();
-+    resp.task_id = 0;
-+    resp.err = new char[size + 1];
-+    memcpy(resp.err, msg.c_str(), size + 1);
-+    return resp;
 +}
 +
-+ext_task_result llama_server_completion_next_result(const int task_id) {
++void llama_server_completion_next_result(const int task_id, ext_server_task_result_t *resp) {
++    assert(llama != NULL && resp != NULL);
 +    std::string msg;
-+    ext_task_result resp = {-1,false,false,NULL};
-+    try {
-+        task_result result = llama.next_result(task_id);
-+        std::string result_json = result.result_json.dump(-1, ' ', false, json::error_handler_t::replace);
-+        const std::string::size_type size = result_json.size();
-+        resp.id = result.id;
-+        resp.stop = result.stop;
-+        resp.error = result.error;
-+        resp.result_json = new char[size + 1];
-+        memcpy(resp.result_json, result_json.c_str(), size + 1);
-+        if (result.error) {
-+            llama.request_cancel(task_id);
-+        } else if (result.stop) {
-+            llama.request_cancel(task_id);
-+        }
-+        return resp;
-+    } catch (std::exception &e) {
-+        msg = e.what(); // TODO - json?
-+    } catch (...) {
-+        msg = "Unknown Exception during completion";
-+    }
-+    resp.error = true;
-+    const std::string::size_type size = msg.size();
-+    resp.result_json = new char[size + 1];
-+    memcpy(resp.result_json, msg.c_str(), size + 1);
-+    return resp;
-+}
-+
-+ext_server_err llama_server_completion_cancel(const int task_id) {
-+    try {
-+        llama.request_cancel(task_id);
-+    } catch (std::exception &e) {
-+        return makeErr(1, e.what());
-+    } catch (...) {
-+        return makeErr(1, "Unknown Exception running llama server");
-+    }
-+    return makeErr(0, "");
-+}
-+
-+
-+ext_server_err llama_server_tokenize(const char *json_req, ext_server_resp *resp) {
++    resp->id = -1;
++    resp->stop = false;
++    resp->error = false;
 +    resp->json_resp = NULL;
++    std::string result_json;
++    try {
++        task_result result = llama->next_result(task_id);
++        result_json = result.result_json.dump(-1, ' ', false, json::error_handler_t::replace);
++        resp->id = result.id;
++        resp->stop = result.stop;
++        resp->error = result.error;
++        if (result.error) {
++            llama->request_cancel(task_id);
++        } else if (result.stop) {
++            llama->request_cancel(task_id);
++        }
++    } catch (std::exception &e) {
++        resp->error = true;
++        resp->id = -1;
++        result_json = "{\"error\":\"exception " + std::string(e.what()) + "\"}";
++    } catch (...) {
++        resp->error = true;
++        resp->id = -1;
++        result_json = "{\"error\":\"Unknown exception during completion\"}";
++    }
++    const std::string::size_type size = result_json.size() + 1;
++    resp->json_resp = new char[size];
++    snprintf(resp->json_resp, size, "%s", result_json.c_str());
++}
++
++void llama_server_release_task_result(ext_server_task_result_t *result) {
++    if (result == NULL || result->json_resp == NULL) {
++        return;
++    }
++    delete[] result->json_resp;
++}
++
++void llama_server_completion_cancel(const int task_id, ext_server_resp_t *err) {
++    assert(llama != NULL && err != NULL);
++    err->id = 0;
++    err->msg[0] = '\0';
++    try {
++        llama->request_cancel(task_id);
++    } catch (std::exception &e) {
++        err->id = -1;
++        snprintf(err->msg, err->msg_len, "exception %s", e.what());
++    } catch (...) {
++        err->id = -1;
++        snprintf(err->msg, err->msg_len, "Unknown exception completion cancel in llama server");
++    }
++}
++
++void llama_server_tokenize(const char *json_req, char **json_resp, ext_server_resp_t *err) {
++    assert(llama != NULL && json_req != NULL && json_resp != NULL && err != NULL);
++    *json_resp = NULL;
++    err->id = 0;
++    err->msg[0] = '\0';
 +    try {
 +        const json body = json::parse(json_req);
 +        std::vector<llama_token> tokens;
 +        if (body.count("content") != 0)
 +        {
-+            tokens = llama.tokenize(body["content"], false);
++            tokens = llama->tokenize(body["content"], false);
 +        }
 +        const json data = format_tokenizer_response(tokens);
 +        std::string result_json = data.dump();
-+        const std::string::size_type size = result_json.size();
-+        resp->json_resp = new char[size + 1];
-+        memcpy(resp->json_resp, result_json.c_str(), size + 1);
++        const std::string::size_type size = result_json.size() + 1;
++        *json_resp = new char[size];
++        snprintf(*json_resp, size, "%s", result_json.c_str());
 +    } catch (std::exception &e) {
-+        return makeErr(1, e.what());
++        err->id = -1;
++        snprintf(err->msg, err->msg_len, "exception %s", e.what());
 +    } catch (...) {
-+        return makeErr(1, "Unknown Exception during tokenize");
++        err->id = -1;
++        snprintf(err->msg, err->msg_len, "Unknown exception during tokenize");
 +    }
-+    return makeErr(0, "");
 +}
 +
-+ext_server_err llama_server_detokenize(const char *json_req, ext_server_resp *resp) {
-+    resp->json_resp = NULL;
++void llama_server_release_json_resp(char **json_resp) {
++    if (json_resp == NULL || *json_resp == NULL) {
++        return;
++    }
++    delete[] *json_resp;
++}
++
++void llama_server_detokenize(const char *json_req, char **json_resp, ext_server_resp_t *err) {
++    assert(llama != NULL && json_req != NULL && json_resp != NULL && err != NULL);
++    *json_resp = NULL;
++    err->id = 0;
++    err->msg[0] = '\0';
 +    try {
 +        const json body = json::parse(json_req);
 +        std::string content;
 +        if (body.count("tokens") != 0)
 +        {
 +            const std::vector<llama_token> tokens = body["tokens"];
-+            content = tokens_to_str(llama.ctx, tokens.cbegin(), tokens.cend());
++            content = tokens_to_str(llama->ctx, tokens.cbegin(), tokens.cend());
 +        }
 +        const json data = format_detokenized_response(content);
 +        std::string result_json = data.dump();
-+        const std::string::size_type size = result_json.size();
-+        resp->json_resp = new char[size + 1];
-+        memcpy(resp->json_resp, result_json.c_str(), size + 1);
++        const std::string::size_type size = result_json.size() + 1;
++        *json_resp = new char[size];
++        snprintf(*json_resp, size, "%s", result_json.c_str());
 +    } catch (std::exception &e) {
-+        return makeErr(1, e.what());
++        err->id = -1;
++        snprintf(err->msg, err->msg_len, "exception %s", e.what());
 +    } catch (...) {
-+        return makeErr(1, "Unknown Exception during detokenize");
++        err->id = -1;
++        snprintf(err->msg, err->msg_len, "Unknown exception during detokenize");
 +    }
-+    return makeErr(0, "");
 +}
 +
-+ext_server_err llama_server_embedding(const char *json_req, ext_server_resp *resp) {
-+    resp->json_resp = NULL;
++void llama_server_embedding(const char *json_req, char** json_resp, ext_server_resp_t *err) {
++    assert(llama != NULL && json_req != NULL && json_resp != NULL && err != NULL);
++    *json_resp = NULL;
++    err->id = 0;
++    err->msg[0] = '\0';
 +    try {
 +        const json body = json::parse(json_req);
 +        json prompt;
@@ -299,28 +325,29 @@ index 895f751..f939590 100644
 +        {
 +            prompt = "";
 +        }
-+        const int task_id = llama.request_completion({ {"prompt", prompt}, { "n_predict", 0} }, false, true, -1);
-+        task_result result = llama.next_result(task_id);
++        const int task_id = llama->request_completion({ {"prompt", prompt}, { "n_predict", 0} }, false, true, -1);
++        task_result result = llama->next_result(task_id);
 +        std::string result_json = result.result_json.dump();
-+        const std::string::size_type size = result_json.size();
-+        resp->json_resp = new char[size + 1];
-+        memcpy(resp->json_resp, result_json.c_str(), size + 1);
++        const std::string::size_type size = result_json.size() + 1;
++        *json_resp = new char[size];
++        snprintf(*json_resp, size, "%s", result_json.c_str());
 +    } catch (std::exception &e) {
-+        return makeErr(1, e.what());
++        err->id = -1;
++        snprintf(err->msg, err->msg_len, "exception %s", e.what());
 +    } catch (...) {
-+        return makeErr(1, "Unknown Exception during detokenize");
++        err->id = -1;
++        snprintf(err->msg, err->msg_len, "Unknown exception during embedding");
 +    }
-+    return makeErr(0, "");
 +}
 +
 +#endif // LLAMA_SERVER_LIBRARY
 \ No newline at end of file
 diff --git a/examples/server/server.h b/examples/server/server.h
 new file mode 100644
-index 0000000..4d03b1e
+index 0000000..d22f1b6
 --- /dev/null
 +++ b/examples/server/server.h
-@@ -0,0 +1,83 @@
+@@ -0,0 +1,89 @@
 +#if defined(LLAMA_SERVER_LIBRARY)
 +#ifndef LLAMA_SERVER_H
 +#define LLAMA_SERVER_H
@@ -336,20 +363,23 @@ index 0000000..4d03b1e
 +extern "C"
 +{
 +#endif
-+    // TODO - clean the type def's up a bit for better consistency
-+    typedef struct ext_server_err {
-+        uint32_t code; // 0 on success, > 0 on error
-+        char *err; // null if code == 0; else contains error message.  Caller responsible for freeing memory
-+    } ext_server_err;
++    typedef struct ext_server_resp {
++        int id; // < 0 on error
++        size_t msg_len; // caller must allocate msg and set msg_len
++        char *msg;
++    } ext_server_resp_t;
 +
++    // Allocated and freed by caller
 +    typedef struct ext_server_lora_adapter {
 +        char *adapter;
 +        float scale;
 +        struct ext_server_lora_adapter *next;
-+    } ext_server_lora_adapter;
++    } ext_server_lora_adapter_t;
++
++    // Allocated and freed by caller
 +    typedef struct ext_server_params
 +    {
-+        char *model;
++        char *model;            
 +        uint32_t n_ctx;         // text context, 0 = from model
 +        uint32_t n_batch;       // prompt processing maximum batch size
 +        uint32_t n_threads;     // number of threads to use for generation
@@ -363,40 +393,43 @@ index 0000000..4d03b1e
 +        bool use_mmap;          // use mmap if possible
 +        bool numa;              // attempt optimizations that help on some NUMA systems
 +        bool embedding;         // get only sentence embedding
-+        ext_server_lora_adapter* lora_adapters;
-+    } ext_server_params;
++        ext_server_lora_adapter_t* lora_adapters;
++        char *mmproj;
++    } ext_server_params_t;
 +
-+    // Initialize the server once per process
-+    ext_server_err llama_server_init(ext_server_params *sparams);
-+
-+    // Run the main loop
-+    void llama_server_start();
-+    // Stop the main loop
-+    void llama_server_stop();
-+
-+    typedef struct ext_task_result
++    typedef struct ext_server_task_result
 +    {
 +        int id;
 +        bool stop;
 +        bool error;
-+        char* result_json; // caller responsible to free this memory
-+    } ext_task_result;
-+    
-+    typedef struct ext_server_completion_resp {
-+        int task_id; // < 0 on error, >= 0 on success
-+        char *err; // null if task_id >= 0; else contains error message.  Caller responsible for freeing memory
-+    } ext_server_completion_resp;
-+    ext_server_completion_resp llama_server_completion(const char *json_req);
-+    ext_task_result llama_server_completion_next_result(const int task_id);
-+    ext_server_err llama_server_completion_cancel(const int task_id);
++        char* json_resp; // null terminated, memory managed by ext_server
++    } ext_server_task_result_t;
 +
-+    // Caller responsible for freeing json_resp
-+    typedef struct ext_server_resp {
-+        char *json_resp; // Caller responsible for freeing string
-+    } ext_server_resp;
-+    ext_server_err llama_server_tokenize(const char *json_req, ext_server_resp *resp);
-+    ext_server_err llama_server_detokenize(const char *json_req, ext_server_resp *resp);
-+    ext_server_err llama_server_embedding(const char *json_req, ext_server_resp *resp);
++    // Initialize the server once per process
++    // err->id = 0 for success and err->msg[0] = NULL
++    // err->id != 0 for failure, and err->msg contains error message
++    void llama_server_init(ext_server_params_t *sparams, ext_server_resp_t *err);
++
++    // Run the main loop, called once per init
++    void llama_server_start();
++    // Stop the main loop and free up resources allocated in init and start.  Init must be called again to reuse
++    void llama_server_stop();
++
++    // json_req null terminated string, memory managed by caller
++    // resp->id >= 0 on success (task ID)
++    // resp->id < 0 on error, and resp->msg contains error message
++    void llama_server_completion(const char *json_req, ext_server_resp_t *resp);
++
++    // Caller must call llama_server_release_task_result to free resp->json_resp
++    void llama_server_completion_next_result(const int task_id, ext_server_task_result_t *result);
++    void llama_server_completion_cancel(const int task_id, ext_server_resp_t *err);
++    void llama_server_release_task_result(ext_server_task_result_t *result);
++
++    // Caller must call llama_server_releaes_json_resp to free json_resp if err.id < 0
++    void llama_server_tokenize(const char *json_req, char **json_resp, ext_server_resp_t *err);
++    void llama_server_detokenize(const char *json_req, char **json_resp, ext_server_resp_t *err);
++    void llama_server_embedding(const char *json_req, char** json_resp, ext_server_resp_t *err);
++    void llama_server_release_json_resp(char **json_resp);
 +
 +#ifdef __cplusplus
 +}
@@ -406,10 +439,10 @@ index 0000000..4d03b1e
 +#endif // LLAMA_SERVER_LIBRARY
 \ No newline at end of file
 diff --git a/ggml-cuda.cu b/ggml-cuda.cu
-index 85f7a29..ce51364 100644
+index 9e1acd3..ea64b55 100644
 --- a/ggml-cuda.cu
 +++ b/ggml-cuda.cu
-@@ -6410,6 +6410,7 @@ static cudaError_t ggml_cuda_cpy_tensor_2d(
+@@ -6505,6 +6505,7 @@ static cudaError_t ggml_cuda_cpy_tensor_2d(
          CUDA_CHECK(cudaGetDevice(&id));
          src_ptr = (char *) extra->data_device[id];
      } else {

llm/llama.go

@@ -3,6 +3,7 @@ package llm
 import (
 	"bytes"
 	"context"
+	_ "embed"
 	"errors"
 	"fmt"
 	"os"
@@ -112,12 +113,6 @@ type ImageData struct {
 	ID   int    `json:"id"`
 }
 
-type llama struct {
-	api.Options
-	ImageData []ImageData
-	Running
-}
-
 var (
 	errNvidiaSMI     = errors.New("warning: gpu support may not be enabled, check that you have installed GPU drivers: nvidia-smi command failed")
 	errAvailableVRAM = errors.New("not enough VRAM available, falling back to CPU only")
@@ -166,7 +161,8 @@ type prediction struct {
 }
 
 const maxBufferSize = 512 * format.KiloByte
-const maxRetries = 6
+const maxRetries = 3
+const retryDelay = 1 * time.Second
 
 type PredictOpts struct {
 	Prompt string

llm/llm.go

@@ -11,6 +11,7 @@ import (
 
 	"github.com/jmorganca/ollama/api"
 	"github.com/jmorganca/ollama/format"
+	"github.com/jmorganca/ollama/gpu"
 )
 
 type LLM interface {
@@ -19,7 +20,6 @@ type LLM interface {
 	Encode(context.Context, string) ([]int, error)
 	Decode(context.Context, []int) (string, error)
 	Close()
-	Ping(context.Context) error
 }
 
 func New(workDir, model string, adapters, projectors []string, opts api.Options) (LLM, error) {
@@ -78,5 +78,17 @@ func New(workDir, model string, adapters, projectors []string, opts api.Options)
 	opts.NumGQA = 0
 	opts.RopeFrequencyBase = 0.0
 	opts.RopeFrequencyScale = 0.0
-	return newLlamaExtServer(model, adapters, projectors, ggml.NumLayers(), opts)
+	gpuInfo := gpu.GetGPUInfo()
+	switch gpuInfo.Driver {
+	case "ROCM":
+		return newRocmShimExtServer(model, adapters, projectors, ggml.NumLayers(), opts)
+	default:
+		// Rely on the built-in CUDA based server which will fall back to CPU
+		return newLlamaExtServer(model, adapters, projectors, ggml.NumLayers(), opts)
+	}
+}
+
+// Give any native cgo implementations an opportunity to initialize
+func Init(workdir string) error {
+	return nativeInit(workdir)
 }

llm/rocm_shim.c

@@ -0,0 +1,134 @@
+#include "rocm_shim.h"
+
+#include <stdio.h>
+#include <string.h>
+
+#ifndef _WIN32
+#include <dlfcn.h>
+#define LOAD_LIBRARY(lib, flags) dlopen(lib, flags)
+#define LOAD_SYMBOL(handle, sym) dlsym(handle, sym)
+#define LOAD_ERR() dlerror()
+#define UNLOAD_LIBRARY(handle) dlclose(handle)
+#else
+#include <windows.h>
+#define LOAD_LIBRARY(lib, flags) LoadLibrary(lib)
+#define LOAD_SYMBOL(handle, sym) GetProcAddress(handle, sym)
+#define UNLOAD_LIBRARY(handle) FreeLibrary(handle)
+// TODO - refactor this with proper error message handling on windows
+inline static char *LOAD_ERR() {
+  static char errbuf[8];
+  snprintf(errbuf, 8, "0x%lx", GetLastError());
+  return errbuf;
+}
+#endif
+
+void rocm_shim_init(const char *libPath, struct rocm_llama_server *s,
+                    ext_server_resp_t *err) {
+  int i = 0;
+  struct lookup {
+    char *s;
+    void **p;
+  } l[] = {
+      {"llama_server_init", (void *)&s->llama_server_init},
+      {"llama_server_start", (void *)&s->llama_server_start},
+      {"llama_server_stop", (void *)&s->llama_server_stop},
+      {"llama_server_completion", (void *)&s->llama_server_completion},
+      {"llama_server_completion_next_result",
+       (void *)&s->llama_server_completion_next_result},
+      {"llama_server_completion_cancel",
+       (void *)&s->llama_server_completion_cancel},
+      {"llama_server_release_task_result",
+       (void *)&s->llama_server_release_task_result},
+      {"llama_server_tokenize", (void *)&s->llama_server_tokenize},
+      {"llama_server_detokenize", (void *)&s->llama_server_detokenize},
+      {"llama_server_embedding", (void *)&s->llama_server_embedding},
+      {"llama_server_release_json_resp",
+       (void *)&s->llama_server_release_json_resp},
+      {"", NULL},
+  };
+
+  printf("Lazy loading %s library\n", libPath);
+  s->handle = LOAD_LIBRARY(libPath, RTLD_LAZY);
+  if (!s->handle) {
+    err->id = -1;
+    snprintf(
+        err->msg, err->msg_len,
+        "Unable to load rocm server library: %s (If you have a Radeon card, "
+        "did you install the ROCM libraries?)",
+        LOAD_ERR());
+    return;
+  }
+
+  for (i = 0; l[i].p != NULL; i++) {
+    *l[i].p = LOAD_SYMBOL(s->handle, l[i].s);
+    if (!l[i].p) {
+      UNLOAD_LIBRARY(s->handle);
+      err->id = -1;
+      snprintf(err->msg, err->msg_len, "symbol lookup for %s failed: %s",
+               l[i].s, LOAD_ERR());
+      return;
+    }
+  }
+}
+
+inline void rocm_shim_llama_server_init(struct rocm_llama_server s,
+                                        ext_server_params_t *sparams,
+                                        ext_server_resp_t *err) {
+  s.llama_server_init(sparams, err);
+}
+
+inline void rocm_shim_llama_server_start(struct rocm_llama_server s) {
+  s.llama_server_start();
+}
+
+inline void rocm_shim_llama_server_stop(struct rocm_llama_server s) {
+  s.llama_server_stop();
+}
+
+inline void rocm_shim_llama_server_completion(struct rocm_llama_server s,
+                                              const char *json_req,
+                                              ext_server_resp_t *resp) {
+  s.llama_server_completion(json_req, resp);
+}
+
+inline void rocm_shim_llama_server_completion_next_result(
+    struct rocm_llama_server s, const int task_id,
+    ext_server_task_result_t *result) {
+  s.llama_server_completion_next_result(task_id, result);
+}
+
+inline void rocm_shim_llama_server_completion_cancel(struct rocm_llama_server s,
+                                                     const int task_id,
+                                                     ext_server_resp_t *err) {
+  s.llama_server_completion_cancel(task_id, err);
+}
+inline void rocm_shim_llama_server_release_task_result(
+    struct rocm_llama_server s, ext_server_task_result_t *result) {
+  s.llama_server_release_task_result(result);
+}
+
+inline void rocm_shim_llama_server_tokenize(struct rocm_llama_server s,
+                                            const char *json_req,
+                                            char **json_resp,
+                                            ext_server_resp_t *err) {
+  s.llama_server_tokenize(json_req, json_resp, err);
+}
+
+inline void rocm_shim_llama_server_detokenize(struct rocm_llama_server s,
+                                              const char *json_req,
+                                              char **json_resp,
+                                              ext_server_resp_t *err) {
+  s.llama_server_detokenize(json_req, json_resp, err);
+}
+
+inline void rocm_shim_llama_server_embedding(struct rocm_llama_server s,
+                                             const char *json_req,
+                                             char **json_resp,
+                                             ext_server_resp_t *err) {
+  s.llama_server_embedding(json_req, json_resp, err);
+}
+
+inline void rocm_shim_llama_server_release_json_resp(struct rocm_llama_server s,
+                                                     char **json_resp) {
+  s.llama_server_release_json_resp(json_resp);
+}

llm/rocm_shim.h

@@ -0,0 +1,73 @@
+#include <stdlib.h>
+
+#include "server.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+struct rocm_llama_server {
+  void *handle;
+  void (*llama_server_init)(ext_server_params_t *sparams,
+                            ext_server_resp_t *err);
+  void (*llama_server_start)();
+  void (*llama_server_stop)();
+  void (*llama_server_completion)(const char *json_req,
+                                  ext_server_resp_t *resp);
+  void (*llama_server_completion_next_result)(const int task_id,
+                                              ext_server_task_result_t *result);
+  void (*llama_server_completion_cancel)(const int task_id,
+                                         ext_server_resp_t *err);
+  void (*llama_server_release_task_result)(ext_server_task_result_t *result);
+  void (*llama_server_tokenize)(const char *json_req, char **json_resp,
+                                ext_server_resp_t *err);
+  void (*llama_server_detokenize)(const char *json_req, char **json_resp,
+                                  ext_server_resp_t *err);
+  void (*llama_server_embedding)(const char *json_req, char **json_resp,
+                                 ext_server_resp_t *err);
+  void (*llama_server_release_json_resp)(char **json_resp);
+};
+
+void rocm_shim_init(const char *libPath, struct rocm_llama_server *s,
+                    ext_server_resp_t *err);
+
+// No good way to call C function pointers from Go so inline the indirection
+void rocm_shim_llama_server_init(struct rocm_llama_server s,
+                                 ext_server_params_t *sparams,
+                                 ext_server_resp_t *err);
+
+void rocm_shim_llama_server_start(struct rocm_llama_server s);
+
+void rocm_shim_llama_server_stop(struct rocm_llama_server s);
+
+void rocm_shim_llama_server_completion(struct rocm_llama_server s,
+                                       const char *json_req,
+                                       ext_server_resp_t *resp);
+
+void rocm_shim_llama_server_completion_next_result(
+    struct rocm_llama_server s, const int task_id,
+    ext_server_task_result_t *result);
+
+void rocm_shim_llama_server_completion_cancel(struct rocm_llama_server s,
+                                              const int task_id,
+                                              ext_server_resp_t *err);
+
+void rocm_shim_llama_server_release_task_result(
+    struct rocm_llama_server s, ext_server_task_result_t *result);
+
+void rocm_shim_llama_server_tokenize(struct rocm_llama_server s,
+                                     const char *json_req, char **json_resp,
+                                     ext_server_resp_t *err);
+
+void rocm_shim_llama_server_detokenize(struct rocm_llama_server s,
+                                       const char *json_req, char **json_resp,
+                                       ext_server_resp_t *err);
+
+void rocm_shim_llama_server_embedding(struct rocm_llama_server s,
+                                      const char *json_req, char **json_resp,
+                                      ext_server_resp_t *err);
+void rocm_shim_llama_server_release_json_resp(struct rocm_llama_server s,
+                                              char **json_resp);
+
+#ifdef __cplusplus
+}
+#endif

llm/shim_darwin.go

@@ -0,0 +1,18 @@
+package llm
+
+import (
+	"fmt"
+
+	"github.com/jmorganca/ollama/api"
+)
+
+// no-op stubs for mac
+
+func newRocmShimExtServer(model string, adapters, projectors []string, numLayers int64, opts api.Options) (extServer, error) {
+	// should never happen...
+	return nil, fmt.Errorf("ROCM GPUs not supported on Mac")
+}
+
+func nativeInit(workDir string) error {
+	return nil
+}

llm/shim_ext_server.go

@@ -0,0 +1,212 @@
+//go:build !darwin
+
+package llm
+
+/*
+
+#include <stdlib.h>
+#include "rocm_shim.h"
+
+*/
+import "C"
+import (
+	"context"
+	"embed"
+	"errors"
+	"fmt"
+	"io"
+	"io/fs"
+	"log"
+	"os"
+	"path/filepath"
+	"runtime"
+	"sync"
+	"unsafe"
+
+	"github.com/jmorganca/ollama/api"
+)
+
+//go:embed llama.cpp/gguf/build/*/lib/*
+var libEmbed embed.FS
+
+var RocmShimMissing = fmt.Errorf("ROCm shim library not included in this build of ollama. Radeon GPUs are not supported")
+var NoShim = true
+
+type shimExtServer struct {
+	s       C.struct_rocm_llama_server
+	options api.Options
+}
+
+// Note: current implementation does not support concurrent instantiations
+var shimMutex sync.Mutex
+var llm *shimExtServer
+
+func (llm *shimExtServer) llama_server_init(sparams *C.ext_server_params_t, err *C.ext_server_resp_t) {
+	C.rocm_shim_llama_server_init(llm.s, sparams, err)
+}
+func (llm *shimExtServer) llama_server_start() {
+	C.rocm_shim_llama_server_start(llm.s)
+}
+func (llm *shimExtServer) llama_server_stop() {
+	C.rocm_shim_llama_server_stop(llm.s)
+}
+
+func (llm *shimExtServer) llama_server_completion(json_req *C.char, resp *C.ext_server_resp_t) {
+	C.rocm_shim_llama_server_completion(llm.s, json_req, resp)
+}
+func (llm *shimExtServer) llama_server_completion_next_result(task_id C.int, resp *C.ext_server_task_result_t) {
+	C.rocm_shim_llama_server_completion_next_result(llm.s, task_id, resp)
+}
+func (llm *shimExtServer) llama_server_completion_cancel(task_id C.int, err *C.ext_server_resp_t) {
+	C.rocm_shim_llama_server_completion_cancel(llm.s, task_id, err)
+}
+func (llm *shimExtServer) llama_server_release_task_result(result *C.ext_server_task_result_t) {
+	C.rocm_shim_llama_server_release_task_result(llm.s, result)
+}
+
+func (llm *shimExtServer) llama_server_tokenize(json_req *C.char, json_resp **C.char, err *C.ext_server_resp_t) {
+	C.rocm_shim_llama_server_tokenize(llm.s, json_req, json_resp, err)
+}
+func (llm *shimExtServer) llama_server_detokenize(json_req *C.char, json_resp **C.char, err *C.ext_server_resp_t) {
+	C.rocm_shim_llama_server_detokenize(llm.s, json_req, json_resp, err)
+}
+func (llm *shimExtServer) llama_server_embedding(json_req *C.char, json_resp **C.char, err *C.ext_server_resp_t) {
+	C.rocm_shim_llama_server_embedding(llm.s, json_req, json_resp, err)
+}
+func (llm *shimExtServer) llama_server_release_json_resp(json_resp **C.char) {
+	C.rocm_shim_llama_server_release_json_resp(llm.s, json_resp)
+}
+
+func newRocmShimExtServer(model string, adapters, projectors []string, numLayers int64, opts api.Options) (extServer, error) {
+	if NoShim {
+		return nil, RocmShimMissing
+	}
+	log.Printf("Loading ROCM llm server")
+	if llm == nil {
+		return nil, fmt.Errorf("nativeInit wasnt called or libary load failed")
+	}
+	llm.options = opts
+	return newExtServer(llm, model, adapters, projectors, numLayers, opts)
+}
+
+func (llm *shimExtServer) Predict(ctx context.Context, pred PredictOpts, fn func(PredictResult)) error {
+	return predict(llm, llm.options, ctx, pred, fn)
+}
+
+func (llm *shimExtServer) Encode(ctx context.Context, prompt string) ([]int, error) {
+	return encode(llm, ctx, prompt)
+}
+
+func (llm *shimExtServer) Decode(ctx context.Context, tokens []int) (string, error) {
+	return decode(llm, ctx, tokens)
+}
+
+func (llm *shimExtServer) Embedding(ctx context.Context, input string) ([]float64, error) {
+	return embedding(llm, ctx, input)
+}
+
+func (llm *shimExtServer) Close() {
+	close(llm)
+}
+
+func nativeInit(workdir string) error {
+	err := extractLib(workdir)
+	if err != nil {
+		if err == RocmShimMissing {
+			log.Printf("%s", err)
+			return nil
+		}
+		return err
+	}
+
+	// Verify we have permissions - either running as root, or we have group access to the driver
+	fd, err := os.OpenFile("/dev/kfd", os.O_RDWR, 0666)
+	if err != nil {
+		if errors.Is(err, fs.ErrPermission) {
+			log.Fatalf("Radeon card detected, but permissions not set up properly.  Either run ollama as root, or add you user account to the render group.")
+			return err
+		} else if errors.Is(err, fs.ErrNotExist) {
+			// expected behavior without a radeon card
+			return nil
+		}
+
+		return fmt.Errorf("failed to check permission on /dev/kfd: %w", err)
+	}
+	fd.Close()
+
+	shimMutex.Lock()
+	defer shimMutex.Unlock()
+	if llm != nil {
+		return nil
+	}
+	var libName string
+	switch runtime.GOOS {
+	case "darwin":
+		// shouldn't happen
+		return nil
+	case "linux":
+		libName = "librocm_server.so"
+	case "windows":
+		libName = "rocm_server.dll"
+	default:
+		// shouldn't happen
+		return nil
+	}
+	libPath := C.CString(filepath.Join(workdir, libName))
+	defer C.free(unsafe.Pointer(libPath))
+	resp := newExtServerResp(128)
+	defer freeExtServerResp(resp)
+	var srv C.struct_rocm_llama_server
+	C.rocm_shim_init(libPath, &srv, &resp)
+	if resp.id < 0 {
+		// TODO - consider softening this failure mode to allow fall-back to the CUDA based built-in llm
+		//        and run against CPU
+		return fmt.Errorf("Unable to load AMD GPU library: %s", C.GoString(resp.msg))
+	}
+	llm = &shimExtServer{
+		s:       srv,
+		options: api.DefaultOptions(),
+	}
+	return nil
+}
+
+func extractLib(workDir string) error {
+	files, err := fs.Glob(libEmbed, "llama.cpp/gguf/build/*/lib/*rocm_server*")
+	if err != nil || len(files) == 0 {
+		// this is expected, ollama may be compiled without shim library packed in
+		return RocmShimMissing
+	}
+
+	if len(files) != 1 {
+		// Shouldn't happen, but just use the first one we find
+		log.Printf("WARNING: multiple rocm libraries detected - using %s", files[0])
+	}
+
+	srcFile, err := libEmbed.Open(files[0])
+	if err != nil {
+		return fmt.Errorf("read ROCm shim %s: %v", files[0], err)
+	}
+	defer srcFile.Close()
+	if err := os.MkdirAll(workDir, 0o755); err != nil {
+		return fmt.Errorf("create ROCm shim temp dir %s: %v", workDir, err)
+	}
+
+	destFile := filepath.Join(workDir, filepath.Base(files[0]))
+
+	_, err = os.Stat(destFile)
+	switch {
+	case errors.Is(err, os.ErrNotExist):
+		destFile, err := os.OpenFile(destFile, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o755)
+		if err != nil {
+			return fmt.Errorf("write ROCm shim %s: %v", files[0], err)
+		}
+		defer destFile.Close()
+		if _, err := io.Copy(destFile, srcFile); err != nil {
+			return fmt.Errorf("copy ROCm shim %s: %v", files[0], err)
+		}
+	case err != nil:
+		return fmt.Errorf("stat ROCm shim %s: %v", files[0], err)
+	}
+	NoShim = false
+	return nil
+}

scripts/build_linux.sh

@@ -8,7 +8,7 @@ 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 --platform=linux/$TARGETARCH --build-arg=VERSION --build-arg=GOFLAGS -f Dockerfile.build -t builder:$TARGETARCH .
+    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

scripts/build_remote.py

@@ -0,0 +1,68 @@
+#!/usr/bin/env python3
+import subprocess
+import sys
+from urllib.parse import urlparse
+from git import Repo
+
+# Helper script to be able to build on remote repos using git to push local changes
+# (e.g. particularly helpful to target a remote windows build system)
+#
+# Typical windows remote git config looks like this:
+#
+#[remote "windows-pa"]
+#        url = jdoe@desktop-foo:C:/Users/Jdoe/code/ollama
+#        fetch = +refs/heads/*:refs/remotes/windows-pa/*
+#        uploadpack = powershell git upload-pack
+#        receivepack = powershell git receive-pack
+#
+
+# TODO - add argpare and make this more configurable 
+# - force flag becomes optional
+# - generate, build or test ...
+
+# Note: remote repo will need this run once:
+# git config --local receive.denyCurrentBranch updateInstead
+repo = Repo(".")
+
+# On linux, add links in /usr/local/bin to the go binaries to avoid needing this
+# GoCmd = "/usr/local/go/bin/go" 
+GoCmd = "go" 
+
+if repo.is_dirty():
+    print("Tree is dirty.  Commit your changes before running this script")
+    sys.exit(1)
+
+if len(sys.argv) != 2:
+    print("Please specify the remote name: " + ', '.join([r.name for r in repo.remotes]))
+    sys.exit(1)
+remote_name = sys.argv[1]
+
+remote = {r.name: r for r in repo.remotes}[remote_name]
+raw_url = list(remote.urls)[0]
+url = urlparse(raw_url)
+# Windows urls don't quite parse properly
+if url.scheme == "" and url.netloc == "":
+    url = urlparse("ssh://" + raw_url)
+print("URL: " + str(url))
+netloc = url.netloc.split(":")[0]
+path = url.path
+branch_name = repo.active_branch.name
+
+print("Force pushing content to remote...")
+# Use with care given the force push
+remote.push(force=True).raise_if_error()
+
+print("Ensuring correct branch checked out on remote via ssh...")
+subprocess.check_call(['ssh', netloc, 'cd', path, ';', 'git', 'checkout', branch_name])
+
+
+# TODO - add some hardening to try to figure out how to set up the path properly
+# subprocess.check_call(['ssh', netloc, 'cd', path, ';', 'env'])
+# TODO - or consider paramiko maybe
+
+print("Performing generate")
+subprocess.check_call(['ssh', netloc, 'cd', path, ';', GoCmd, 'generate', './...'])
+
+print("Building")
+subprocess.check_call(['ssh', netloc, 'cd', path, ';', GoCmd, 'build', '.'])
+

scripts/setup_integration_tests.sh

@@ -32,4 +32,4 @@ for LAYER in $(cat ${OLLAMA_MODELS}/manifests/${REGISTRY}/${TEST_MODEL}/${TEST_M
     curl -L -C - --header "${ACCEPT_HEADER}" \
         -o ${OLLAMA_MODELS}/blobs/${LAYER} \
         ${REGISTRY_SCHEME}://${REGISTRY}/v2/${TEST_MODEL}/blobs/${LAYER}
-done
+done

server/llm_test.go

@@ -2,14 +2,17 @@ package server
 
 import (
 	"context"
+	"os"
 	"strings"
 	"sync"
 	"testing"
 	"time"
 
 	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
 
 	"github.com/jmorganca/ollama/api"
+	"github.com/jmorganca/ollama/llm"
 )
 
 // TODO - this would ideally be in the llm package, but that would require some refactoring of interfaces in the server
@@ -33,12 +36,16 @@ var (
 	}
 	resp = [2]string{
 		"once upon a time",
-		"fourth thursday",
+		"united states thanksgiving",
 	}
 )
 
 func TestIntegrationSimpleOrcaMini(t *testing.T) {
 	SkipIFNoTestData(t)
+	workDir, err := os.MkdirTemp("", "ollama")
+	require.NoError(t, err)
+	defer os.RemoveAll(workDir)
+	require.NoError(t, llm.Init(workDir))
 	ctx, cancel := context.WithTimeout(context.Background(), time.Second*60)
 	defer cancel()
 	opts := api.DefaultOptions()
@@ -56,7 +63,13 @@ func TestIntegrationSimpleOrcaMini(t *testing.T) {
 // get true concurrency working with n_parallel support in the backend
 func TestIntegrationConcurrentPredictOrcaMini(t *testing.T) {
 	SkipIFNoTestData(t)
+
 	t.Skip("concurrent prediction on single runner not currently supported")
+
+	workDir, err := os.MkdirTemp("", "ollama")
+	require.NoError(t, err)
+	defer os.RemoveAll(workDir)
+	require.NoError(t, llm.Init(workDir))
 	ctx, cancel := context.WithTimeout(context.Background(), time.Second*60)
 	defer cancel()
 	opts := api.DefaultOptions()
@@ -79,6 +92,10 @@ func TestIntegrationConcurrentPredictOrcaMini(t *testing.T) {
 
 func TestIntegrationConcurrentRunnersOrcaMini(t *testing.T) {
 	SkipIFNoTestData(t)
+	workDir, err := os.MkdirTemp("", "ollama")
+	require.NoError(t, err)
+	defer os.RemoveAll(workDir)
+	require.NoError(t, llm.Init(workDir))
 	ctx, cancel := context.WithTimeout(context.Background(), time.Second*60)
 	defer cancel()
 	opts := api.DefaultOptions()
@@ -87,6 +104,7 @@ func TestIntegrationConcurrentRunnersOrcaMini(t *testing.T) {
 	var wg sync.WaitGroup
 	wg.Add(len(req))
 
+	t.Logf("Running %d concurrently", len(req))
 	for i := 0; i < len(req); i++ {
 		go func(i int) {
 			defer wg.Done()

server/routes.go

@@ -25,6 +25,7 @@ import (
 	"github.com/gin-gonic/gin"
 
 	"github.com/jmorganca/ollama/api"
+	"github.com/jmorganca/ollama/gpu"
 	"github.com/jmorganca/ollama/llm"
 	"github.com/jmorganca/ollama/parser"
 	"github.com/jmorganca/ollama/version"
@@ -81,20 +82,6 @@ func load(c *gin.Context, modelName string, reqOpts map[string]interface{}, sess
 		return nil, err
 	}
 
-	ctx := c.Request.Context()
-
-	// check if the loaded model is still running in a subprocess, in case something unexpected happened
-	if loaded.runner != nil {
-		if err := loaded.runner.Ping(ctx); err != nil {
-			log.Print("loaded llm process not responding, closing now")
-			// the subprocess is no longer running, so close it
-			loaded.runner.Close()
-			loaded.runner = nil
-			loaded.Model = nil
-			loaded.Options = nil
-		}
-	}
-
 	needLoad := loaded.runner == nil || // is there a model loaded?
 		loaded.ModelPath != model.ModelPath || // has the base model changed?
 		!reflect.DeepEqual(loaded.AdapterPaths, model.AdapterPaths) || // have the adapters changed?
@@ -905,9 +892,12 @@ func Serve(ln net.Listener) error {
 		os.Exit(0)
 	}()
 
-	if runtime.GOOS == "linux" {
+	if err := llm.Init(s.WorkDir); err != nil {
+		return fmt.Errorf("unable to initialize llm library %w", err)
+	}
+	if runtime.GOOS == "linux" { // TODO - windows too
 		// check compatibility to log warnings
-		if _, err := llm.CheckVRAM(); err != nil {
+		if _, err := gpu.CheckVRAM(); err != nil {
 			log.Print(err.Error())
 		}
 	}