Skip to content

GitLab

Unverified Commit bc8909fb authored by Daniel Hiltgen's avatar Daniel Hiltgen Committed by GitHub
Browse files

Use runners for GPU discovery (#12090) 

This revamps how we discover GPUs in the system by leveraging the Ollama
runner.  This should eliminate inconsistency between our GPU discovery and the
runners capabilities at runtime, particularly for cases where we try to filter
out unsupported GPUs.  Now the runner does that implicitly based on the actual
device list.  In some cases free VRAM reporting can be unreliable which can
leaad to scheduling mistakes, so this also includes a patch to leverage more
reliable VRAM reporting libraries if available.

Automatic workarounds have been removed as only one GPU leveraged this, which
is now documented. This GPU will soon fall off the support matrix with the next
ROCm bump.

Additional cleanup of the scheduler and discovery packages can be done in the
future once we have switched on the new memory management code, and removed
support for the llama runner.
parent 6b50f2b9
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 155 additions and 2759 deletions
CMakeLists.txt
View file @ bc8909fb
......@@ -117,7 +117,6 @@ if(CMAKE_HIP_COMPILER)
target_compile_definitions(ggml-hip PRIVATE GGML_HIP_NO_VMM)
set(OLLAMA_HIP_INSTALL_DIR ${OLLAMA_INSTALL_DIR}/rocm)
install(TARGETS ggml-hip
RUNTIME_DEPENDENCY_SET rocm
RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP
......@@ -128,13 +127,13 @@ if(CMAKE_HIP_COMPILER)
PRE_INCLUDE_REGEXES hipblas rocblas amdhip64 rocsolver amd_comgr hsa-runtime64 rocsparse tinfo rocprofiler-register drm drm_amdgpu numa elf
PRE_EXCLUDE_REGEXES ".*"
POST_EXCLUDE_REGEXES "system32"
RUNTIME DESTINATION ${OLLAMA_HIP_INSTALL_DIR} COMPONENT HIP
LIBRARY DESTINATION ${OLLAMA_HIP_INSTALL_DIR} COMPONENT HIP
RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP
LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP
)
foreach(HIP_LIB_BIN_INSTALL_DIR IN ITEMS ${HIP_BIN_INSTALL_DIR} ${HIP_LIB_INSTALL_DIR})
if(EXISTS ${HIP_LIB_BIN_INSTALL_DIR}/rocblas)
install(DIRECTORY ${HIP_LIB_BIN_INSTALL_DIR}/rocblas DESTINATION ${OLLAMA_HIP_INSTALL_DIR} COMPONENT HIP)
install(DIRECTORY ${HIP_LIB_BIN_INSTALL_DIR}/rocblas DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP)
break()
endif()
endforeach()
......
Dockerfile
View file @ bc8909fb
......@@ -77,7 +77,7 @@ FROM base AS rocm-6
ENV PATH=/opt/rocm/hcc/bin:/opt/rocm/hip/bin:/opt/rocm/bin:/opt/rocm/hcc/bin:$PATH
ARG PARALLEL
RUN --mount=type=cache,target=/root/.ccache \
cmake --preset 'ROCm 6' \
cmake --preset 'ROCm 6' -DOLLAMA_RUNNER_DIR="rocm" \
&& cmake --build --parallel ${PARALLEL} --preset 'ROCm 6' \
&& cmake --install build --component HIP --strip --parallel ${PARALLEL}
......@@ -89,7 +89,7 @@ COPY CMakeLists.txt CMakePresets.json .
COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
ARG PARALLEL
RUN --mount=type=cache,target=/root/.ccache \
cmake --preset 'JetPack 5' \
cmake --preset 'JetPack 5' -DOLLAMA_RUNNER_DIR="cuda_jetpack5" \
&& cmake --build --parallel ${PARALLEL} --preset 'JetPack 5' \
&& cmake --install build --component CUDA --strip --parallel ${PARALLEL}
......@@ -101,7 +101,7 @@ COPY CMakeLists.txt CMakePresets.json .
COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
ARG PARALLEL
RUN --mount=type=cache,target=/root/.ccache \
cmake --preset 'JetPack 6' \
cmake --preset 'JetPack 6' -DOLLAMA_RUNNER_DIR="cuda_jetpack6" \
&& cmake --build --parallel ${PARALLEL} --preset 'JetPack 6' \
&& cmake --install build --component CUDA --strip --parallel ${PARALLEL}
......@@ -128,8 +128,8 @@ FROM --platform=linux/arm64 scratch AS arm64
# COPY --from=cuda-11 dist/lib/ollama/ /lib/ollama/
COPY --from=cuda-12 dist/lib/ollama /lib/ollama/
COPY --from=cuda-13 dist/lib/ollama/ /lib/ollama/
COPY --from=jetpack-5 dist/lib/ollama /lib/ollama/cuda_jetpack5
COPY --from=jetpack-6 dist/lib/ollama /lib/ollama/cuda_jetpack6
COPY --from=jetpack-5 dist/lib/ollama/ /lib/ollama/
COPY --from=jetpack-6 dist/lib/ollama/ /lib/ollama/
FROM scratch AS rocm
COPY --from=rocm-6 dist/lib/ollama /lib/ollama
......
discover/amd_common.go deleted 100644 → 0
View file @ 6b50f2b9
//go:build linux || windows
package discover
import (
"errors"
"log/slog"
"os"
"path/filepath"
"runtime"
"strings"
)
// Determine if the given ROCm lib directory is usable by checking for existence of some glob patterns
func rocmLibUsable(libDir string) bool {
slog.Debug("evaluating potential rocm lib dir " + libDir)
for _, g := range ROCmLibGlobs {
res, _ := filepath.Glob(filepath.Join(libDir, g))
if len(res) == 0 {
return false
}
}
return true
}
func GetSupportedGFX(libDir string) ([]string, error) {
var ret []string
files, err := filepath.Glob(filepath.Join(libDir, "rocblas", "library", "TensileLibrary_lazy_gfx*.dat"))
if err != nil {
return nil, err
}
for _, file := range files {
ret = append(ret, strings.TrimSuffix(strings.TrimPrefix(filepath.Base(file), "TensileLibrary_lazy_"), ".dat"))
}
return ret, nil
}
func commonAMDValidateLibDir() (string, error) {
// Favor our bundled version
// Installer payload location if we're running the installed binary
rocmTargetDir := filepath.Join(LibOllamaPath, "rocm")
if rocmLibUsable(rocmTargetDir) {
slog.Debug("detected ROCM next to ollama executable " + rocmTargetDir)
return rocmTargetDir, nil
}
// Prefer explicit HIP env var
hipPath := os.Getenv("HIP_PATH")
if hipPath != "" {
hipLibDir := filepath.Join(hipPath, "bin")
if rocmLibUsable(hipLibDir) {
slog.Debug("detected ROCM via HIP_PATH=" + hipPath)
return hipLibDir, nil
}
}
// Scan the LD_LIBRARY_PATH or PATH
pathEnv := "LD_LIBRARY_PATH"
if runtime.GOOS == "windows" {
pathEnv = "PATH"
}
paths := os.Getenv(pathEnv)
for _, path := range filepath.SplitList(paths) {
d, err := filepath.Abs(path)
if err != nil {
continue
}
if rocmLibUsable(d) {
return d, nil
}
}
// Well known location(s)
for _, path := range RocmStandardLocations {
if rocmLibUsable(path) {
return path, nil
}
}
return "", errors.New("no suitable rocm found, falling back to CPU")
}
discover/amd_hip_windows.go deleted 100644 → 0
View file @ 6b50f2b9
package discover
import (
"errors"
"fmt"
"log/slog"
"syscall"
"unsafe"
"golang.org/x/sys/windows"
)
const (
hipSuccess = 0
hipErrorNoDevice = 100
)
type hipDevicePropMinimal struct {
Name [256]byte
unused1 [140]byte
GcnArchName [256]byte // gfx####
iGPU int // Doesn't seem to actually report correctly
unused2 [128]byte
}
// Wrap the amdhip64.dll library for GPU discovery
type HipLib struct {
dll windows.Handle
hipGetDeviceCount uintptr
hipGetDeviceProperties uintptr
hipMemGetInfo uintptr
hipSetDevice uintptr
hipDriverGetVersion uintptr
}
func NewHipLib() (*HipLib, error) {
// At runtime we depend on v6, so discover GPUs with the same library for a consistent set of GPUs
h, err := windows.LoadLibrary("amdhip64_6.dll")
if err != nil {
return nil, fmt.Errorf("unable to load amdhip64_6.dll, please make sure to upgrade to the latest amd driver: %w", err)
}
hl := &HipLib{}
hl.dll = h
hl.hipGetDeviceCount, err = windows.GetProcAddress(hl.dll, "hipGetDeviceCount")
if err != nil {
return nil, err
}
hl.hipGetDeviceProperties, err = windows.GetProcAddress(hl.dll, "hipGetDeviceProperties")
if err != nil {
return nil, err
}
hl.hipMemGetInfo, err = windows.GetProcAddress(hl.dll, "hipMemGetInfo")
if err != nil {
return nil, err
}
hl.hipSetDevice, err = windows.GetProcAddress(hl.dll, "hipSetDevice")
if err != nil {
return nil, err
}
hl.hipDriverGetVersion, err = windows.GetProcAddress(hl.dll, "hipDriverGetVersion")
if err != nil {
return nil, err
}
return hl, nil
}
// The hip library only evaluates the ROCR_VISIBLE_DEVICES variable at startup
// so we have to unload/reset the library after we do our initial discovery
// to make sure our updates to that variable are processed by llama.cpp
func (hl *HipLib) Release() {
err := windows.FreeLibrary(hl.dll)
if err != nil {
slog.Warn("failed to unload amdhip64.dll", "error", err)
}
hl.dll = 0
}
func (hl *HipLib) AMDDriverVersion() (driverMajor, driverMinor int, err error) {
if hl.dll == 0 {
return 0, 0, errors.New("dll has been unloaded")
}
var version int
status, _, err := syscall.SyscallN(hl.hipDriverGetVersion, uintptr(unsafe.Pointer(&version)))
if status != hipSuccess {
return 0, 0, fmt.Errorf("failed call to hipDriverGetVersion: %d %s", status, err)
}
slog.Debug("hipDriverGetVersion", "version", version)
driverMajor = version / 10000000
driverMinor = (version - (driverMajor * 10000000)) / 100000
return driverMajor, driverMinor, nil
}
func (hl *HipLib) HipGetDeviceCount() int {
if hl.dll == 0 {
slog.Error("dll has been unloaded")
return 0
}
var count int
status, _, err := syscall.SyscallN(hl.hipGetDeviceCount, uintptr(unsafe.Pointer(&count)))
if status == hipErrorNoDevice {
slog.Info("AMD ROCm reports no devices found")
return 0
}
if status != hipSuccess {
slog.Warn("failed call to hipGetDeviceCount", "status", status, "error", err)
}
return count
}
func (hl *HipLib) HipSetDevice(device int) error {
if hl.dll == 0 {
return errors.New("dll has been unloaded")
}
status, _, err := syscall.SyscallN(hl.hipSetDevice, uintptr(device))
if status != hipSuccess {
return fmt.Errorf("failed call to hipSetDevice: %d %s", status, err)
}
return nil
}
func (hl *HipLib) HipGetDeviceProperties(device int) (*hipDevicePropMinimal, error) {
if hl.dll == 0 {
return nil, errors.New("dll has been unloaded")
}
var props hipDevicePropMinimal
status, _, err := syscall.SyscallN(hl.hipGetDeviceProperties, uintptr(unsafe.Pointer(&props)), uintptr(device))
if status != hipSuccess {
return nil, fmt.Errorf("failed call to hipGetDeviceProperties: %d %s", status, err)
}
return &props, nil
}
// free, total, err
func (hl *HipLib) HipMemGetInfo() (uint64, uint64, error) {
if hl.dll == 0 {
return 0, 0, errors.New("dll has been unloaded")
}
var totalMemory uint64
var freeMemory uint64
status, _, err := syscall.SyscallN(hl.hipMemGetInfo, uintptr(unsafe.Pointer(&freeMemory)), uintptr(unsafe.Pointer(&totalMemory)))
if status != hipSuccess {
return 0, 0, fmt.Errorf("failed call to hipMemGetInfo: %d %s", status, err)
}
return freeMemory, totalMemory, nil
}
discover/amd_linux.go deleted 100644 → 0
View file @ 6b50f2b9
package discover
import (
"bufio"
"errors"
"fmt"
"io"
"io/fs"
"log/slog"
"os"
"path/filepath"
"regexp"
"slices"
"sort"
"strconv"
"strings"
"github.com/ollama/ollama/envconfig"
"github.com/ollama/ollama/format"
)
// Discovery logic for AMD/ROCm GPUs
const (
DriverVersionFile = "/sys/module/amdgpu/version"
AMDNodesSysfsDir = "/sys/class/kfd/kfd/topology/nodes/"
GPUPropertiesFileGlob = AMDNodesSysfsDir + "*/properties"
// Prefix with the node dir
GPUTotalMemoryFileGlob = "mem_banks/*/properties" // size_in_bytes line
// Direct Rendering Manager sysfs location
DRMDeviceDirGlob = "/sys/class/drm/card*/device"
DRMTotalMemoryFile = "mem_info_vram_total"
DRMUsedMemoryFile = "mem_info_vram_used"
// In hex; properties file is in decimal
DRMUniqueIDFile = "unique_id"
DRMVendorFile = "vendor"
DRMDeviceFile = "device"
)
var (
// Used to validate if the given ROCm lib is usable
ROCmLibGlobs = []string{"libhipblas.so.2*", "rocblas"} // TODO - probably include more coverage of files here...
RocmStandardLocations = []string{"/opt/rocm/lib", "/usr/lib64"}
)
// Gather GPU information from the amdgpu driver if any supported GPUs are detected
// Only called once during bootstrap
func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
resp := []RocmGPUInfo{}
if !AMDDetected() {
return resp, fmt.Errorf("AMD GPUs not detected")
}
// Opportunistic logging of driver version to aid in troubleshooting
driverMajor, driverMinor, err := AMDDriverVersion()
if err != nil {
// TODO - if we see users crash and burn with the upstreamed kernel this can be adjusted to hard-fail rocm support and fallback to CPU
slog.Warn("ollama recommends running the https://www.amd.com/en/support/download/linux-drivers.html", "error", err)
}
// Determine if the user has already pre-selected which GPUs to look at, then ignore the others
var visibleDevices []string
hipVD := envconfig.HipVisibleDevices() // zero based index only
rocrVD := envconfig.RocrVisibleDevices() // zero based index or UUID
gpuDO := envconfig.GpuDeviceOrdinal() // zero based index
switch {
case rocrVD != "":
visibleDevices = strings.Split(rocrVD, ",")
case hipVD != "":
visibleDevices = strings.Split(hipVD, ",")
case gpuDO != "":
visibleDevices = strings.Split(gpuDO, ",")
}
gfxOverride := envconfig.HsaOverrideGfxVersion()
var supported []string
var libDir string
// The amdgpu driver always exposes the host CPU(s) first, but we have to skip them and subtract
// from the other IDs to get alignment with the HIP libraries expectations (zero is the first GPU, not the CPU)
matches, _ := filepath.Glob(GPUPropertiesFileGlob)
sort.Slice(matches, func(i, j int) bool {
// /sys/class/kfd/kfd/topology/nodes/<number>/properties
a, err := strconv.ParseInt(filepath.Base(filepath.Dir(matches[i])), 10, 64)
if err != nil {
slog.Debug("parse err", "error", err, "match", matches[i])
return false
}
b, err := strconv.ParseInt(filepath.Base(filepath.Dir(matches[j])), 10, 64)
if err != nil {
slog.Debug("parse err", "error", err, "match", matches[i])
return false
}
return a < b
})
gpuCount := 0
gpuOrdinalID := 0
for _, match := range matches {
slog.Debug("evaluating amdgpu node " + match)
fp, err := os.Open(match)
if err != nil {
slog.Debug("failed to open sysfs node", "file", match, "error", err)
continue
}
defer fp.Close()
scanner := bufio.NewScanner(fp)
isCPU := false
var major, minor, patch uint64
var vendor, device, uniqueID uint64
for scanner.Scan() {
line := strings.TrimSpace(scanner.Text())
// Note: we could also use "cpu_cores_count X" where X is greater than zero to detect CPUs
if strings.HasPrefix(line, "gfx_target_version") {
ver := strings.Fields(line)
// Detect CPUs
if len(ver) == 2 && ver[1] == "0" {
slog.Debug("detected CPU " + match)
isCPU = true
break
}
if len(ver) != 2 || len(ver[1]) < 5 {
slog.Warn("malformed "+match, "gfx_target_version", line)
// If this winds up being a CPU, our offsets may be wrong
continue
}
l := len(ver[1])
var err1, err2, err3 error
patch, err1 = strconv.ParseUint(ver[1][l-2:l], 10, 32)
minor, err2 = strconv.ParseUint(ver[1][l-4:l-2], 10, 32)
major, err3 = strconv.ParseUint(ver[1][:l-4], 10, 32)
if err1 != nil || err2 != nil || err3 != nil {
slog.Debug("malformed int " + line)
continue
}
} else if strings.HasPrefix(line, "vendor_id") {
ver := strings.Fields(line)
if len(ver) != 2 {
slog.Debug("malformed", "vendor_id", line)
continue
}
vendor, err = strconv.ParseUint(ver[1], 10, 64)
if err != nil {
slog.Debug("malformed", "vendor_id", line, "error", err)
}
} else if strings.HasPrefix(line, "device_id") {
ver := strings.Fields(line)
if len(ver) != 2 {
slog.Debug("malformed", "device_id", line)
continue
}
device, err = strconv.ParseUint(ver[1], 10, 64)
if err != nil {
slog.Debug("malformed", "device_id", line, "error", err)
}
} else if strings.HasPrefix(line, "unique_id") {
ver := strings.Fields(line)
if len(ver) != 2 {
slog.Debug("malformed", "unique_id", line)
continue
}
uniqueID, err = strconv.ParseUint(ver[1], 10, 64)
if err != nil {
slog.Debug("malformed", "unique_id", line, "error", err)
}
}
// TODO - any other properties we want to extract and record?
// vendor_id + device_id -> pci lookup for "Name"
// Other metrics that may help us understand relative performance between multiple GPUs
}
// Note: while ./mem_banks/*/used_memory exists, it doesn't appear to take other VRAM consumers
// into consideration, so we instead map the device over to the DRM driver sysfs nodes which
// do reliably report VRAM usage.
if isCPU {
continue
}
// Skip over any GPUs that are masked
if major == 0 && minor == 0 && patch == 0 {
slog.Debug("skipping gpu with gfx000")
continue
}
// Look up the memory for the current node
totalMemory := uint64(0)
usedMemory := uint64(0)
var usedFile string
mapping := []struct {
id uint64
filename string
}{
{vendor, DRMVendorFile},
{device, DRMDeviceFile},
{uniqueID, DRMUniqueIDFile}, // Not all devices will report this
}
slog.Debug("mapping amdgpu to drm sysfs nodes", "amdgpu", match, "vendor", vendor, "device", device, "unique_id", uniqueID)
// Map over to DRM location to find the total/free memory
drmMatches, _ := filepath.Glob(DRMDeviceDirGlob)
for _, devDir := range drmMatches {
matched := true
for _, m := range mapping {
if m.id == 0 {
// Null ID means it didn't populate, so we can't use it to match
continue
}
filename := filepath.Join(devDir, m.filename)
buf, err := os.ReadFile(filename)
if err != nil {
slog.Debug("failed to read sysfs node", "file", filename, "error", err)
matched = false
break
}
// values here are in hex, strip off the lead 0x and parse so we can compare the numeric (decimal) values in amdgpu
cmp, err := strconv.ParseUint(strings.TrimPrefix(strings.TrimSpace(string(buf)), "0x"), 16, 64)
if err != nil {
slog.Debug("failed to parse sysfs node", "file", filename, "error", err)
matched = false
break
}
if cmp != m.id {
matched = false
break
}
}
if !matched {
continue
}
// Found the matching DRM directory
slog.Debug("matched", "amdgpu", match, "drm", devDir)
totalFile := filepath.Join(devDir, DRMTotalMemoryFile)
buf, err := os.ReadFile(totalFile)
if err != nil {
slog.Debug("failed to read sysfs node", "file", totalFile, "error", err)
break
}
totalMemory, err = strconv.ParseUint(strings.TrimSpace(string(buf)), 10, 64)
if err != nil {
slog.Debug("failed to parse sysfs node", "file", totalFile, "error", err)
break
}
usedFile = filepath.Join(devDir, DRMUsedMemoryFile)
usedMemory, err = getFreeMemory(usedFile)
if err != nil {
slog.Debug("failed to update used memory", "error", err)
}
break
}
var name string
// TODO - PCI ID lookup
if vendor > 0 && device > 0 {
name = fmt.Sprintf("%04x:%04x", vendor, device)
}
// Favor UUIDs if available to reduce possibility of getting the numeric IDs wrong
var ID string
if uniqueID != 0 {
ID = fmt.Sprintf("GPU-%016x", uniqueID)
} else {
ID = strconv.Itoa(gpuOrdinalID)
}
gpuInfo := RocmGPUInfo{
GpuInfo: GpuInfo{
Library: "rocm",
memInfo: memInfo{
TotalMemory: totalMemory,
FreeMemory: (totalMemory - usedMemory),
},
ID: ID,
filterID: gpuOrdinalID,
Name: name,
Compute: fmt.Sprintf("gfx%d%x%x", major, minor, patch),
MinimumMemory: rocmMinimumMemory,
DriverMajor: driverMajor,
DriverMinor: driverMinor,
},
usedFilepath: usedFile,
index: gpuCount,
}
// Keep track of numeric IDs based on valid GPUs
gpuCount += 1
// If the user wants to filter to a subset of devices, filter out if we aren't a match
if len(visibleDevices) > 0 {
include := false
for _, visible := range visibleDevices {
if (uniqueID != 0 && visible == gpuInfo.ID) || visible == strconv.Itoa(gpuInfo.index) {
include = true
break
}
}
if !include {
reason := "filtering out device per user request"
slog.Info(reason, "id", gpuInfo.ID, "index", gpuInfo.index, "visible_devices", visibleDevices)
unsupportedGPUs = append(unsupportedGPUs, UnsupportedGPUInfo{
GpuInfo: gpuInfo.GpuInfo,
Reason: reason,
})
continue
}
}
// Ordinal IDs are based on the visible GPUs
gpuOrdinalID += 1
// iGPU detection, remove this check once we can support an iGPU variant of the rocm library
if totalMemory < IGPUMemLimit {
reason := "unsupported Radeon iGPU detected skipping"
slog.Info(reason, "id", gpuInfo.ID, "total", format.HumanBytes2(totalMemory))
unsupportedGPUs = append(unsupportedGPUs, UnsupportedGPUInfo{
GpuInfo: gpuInfo.GpuInfo,
Reason: reason,
})
continue
}
minVer, err := strconv.Atoi(RocmComputeMajorMin)
if err != nil {
slog.Error("invalid RocmComputeMajorMin setting", "value", RocmComputeMajorMin, "error", err)
}
if int(major) < minVer {
reason := fmt.Sprintf("amdgpu too old gfx%d%x%x", major, minor, patch)
slog.Warn(reason, "gpu", gpuInfo.ID)
unsupportedGPUs = append(unsupportedGPUs, UnsupportedGPUInfo{
GpuInfo: gpuInfo.GpuInfo,
Reason: reason,
})
continue
}
slog.Debug("amdgpu memory", "gpu", gpuInfo.ID, "total", format.HumanBytes2(totalMemory))
slog.Debug("amdgpu memory", "gpu", gpuInfo.ID, "available", format.HumanBytes2(totalMemory-usedMemory))
// Final validation is gfx compatibility - load the library if we haven't already loaded it
// even if the user overrides, we still need to validate the library
if libDir == "" {
libDir, err = AMDValidateLibDir()
if err != nil {
err = fmt.Errorf("unable to verify rocm library: %w", err)
slog.Warn(err.Error())
unsupportedGPUs = append(unsupportedGPUs, UnsupportedGPUInfo{
GpuInfo: gpuInfo.GpuInfo,
Reason: err.Error(),
})
return nil, err
}
}
gpuInfo.DependencyPath = []string{libDir}
if gfxOverride == "" {
// Only load supported list once
if len(supported) == 0 {
supported, err = GetSupportedGFX(libDir)
if err != nil {
err = fmt.Errorf("failed to lookup supported GFX types: %w", err)
slog.Warn(err.Error())
unsupportedGPUs = append(unsupportedGPUs, UnsupportedGPUInfo{
GpuInfo: gpuInfo.GpuInfo,
Reason: err.Error(),
})
return nil, err
}
slog.Debug("rocm supported GPUs", "types", supported)
}
gfx := gpuInfo.Compute
if !slices.Contains[[]string, string](supported, gfx) {
reason := fmt.Sprintf("amdgpu is not supported (supported types:%s)", supported)
slog.Warn(reason, "gpu_type", gfx, "gpu", gpuInfo.ID, "library", libDir)
unsupportedGPUs = append(unsupportedGPUs, UnsupportedGPUInfo{
GpuInfo: gpuInfo.GpuInfo,
Reason: reason,
})
// TODO - consider discrete markdown just for ROCM troubleshooting?
slog.Warn("See https://github.com/ollama/ollama/blob/main/docs/gpu.md#overrides for HSA_OVERRIDE_GFX_VERSION usage")
continue
} else {
slog.Info("amdgpu is supported", "gpu", gpuInfo.ID, "gpu_type", gfx)
}
} else {
slog.Info("skipping rocm gfx compatibility check", "HSA_OVERRIDE_GFX_VERSION", gfxOverride)
}
// Check for env var workarounds
if name == "1002:687f" { // Vega RX 56
gpuInfo.EnvWorkarounds = append(gpuInfo.EnvWorkarounds, "HSA_ENABLE_SDMA=0")
}
// The GPU has passed all the verification steps and is supported
resp = append(resp, gpuInfo)
}
if len(resp) == 0 {
err := fmt.Errorf("no compatible amdgpu devices detected")
slog.Info(err.Error())
return nil, err
}
if err := verifyKFDDriverAccess(); err != nil {
err = fmt.Errorf("amdgpu devices detected but permission problems block access: %w", err)
slog.Error(err.Error())
return nil, err
}
return resp, nil
}
// Quick check for AMD driver so we can skip amdgpu discovery if not present
func AMDDetected() bool {
// Some driver versions (older?) don't have a version file, so just lookup the parent dir
sysfsDir := filepath.Dir(DriverVersionFile)
_, err := os.Stat(sysfsDir)
if errors.Is(err, os.ErrNotExist) {
slog.Debug("amdgpu driver not detected " + sysfsDir)
return false
} else if err != nil {
slog.Debug("error looking up amd driver", "path", sysfsDir, "error", err)
return false
}
return true
}
// Prefer to use host installed ROCm, as long as it meets our minimum requirements
// failing that, tell the user how to download it on their own
func AMDValidateLibDir() (string, error) {
libDir, err := commonAMDValidateLibDir()
if err == nil {
return libDir, nil
}
// Well known ollama installer path
installedRocmDir := "/usr/share/ollama/lib/rocm"
if rocmLibUsable(installedRocmDir) {
return installedRocmDir, nil
}
// If we still haven't found a usable rocm, the user will have to install it on their own
slog.Warn("amdgpu detected, but no compatible rocm library found. Either install rocm v6, or follow manual install instructions at https://github.com/ollama/ollama/blob/main/docs/linux.md#manual-install")
return "", errors.New("no suitable rocm found, falling back to CPU")
}
func AMDDriverVersion() (driverMajor, driverMinor int, err error) {
_, err = os.Stat(DriverVersionFile)
if err != nil {
return 0, 0, fmt.Errorf("amdgpu version file missing: %s %w", DriverVersionFile, err)
}
fp, err := os.Open(DriverVersionFile)
if err != nil {
return 0, 0, err
}
defer fp.Close()
verString, err := io.ReadAll(fp)
if err != nil {
return 0, 0, err
}
pattern := `\A(\d+)\.(\d+).*`
regex := regexp.MustCompile(pattern)
match := regex.FindStringSubmatch(string(verString))
if len(match) < 2 {
return 0, 0, fmt.Errorf("malformed version string %s", string(verString))
}
driverMajor, err = strconv.Atoi(match[1])
if err != nil {
return 0, 0, err
}
driverMinor, err = strconv.Atoi(match[2])
if err != nil {
return 0, 0, err
}
return driverMajor, driverMinor, nil
}
func (gpus RocmGPUInfoList) RefreshFreeMemory() error {
if len(gpus) == 0 {
return nil
}
for i := range gpus {
usedMemory, err := getFreeMemory(gpus[i].usedFilepath)
if err != nil {
return err
}
slog.Debug("updating rocm free memory", "gpu", gpus[i].ID, "name", gpus[i].Name, "before", format.HumanBytes2(gpus[i].FreeMemory), "now", format.HumanBytes2(gpus[i].TotalMemory-usedMemory))
gpus[i].FreeMemory = gpus[i].TotalMemory - usedMemory
}
return nil
}
func getFreeMemory(usedFile string) (uint64, error) {
buf, err := os.ReadFile(usedFile)
if err != nil {
return 0, fmt.Errorf("failed to read sysfs node %s %w", usedFile, err)
}
usedMemory, err := strconv.ParseUint(strings.TrimSpace(string(buf)), 10, 64)
if err != nil {
slog.Debug("failed to parse sysfs node", "file", usedFile, "error", err)
return 0, fmt.Errorf("failed to parse sysfs node %s %w", usedFile, err)
}
return usedMemory, nil
}
func verifyKFDDriverAccess() error {
// 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, 0o666)
if err != nil {
if errors.Is(err, fs.ErrPermission) {
return fmt.Errorf("permissions not set up properly. Either run ollama as root, or add you user account to the render group. %w", err)
} else if errors.Is(err, fs.ErrNotExist) {
// Container runtime failure?
return fmt.Errorf("kfd driver not loaded. If running in a container, remember to include '--device /dev/kfd --device /dev/dri'")
}
return fmt.Errorf("failed to check permission on /dev/kfd: %w", err)
}
fd.Close()
return nil
}
func rocmGetVisibleDevicesEnv(gpuInfo []GpuInfo) string {
ids := []string{}
for _, info := range gpuInfo {
if info.Library != "rocm" {
continue
}
// If the devices requires a numeric ID, for filtering purposes, we use the unfiltered ID number
if _, err := strconv.Atoi(info.ID); err == nil {
ids = append(ids, fmt.Sprintf("%d", info.filterID))
} else {
ids = append(ids, info.ID)
}
}
if len(ids) == 0 {
return ""
}
// There are 3 potential env vars to use to select GPUs.
// ROCR_VISIBLE_DEVICES supports UUID or numeric so is our preferred on linux
// GPU_DEVICE_ORDINAL supports numeric IDs only
// HIP_VISIBLE_DEVICES supports numeric IDs only
return "ROCR_VISIBLE_DEVICES=" + strings.Join(ids, ",")
}
discover/amd_windows.go deleted 100644 → 0
View file @ 6b50f2b9
package discover
import (
"bytes"
"errors"
"fmt"
"log/slog"
"path/filepath"
"slices"
"strconv"
"strings"
"github.com/ollama/ollama/envconfig"
"github.com/ollama/ollama/format"
)
const (
// TODO We're lookinng for this exact name to detect iGPUs since hipGetDeviceProperties never reports integrated==true
iGPUName = "AMD Radeon(TM) Graphics"
)
var (
// Used to validate if the given ROCm lib is usable
ROCmLibGlobs = []string{"hipblas.dll", "rocblas"} // This is not sufficient to discern v5 vs v6
RocmStandardLocations = []string{"C:\\Program Files\\AMD\\ROCm\\6.1\\bin"} // TODO glob?
)
// Only called once during bootstrap
func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
resp := []RocmGPUInfo{}
hl, err := NewHipLib()
if err != nil {
slog.Debug(err.Error())
return nil, err
}
defer hl.Release()
driverMajor, driverMinor, err := hl.AMDDriverVersion()
if err != nil {
// For now this is benign, but we may eventually need to fail compatibility checks
slog.Debug("error looking up amd driver version", "error", err)
}
// Note: the HIP library automatically handles subsetting to any *_VISIBLE_DEVICES the user specified
count := hl.HipGetDeviceCount()
if count == 0 {
err := fmt.Errorf("no compatible amdgpu devices detected")
slog.Info(err.Error())
return nil, err
}
libDir, err := AMDValidateLibDir()
if err != nil {
err = fmt.Errorf("unable to verify rocm library: %w", err)
slog.Warn(err.Error())
return nil, err
}
var supported []string
gfxOverride := envconfig.HsaOverrideGfxVersion()
if gfxOverride == "" {
supported, err = GetSupportedGFX(libDir)
if err != nil {
err = fmt.Errorf("failed to lookup supported GFX types: %w", err)
slog.Warn(err.Error())
return nil, err
}
} else {
slog.Info("skipping rocm gfx compatibility check", "HSA_OVERRIDE_GFX_VERSION", gfxOverride)
}
slog.Debug("detected hip devices", "count", count)
// TODO how to determine the underlying device ID when visible devices is causing this to subset?
for i := range count {
err = hl.HipSetDevice(i)
if err != nil {
slog.Warn("set device", "id", i, "error", err)
continue
}
props, err := hl.HipGetDeviceProperties(i)
if err != nil {
slog.Warn("get properties", "id", i, "error", err)
continue
}
n := bytes.IndexByte(props.Name[:], 0)
name := string(props.Name[:n])
// TODO is UUID actually populated on windows?
// Can luid be used on windows for setting visible devices (and is it actually set?)
n = bytes.IndexByte(props.GcnArchName[:], 0)
gfx := string(props.GcnArchName[:n])
slog.Debug("hip device", "id", i, "name", name, "gfx", gfx)
// slog.Info(fmt.Sprintf("[%d] Integrated: %d", i, props.iGPU)) // DOESN'T REPORT CORRECTLY! Always 0
// TODO Why isn't props.iGPU accurate!?
freeMemory, totalMemory, err := hl.HipMemGetInfo()
if err != nil {
slog.Warn("get mem info", "id", i, "error", err)
continue
}
gpuInfo := RocmGPUInfo{
GpuInfo: GpuInfo{
Library: "rocm",
memInfo: memInfo{
TotalMemory: totalMemory,
FreeMemory: freeMemory,
},
// Free memory reporting on Windows is not reliable until we bump to ROCm v6.2
UnreliableFreeMemory: true,
ID: strconv.Itoa(i), // TODO this is probably wrong if we specify visible devices
filterID: i,
DependencyPath: []string{libDir},
MinimumMemory: rocmMinimumMemory,
Name: name,
Compute: gfx,
DriverMajor: driverMajor,
DriverMinor: driverMinor,
},
index: i,
}
// iGPU detection, remove this check once we can support an iGPU variant of the rocm library
if strings.EqualFold(name, iGPUName) || totalMemory < IGPUMemLimit {
reason := "unsupported Radeon iGPU detected skipping"
slog.Info(reason, "id", gpuInfo.ID, "total", format.HumanBytes2(totalMemory))
unsupportedGPUs = append(unsupportedGPUs, UnsupportedGPUInfo{
GpuInfo: gpuInfo.GpuInfo,
Reason: reason,
})
continue
}
// Strip off Target Features when comparing
if !slices.Contains[[]string, string](supported, strings.Split(gfx, ":")[0]) {
reason := fmt.Sprintf("amdgpu is not supported (supported types:%s)", supported)
slog.Warn(reason, "gpu_type", gfx, "gpu", gpuInfo.ID, "library", libDir)
unsupportedGPUs = append(unsupportedGPUs, UnsupportedGPUInfo{
GpuInfo: gpuInfo.GpuInfo,
Reason: reason,
})
// HSA_OVERRIDE_GFX_VERSION not supported on windows
continue
} else {
slog.Debug("amdgpu is supported", "gpu", i, "gpu_type", gfx)
}
slog.Debug("amdgpu memory", "gpu", i, "total", format.HumanBytes2(totalMemory))
slog.Debug("amdgpu memory", "gpu", i, "available", format.HumanBytes2(freeMemory))
resp = append(resp, gpuInfo)
}
return resp, nil
}
func AMDValidateLibDir() (string, error) {
libDir, err := commonAMDValidateLibDir()
if err == nil {
return libDir, nil
}
// Installer payload (if we're running from some other location)
rocmTargetDir := filepath.Join(LibOllamaPath, "rocm")
if rocmLibUsable(rocmTargetDir) {
slog.Debug("detected ollama installed ROCm at " + rocmTargetDir)
return rocmTargetDir, nil
}
// Should not happen on windows since we include it in the installer, but stand-alone binary might hit this
slog.Warn("amdgpu detected, but no compatible rocm library found. Please install ROCm")
return "", errors.New("no suitable rocm found, falling back to CPU")
}
func (gpus RocmGPUInfoList) RefreshFreeMemory() error {
if len(gpus) == 0 {
return nil
}
hl, err := NewHipLib()
if err != nil {
slog.Debug(err.Error())
return err
}
defer hl.Release()
for i := range gpus {
err := hl.HipSetDevice(gpus[i].index)
if err != nil {
return err
}
freeMemory, _, err := hl.HipMemGetInfo()
if err != nil {
slog.Warn("get mem info", "id", i, "error", err)
continue
}
slog.Debug("updating rocm free memory", "gpu", gpus[i].ID, "name", gpus[i].Name, "before", format.HumanBytes2(gpus[i].FreeMemory), "now", format.HumanBytes2(freeMemory))
gpus[i].FreeMemory = freeMemory
}
return nil
}
func rocmGetVisibleDevicesEnv(gpuInfo []GpuInfo) string {
ids := []string{}
for _, info := range gpuInfo {
if info.Library != "rocm" {
continue
}
// If the devices requires a numeric ID, for filtering purposes, we use the unfiltered ID number
if _, err := strconv.Atoi(info.ID); err == nil {
ids = append(ids, fmt.Sprintf("%d", info.filterID))
} else {
ids = append(ids, info.ID)
}
}
if len(ids) == 0 {
return ""
}
// There are 3 potential env vars to use to select GPUs.
// ROCR_VISIBLE_DEVICES supports UUID or numeric but does not work on Windows
// HIP_VISIBLE_DEVICES supports numeric IDs only
// GPU_DEVICE_ORDINAL supports numeric IDs only
return "HIP_VISIBLE_DEVICES=" + strings.Join(ids, ",")
}
discover/cpu_common.go deleted 100644 → 0
View file @ 6b50f2b9
package discover
import (
"os"
"path/filepath"
"runtime"
"strings"
)
func IsNUMA() bool {
if runtime.GOOS != "linux" {
// numa support in llama.cpp is linux only
return false
}
ids := map[string]any{}
packageIds, _ := filepath.Glob("/sys/devices/system/cpu/cpu*/topology/physical_package_id")
for _, packageId := range packageIds {
id, err := os.ReadFile(packageId)
if err == nil {
ids[strings.TrimSpace(string(id))] = struct{}{}
}
}
return len(ids) > 1
}
discover/gpu_linux.go → discover/cpu_linux.go
View file @ bc8909fb
......@@ -4,7 +4,9 @@ import (
"bufio"
"fmt"
"io"
"log/slog"
"os"
"path/filepath"
"reflect"
"regexp"
"sort"
......@@ -13,47 +15,6 @@ import (
"github.com/ollama/ollama/format"
)
var CudartGlobs = []string{
"/usr/local/cuda/lib64/libcudart.so*",
"/usr/lib/x86_64-linux-gnu/nvidia/current/libcudart.so*",
"/usr/lib/x86_64-linux-gnu/libcudart.so*",
"/usr/lib/wsl/lib/libcudart.so*",
"/usr/lib/wsl/drivers/*/libcudart.so*",
"/opt/cuda/lib64/libcudart.so*",
"/usr/local/cuda*/targets/aarch64-linux/lib/libcudart.so*",
"/usr/lib/aarch64-linux-gnu/nvidia/current/libcudart.so*",
"/usr/lib/aarch64-linux-gnu/libcudart.so*",
"/usr/local/cuda/lib*/libcudart.so*",
"/usr/lib*/libcudart.so*",
"/usr/local/lib*/libcudart.so*",
}
var NvmlGlobs = []string{}
var NvcudaGlobs = []string{
"/usr/local/cuda*/targets/*/lib/libcuda.so*",
"/usr/lib/*-linux-gnu/nvidia/current/libcuda.so*",
"/usr/lib/*-linux-gnu/libcuda.so*",
"/usr/lib/wsl/lib/libcuda.so*",
"/usr/lib/wsl/drivers/*/libcuda.so*",
"/opt/cuda/lib*/libcuda.so*",
"/usr/local/cuda/lib*/libcuda.so*",
"/usr/lib*/libcuda.so*",
"/usr/local/lib*/libcuda.so*",
}
var OneapiGlobs = []string{
"/usr/lib/x86_64-linux-gnu/libze_intel_gpu.so*",
"/usr/lib*/libze_intel_gpu.so*",
}
var (
CudartMgmtName = "libcudart.so*"
NvcudaMgmtName = "libcuda.so*"
NvmlMgmtName = "" // not currently wired on linux
OneapiMgmtName = "libze_intel_gpu.so*"
)
func GetCPUMem() (memInfo, error) {
var mem memInfo
var total, available, free, buffers, cached, freeSwap uint64
......@@ -106,16 +67,17 @@ type linuxCpuInfo struct {
CoreID string `cpuinfo:"core id"`
}
func GetCPUDetails() ([]CPU, error) {
func GetCPUDetails() []CPU {
file, err := os.Open(CpuInfoFilename)
if err != nil {
return nil, err
slog.Warn("failed to get CPU details", "error", err)
return nil
}
defer file.Close()
return linuxCPUDetails(file)
}
func linuxCPUDetails(file io.Reader) ([]CPU, error) {
func linuxCPUDetails(file io.Reader) []CPU {
reColumns := regexp.MustCompile("\t+: ")
scanner := bufio.NewScanner(file)
cpuInfos := []linuxCpuInfo{}
......@@ -194,5 +156,17 @@ func linuxCPUDetails(file io.Reader) ([]CPU, error) {
for _, k := range keys {
result = append(result, *socketByID[k])
}
return result, nil
return result
}
func IsNUMA() bool {
ids := map[string]any{}
packageIds, _ := filepath.Glob("/sys/devices/system/cpu/cpu*/topology/physical_package_id")
for _, packageId := range packageIds {
id, err := os.ReadFile(packageId)
if err == nil {
ids[strings.TrimSpace(string(id))] = struct{}{}
}
}
return len(ids) > 1
}
discover/gpu_linux_test.go → discover/cpu_linux_test.go
View file @ bc8909fb
......@@ -2062,10 +2062,7 @@ power management:
for k, v := range testCases {
t.Run(k, func(t *testing.T) {
buf := bytes.NewBufferString(v.input)
cpus, err := linuxCPUDetails(buf)
if err != nil {
t.Fatal(err)
}
cpus := linuxCPUDetails(buf)
slog.Info("example", "scenario", k, "cpus", cpus)
si := SystemInfo{
......
discover/gpu_windows.go → discover/cpu_windows.go
View file @ bc8909fb
......@@ -26,29 +26,6 @@ var (
GetLogicalProcessorInformationEx = k32.NewProc("GetLogicalProcessorInformationEx")
)
var CudartGlobs = []string{
"c:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v*\\bin\\cudart64_*.dll",
}
var NvmlGlobs = []string{
"c:\\Windows\\System32\\nvml.dll",
}
var NvcudaGlobs = []string{
"c:\\windows\\system*\\nvcuda.dll",
}
var OneapiGlobs = []string{
"c:\\Windows\\System32\\DriverStore\\FileRepository\\*\\ze_intel_gpu64.dll",
}
var (
CudartMgmtName = "cudart64_*.dll"
NvcudaMgmtName = "nvcuda.dll"
NvmlMgmtName = "nvml.dll"
OneapiMgmtName = "ze_intel_gpu64.dll"
)
func GetCPUMem() (memInfo, error) {
memStatus := MEMORYSTATUSEX{length: sizeofMemoryStatusEx}
r1, _, err := globalMemoryStatusExProc.Call(uintptr(unsafe.Pointer(&memStatus)))
......@@ -122,27 +99,22 @@ func (pkg *winPackage) IsMember(target *GROUP_AFFINITY) bool {
}
func getLogicalProcessorInformationEx() ([]byte, error) {
buf := make([]byte, 1)
buf := make([]byte, 1024)
bufSize := len(buf)
ret, _, err := GetLogicalProcessorInformationEx.Call(
uintptr(RelationAll),
uintptr(unsafe.Pointer(&buf[0])),
uintptr(unsafe.Pointer(&bufSize)),
)
if ret != 0 {
return nil, fmt.Errorf("failed to determine size info ret:%d %w", ret, err)
}
buf = make([]byte, bufSize)
ret, _, err = GetLogicalProcessorInformationEx.Call(
uintptr(RelationAll),
uintptr(unsafe.Pointer(&buf[0])),
uintptr(unsafe.Pointer(&bufSize)),
)
if ret == 0 {
return nil, fmt.Errorf("failed to gather processor information ret:%d buflen:%d %w", ret, bufSize, err)
var err error
for range 3 {
var ret uintptr
ret, _, err = GetLogicalProcessorInformationEx.Call(
uintptr(RelationAll),
uintptr(unsafe.Pointer(&buf[0])),
uintptr(unsafe.Pointer(&bufSize)),
)
if ret == 1 && bufSize <= len(buf) {
return buf, nil
}
buf = make([]byte, bufSize)
}
return buf, nil
return nil, fmt.Errorf("unable to determine CPU details: %w", err)
}
func processSystemLogicalProcessorInforationList(buf []byte) []*winPackage {
......@@ -217,10 +189,11 @@ func processSystemLogicalProcessorInforationList(buf []byte) []*winPackage {
return packages
}
func GetCPUDetails() ([]CPU, error) {
func GetCPUDetails() []CPU {
buf, err := getLogicalProcessorInformationEx()
if err != nil {
return nil, err
slog.Warn("failed to get CPU details", "error", err)
return nil
}
packages := processSystemLogicalProcessorInforationList(buf)
cpus := make([]CPU, len(packages))
......@@ -230,5 +203,10 @@ func GetCPUDetails() ([]CPU, error) {
cpus[i].EfficiencyCoreCount = pkg.efficiencyCoreCount
cpus[i].ThreadCount = pkg.threadCount
}
return cpus, nil
return cpus
}
func IsNUMA() bool {
// numa support in ggml is linux only
return false
}
discover/cuda_common.go deleted 100644 → 0
View file @ 6b50f2b9
//go:build linux || windows
package discover
import (
"fmt"
"log/slog"
"os"
"regexp"
"runtime"
"strconv"
"strings"
)
// Jetson devices have JETSON_JETPACK="x.y.z" factory set to the Jetpack version installed.
// Included to drive logic for reducing Ollama-allocated overhead on L4T/Jetson devices.
var CudaTegra string = os.Getenv("JETSON_JETPACK")
func cudaVariant(gpuInfos []CudaGPUInfo) string {
if runtime.GOARCH == "arm64" && runtime.GOOS == "linux" {
if CudaTegra != "" {
ver := strings.Split(CudaTegra, ".")
if len(ver) > 0 {
return "jetpack" + ver[0]
}
} else if data, err := os.ReadFile("/etc/nv_tegra_release"); err == nil {
r := regexp.MustCompile(` R(\d+) `)
m := r.FindSubmatch(data)
if len(m) != 2 {
slog.Info("Unexpected format for /etc/nv_tegra_release. Set JETSON_JETPACK to select version")
} else {
if l4t, err := strconv.Atoi(string(m[1])); err == nil {
// Note: mapping from L4t -> JP is inconsistent (can't just subtract 30)
// https://developer.nvidia.com/embedded/jetpack-archive
switch l4t {
case 35:
return "jetpack5"
case 36:
return "jetpack6"
default:
slog.Info("unsupported L4T version", "nv_tegra_release", string(data))
}
}
}
}
}
// Check GPU compute capability FIRST, lowest common denominator if multi-gpu
for _, gpuInfo := range gpuInfos {
if gpuInfo.computeMajor < 7 || (gpuInfo.computeMajor == 7 && gpuInfo.computeMinor < 5) {
// GPU is Pascal or older (CC <= 7.4) - use CUDA v12 (supports CC 6.1)
return "v12"
}
}
// GPU is Turing or newer (CC >= 7.5) - can use newer CUDA
if len(gpuInfos) > 0 && gpuInfos[0].DriverMajor < 13 {
// The detected driver is older than 580 (Aug 2025)
// Warn if their CC is compatible with v13 and they should upgrade their driver to get better performance
slog.Warn("old CUDA driver detected - please upgrade to a newer driver for best performance", "version", fmt.Sprintf("%d.%d", gpuInfos[0].DriverMajor, gpuInfos[0].DriverMinor))
return "v12"
}
return "v13"
}
discover/gpu.go
View file @ bc8909fb
This diff is collapsed.
discover/gpu_darwin.go
View file @ bc8909fb
//go:build darwin
package discover
/*
......@@ -11,7 +9,6 @@ import "C"
import (
"log/slog"
"runtime"
"syscall"
"github.com/ollama/ollama/format"
......@@ -21,39 +18,6 @@ const (
metalMinimumMemory = 512 * format.MebiByte
)
func GetGPUInfo() GpuInfoList {
mem, _ := GetCPUMem()
if runtime.GOARCH == "amd64" {
return []GpuInfo{
{
Library: "cpu",
memInfo: mem,
},
}
}
info := GpuInfo{
Library: "metal",
ID: "0",
}
info.TotalMemory = uint64(C.getRecommendedMaxVRAM())
// TODO is there a way to gather actual allocated video memory? (currentAllocatedSize doesn't work)
info.FreeMemory = info.TotalMemory
info.MinimumMemory = metalMinimumMemory
return []GpuInfo{info}
}
func GetCPUInfo() GpuInfoList {
mem, _ := GetCPUMem()
return []GpuInfo{
{
Library: "cpu",
memInfo: mem,
},
}
}
func GetCPUMem() (memInfo, error) {
return memInfo{
TotalMemory: uint64(C.getPhysicalMemory()),
......@@ -62,13 +26,7 @@ func GetCPUMem() (memInfo, error) {
}, nil
}
func (l GpuInfoList) GetVisibleDevicesEnv() []string {
// No-op on darwin
return nil
}
func GetSystemInfo() SystemInfo {
mem, _ := GetCPUMem()
func GetCPUDetails() []CPU {
query := "hw.perflevel0.physicalcpu"
perfCores, err := syscall.SysctlUint32(query)
if err != nil {
......@@ -81,19 +39,16 @@ func GetSystemInfo() SystemInfo {
query = "hw.logicalcpu"
logicalCores, _ := syscall.SysctlUint32(query)
return SystemInfo{
System: CPUInfo{
GpuInfo: GpuInfo{
memInfo: mem,
},
CPUs: []CPU{
{
CoreCount: int(perfCores + efficiencyCores),
EfficiencyCoreCount: int(efficiencyCores),
ThreadCount: int(logicalCores),
},
},
return []CPU{
{
CoreCount: int(perfCores + efficiencyCores),
EfficiencyCoreCount: int(efficiencyCores),
ThreadCount: int(logicalCores),
},
GPUs: GetGPUInfo(),
}
}
func IsNUMA() bool {
// numa support in ggml is linux only
return false
}
discover/gpu_info.h deleted 100644 → 0
View file @ 6b50f2b9
#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() strdup(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)
#define LOAD_ERR() ({\
LPSTR messageBuffer = NULL; \
size_t size = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS, \
NULL, GetLastError(), MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&messageBuffer, 0, NULL); \
char *resp = strdup(messageBuffer); \
LocalFree(messageBuffer); \
resp; \
})
#endif
#ifndef LOG
#define LOG(verbose, ...) \
do { \
if (verbose) { \
fprintf(stderr, __VA_ARGS__); \
} \
} while (0)
#endif
#ifdef __cplusplus
extern "C" {
#endif
#define GPU_ID_LEN 64
#define GPU_NAME_LEN 96
typedef struct mem_info {
char *err; // If non-nill, caller responsible for freeing
char gpu_id[GPU_ID_LEN];
char gpu_name[GPU_NAME_LEN];
uint64_t total;
uint64_t free;
uint64_t used;
// Compute Capability
int major;
int minor;
int patch;
} mem_info_t;
void cpu_check_ram(mem_info_t *resp);
#ifdef __cplusplus
}
#endif
#include "gpu_info_cudart.h"
#include "gpu_info_nvcuda.h"
#include "gpu_info_nvml.h"
#include "gpu_info_oneapi.h"
#endif // __GPU_INFO_H__
#endif // __APPLE__
discover/gpu_info_cudart.c deleted 100644 → 0
View file @ 6b50f2b9
#ifndef __APPLE__ // TODO - maybe consider nvidia support on intel macs?
#include <string.h>
#include <inttypes.h>
#include "gpu_info_cudart.h"
void cudart_init(char *cudart_lib_path, cudart_init_resp_t *resp) {
cudartReturn_t ret;
resp->err = NULL;
resp->num_devices = 0;
const int buflen = 256;
char buf[buflen + 1];
int i;
struct lookup {
char *s;
void **p;
} l[] = {
{"cudaSetDevice", (void *)&resp->ch.cudaSetDevice},
{"cudaDeviceSynchronize", (void *)&resp->ch.cudaDeviceSynchronize},
{"cudaDeviceReset", (void *)&resp->ch.cudaDeviceReset},
{"cudaMemGetInfo", (void *)&resp->ch.cudaMemGetInfo},
{"cudaGetDeviceCount", (void *)&resp->ch.cudaGetDeviceCount},
{"cudaDeviceGetAttribute", (void *)&resp->ch.cudaDeviceGetAttribute},
{"cudaDriverGetVersion", (void *)&resp->ch.cudaDriverGetVersion},
{"cudaGetDeviceProperties", (void *)&resp->ch.cudaGetDeviceProperties},
{NULL, NULL},
};
resp->ch.handle = LOAD_LIBRARY(cudart_lib_path, RTLD_LAZY);
if (!resp->ch.handle) {
char *msg = LOAD_ERR();
LOG(resp->ch.verbose, "library %s load err: %s\n", cudart_lib_path, msg);
snprintf(buf, buflen,
"Unable to load %s library to query for Nvidia GPUs: %s",
cudart_lib_path, msg);
free(msg);
resp->err = strdup(buf);
return;
}
for (i = 0; l[i].s != NULL; i++) {
*l[i].p = LOAD_SYMBOL(resp->ch.handle, l[i].s);
if (!*(l[i].p)) {
char *msg = LOAD_ERR();
LOG(resp->ch.verbose, "dlerr: %s\n", msg);
UNLOAD_LIBRARY(resp->ch.handle);
resp->ch.handle = NULL;
snprintf(buf, buflen, "symbol lookup for %s failed: %s", l[i].s,
msg);
free(msg);
resp->err = strdup(buf);
return;
}
}
ret = (*resp->ch.cudaSetDevice)(0);
if (ret != CUDART_SUCCESS) {
LOG(resp->ch.verbose, "cudaSetDevice err: %d\n", ret);
UNLOAD_LIBRARY(resp->ch.handle);
resp->ch.handle = NULL;
if (ret == CUDART_ERROR_INSUFFICIENT_DRIVER) {
resp->err = strdup("your nvidia driver is too old or missing. If you have a CUDA GPU please upgrade to run ollama");
return;
}
snprintf(buf, buflen, "cudart init failure: %d", ret);
resp->err = strdup(buf);
return;
}
int version = 0;
// Report driver version if we're in verbose mode, ignore errors
ret = (*resp->ch.cudaDriverGetVersion)(&version);
if (ret != CUDART_SUCCESS) {
LOG(resp->ch.verbose, "cudaDriverGetVersion failed: %d\n", ret);
} else {
resp->ch.driver_major = version / 1000;
resp->ch.driver_minor = (version - (resp->ch.driver_major * 1000)) / 10;
LOG(resp->ch.verbose, "CUDA driver version: %d-%d\n", resp->ch.driver_major, resp->ch.driver_minor);
}
ret = (*resp->ch.cudaGetDeviceCount)(&resp->num_devices);
if (ret != CUDART_SUCCESS) {
LOG(resp->ch.verbose, "cudaGetDeviceCount err: %d\n", ret);
UNLOAD_LIBRARY(resp->ch.handle);
resp->ch.handle = NULL;
snprintf(buf, buflen, "unable to get device count: %d", ret);
resp->err = strdup(buf);
return;
}
}
void cudart_bootstrap(cudart_handle_t h, int i, mem_info_t *resp) {
resp->err = NULL;
cudartMemory_t memInfo = {0,0,0};
cudartReturn_t ret;
const int buflen = 256;
char buf[buflen + 1];
if (h.handle == NULL) {
resp->err = strdup("cudart handle isn't initialized");
return;
}
ret = (*h.cudaSetDevice)(i);
if (ret != CUDART_SUCCESS) {
snprintf(buf, buflen, "cudart device failed to initialize");
resp->err = strdup(buf);
return;
}
cudaDeviceProp_t props;
ret = (*h.cudaGetDeviceProperties)(&props, i);
if (ret != CUDART_SUCCESS) {
LOG(h.verbose, "[%d] device properties lookup failure: %d\n", i, ret);
snprintf(&resp->gpu_id[0], GPU_ID_LEN, "%d", i);
resp->major = 0;
resp->minor = 0;
} else {
int allNull = 1;
for (int j = 0; j < 16; j++) {
if (props.uuid.bytes[j] != 0) {
allNull = 0;
break;
}
}
if (allNull != 0) {
snprintf(&resp->gpu_id[0], GPU_ID_LEN, "%d", i);
} else {
// GPU-d110a105-ac29-1d54-7b49-9c90440f215b
snprintf(&resp->gpu_id[0], GPU_ID_LEN,
"GPU-%02x%02x%02x%02x-%02x%02x-%02x%02x-%02x%02x-%02x%02x%02x%02x%02x%02x",
props.uuid.bytes[0],
props.uuid.bytes[1],
props.uuid.bytes[2],
props.uuid.bytes[3],
props.uuid.bytes[4],
props.uuid.bytes[5],
props.uuid.bytes[6],
props.uuid.bytes[7],
props.uuid.bytes[8],
props.uuid.bytes[9],
props.uuid.bytes[10],
props.uuid.bytes[11],
props.uuid.bytes[12],
props.uuid.bytes[13],
props.uuid.bytes[14],
props.uuid.bytes[15]
);
}
resp->major = props.major;
resp->minor = props.minor;
// TODO add other useful properties from props
}
ret = (*h.cudaMemGetInfo)(&memInfo.free, &memInfo.total);
if (ret != CUDART_SUCCESS) {
snprintf(buf, buflen, "cudart device memory info lookup failure %d", ret);
resp->err = strdup(buf);
return;
}
resp->total = memInfo.total;
resp->free = memInfo.free;
resp->used = memInfo.used;
LOG(h.verbose, "[%s] CUDA totalMem %" PRId64 "\n", resp->gpu_id, resp->total);
LOG(h.verbose, "[%s] CUDA freeMem %" PRId64 "\n", resp->gpu_id, resp->free);
LOG(h.verbose, "[%s] CUDA usedMem %" PRId64 "\n", resp->gpu_id, resp->used);
LOG(h.verbose, "[%s] Compute Capability %d.%d\n", resp->gpu_id, resp->major, resp->minor);
}
void cudart_release(cudart_handle_t h) {
LOG(h.verbose, "releasing cudart library\n");
UNLOAD_LIBRARY(h.handle);
h.handle = NULL;
}
#endif // __APPLE__
discover/gpu_info_cudart.h deleted 100644 → 0
View file @ 6b50f2b9
#ifndef __APPLE__
#ifndef __GPU_INFO_CUDART_H__
#define __GPU_INFO_CUDART_H__
#include "gpu_info.h"
// Just enough typedef's to dlopen/dlsym for memory information
typedef enum cudartReturn_enum {
CUDART_SUCCESS = 0,
CUDART_ERROR_INVALID_VALUE = 1,
CUDART_ERROR_MEMORY_ALLOCATION = 2,
CUDART_ERROR_INSUFFICIENT_DRIVER = 35,
// Other values omitted for now...
} cudartReturn_t;
typedef enum cudartDeviceAttr_enum {
cudartDevAttrComputeCapabilityMajor = 75,
cudartDevAttrComputeCapabilityMinor = 76,
// TODO - not yet wired up but may be useful for Jetson or other
// integrated GPU scenarios with shared memory
cudaDevAttrIntegrated = 18
} cudartDeviceAttr_t;
typedef void *cudartDevice_t; // Opaque is sufficient
typedef struct cudartMemory_st {
size_t total;
size_t free;
size_t used;
} cudartMemory_t;
typedef struct cudaUUID {
unsigned char bytes[16];
} cudaUUID_t;
typedef struct cudaDeviceProp {
char name[256]; /**< ASCII string identifying device */
cudaUUID_t uuid; /**< 16-byte unique identifier */
char luid[8]; /**< 8-byte locally unique identifier. Value is undefined on TCC and non-Windows platforms */
unsigned int luidDeviceNodeMask; /**< LUID device node mask. Value is undefined on TCC and non-Windows platforms */
size_t totalGlobalMem; /**< Global memory available on device in bytes */
size_t sharedMemPerBlock; /**< Shared memory available per block in bytes */
int regsPerBlock; /**< 32-bit registers available per block */
int warpSize; /**< Warp size in threads */
size_t memPitch; /**< Maximum pitch in bytes allowed by memory copies */
int maxThreadsPerBlock; /**< Maximum number of threads per block */
int maxThreadsDim[3]; /**< Maximum size of each dimension of a block */
int maxGridSize[3]; /**< Maximum size of each dimension of a grid */
int clockRate; /**< Clock frequency in kilohertz */
size_t totalConstMem; /**< Constant memory available on device in bytes */
int major; /**< Major compute capability */
int minor; /**< Minor compute capability */
size_t textureAlignment; /**< Alignment requirement for textures */
size_t texturePitchAlignment; /**< Pitch alignment requirement for texture references bound to pitched memory */
int deviceOverlap; /**< Device can concurrently copy memory and execute a kernel. Deprecated. Use instead asyncEngineCount. */
int multiProcessorCount; /**< Number of multiprocessors on device */
int kernelExecTimeoutEnabled; /**< Specified whether there is a run time limit on kernels */
int integrated; /**< Device is integrated as opposed to discrete */
int canMapHostMemory; /**< Device can map host memory with cudaHostAlloc/cudaHostGetDevicePointer */
int computeMode; /**< Compute mode (See ::cudaComputeMode) */
int maxTexture1D; /**< Maximum 1D texture size */
int maxTexture1DMipmap; /**< Maximum 1D mipmapped texture size */
int maxTexture1DLinear; /**< Deprecated, do not use. Use cudaDeviceGetTexture1DLinearMaxWidth() or cuDeviceGetTexture1DLinearMaxWidth() instead. */
int maxTexture2D[2]; /**< Maximum 2D texture dimensions */
int maxTexture2DMipmap[2]; /**< Maximum 2D mipmapped texture dimensions */
int maxTexture2DLinear[3]; /**< Maximum dimensions (width, height, pitch) for 2D textures bound to pitched memory */
int maxTexture2DGather[2]; /**< Maximum 2D texture dimensions if texture gather operations have to be performed */
int maxTexture3D[3]; /**< Maximum 3D texture dimensions */
int maxTexture3DAlt[3]; /**< Maximum alternate 3D texture dimensions */
int maxTextureCubemap; /**< Maximum Cubemap texture dimensions */
int maxTexture1DLayered[2]; /**< Maximum 1D layered texture dimensions */
int maxTexture2DLayered[3]; /**< Maximum 2D layered texture dimensions */
int maxTextureCubemapLayered[2];/**< Maximum Cubemap layered texture dimensions */
int maxSurface1D; /**< Maximum 1D surface size */
int maxSurface2D[2]; /**< Maximum 2D surface dimensions */
int maxSurface3D[3]; /**< Maximum 3D surface dimensions */
int maxSurface1DLayered[2]; /**< Maximum 1D layered surface dimensions */
int maxSurface2DLayered[3]; /**< Maximum 2D layered surface dimensions */
int maxSurfaceCubemap; /**< Maximum Cubemap surface dimensions */
int maxSurfaceCubemapLayered[2];/**< Maximum Cubemap layered surface dimensions */
size_t surfaceAlignment; /**< Alignment requirements for surfaces */
int concurrentKernels; /**< Device can possibly execute multiple kernels concurrently */
int ECCEnabled; /**< Device has ECC support enabled */
int pciBusID; /**< PCI bus ID of the device */
int pciDeviceID; /**< PCI device ID of the device */
int pciDomainID; /**< PCI domain ID of the device */
int tccDriver; /**< 1 if device is a Tesla device using TCC driver, 0 otherwise */
int asyncEngineCount; /**< Number of asynchronous engines */
int unifiedAddressing; /**< Device shares a unified address space with the host */
int memoryClockRate; /**< Peak memory clock frequency in kilohertz */
int memoryBusWidth; /**< Global memory bus width in bits */
int l2CacheSize; /**< Size of L2 cache in bytes */
int persistingL2CacheMaxSize; /**< Device's maximum l2 persisting lines capacity setting in bytes */
int maxThreadsPerMultiProcessor;/**< Maximum resident threads per multiprocessor */
int streamPrioritiesSupported; /**< Device supports stream priorities */
int globalL1CacheSupported; /**< Device supports caching globals in L1 */
int localL1CacheSupported; /**< Device supports caching locals in L1 */
size_t sharedMemPerMultiprocessor; /**< Shared memory available per multiprocessor in bytes */
int regsPerMultiprocessor; /**< 32-bit registers available per multiprocessor */
int managedMemory; /**< Device supports allocating managed memory on this system */
int isMultiGpuBoard; /**< Device is on a multi-GPU board */
int multiGpuBoardGroupID; /**< Unique identifier for a group of devices on the same multi-GPU board */
int hostNativeAtomicSupported; /**< Link between the device and the host supports native atomic operations */
int singleToDoublePrecisionPerfRatio; /**< Ratio of single precision performance (in floating-point operations per second) to double precision performance */
int pageableMemoryAccess; /**< Device supports coherently accessing pageable memory without calling cudaHostRegister on it */
int concurrentManagedAccess; /**< Device can coherently access managed memory concurrently with the CPU */
int computePreemptionSupported; /**< Device supports Compute Preemption */
int canUseHostPointerForRegisteredMem; /**< Device can access host registered memory at the same virtual address as the CPU */
int cooperativeLaunch; /**< Device supports launching cooperative kernels via ::cudaLaunchCooperativeKernel */
int cooperativeMultiDeviceLaunch; /**< Deprecated, cudaLaunchCooperativeKernelMultiDevice is deprecated. */
size_t sharedMemPerBlockOptin; /**< Per device maximum shared memory per block usable by special opt in */
int pageableMemoryAccessUsesHostPageTables; /**< Device accesses pageable memory via the host's page tables */
int directManagedMemAccessFromHost; /**< Host can directly access managed memory on the device without migration. */
int maxBlocksPerMultiProcessor; /**< Maximum number of resident blocks per multiprocessor */
int accessPolicyMaxWindowSize; /**< The maximum value of ::cudaAccessPolicyWindow::num_bytes. */
size_t reservedSharedMemPerBlock; /**< Shared memory reserved by CUDA driver per block in bytes */
} cudaDeviceProp_t;
typedef struct cudart_handle {
void *handle;
uint16_t verbose;
int driver_major;
int driver_minor;
cudartReturn_t (*cudaSetDevice)(int device);
cudartReturn_t (*cudaDeviceSynchronize)(void);
cudartReturn_t (*cudaDeviceReset)(void);
cudartReturn_t (*cudaMemGetInfo)(size_t *, size_t *);
cudartReturn_t (*cudaGetDeviceCount)(int *);
cudartReturn_t (*cudaDeviceGetAttribute)(int* value, cudartDeviceAttr_t attr, int device);
cudartReturn_t (*cudaDriverGetVersion) (int *driverVersion);
cudartReturn_t (*cudaGetDeviceProperties) (cudaDeviceProp_t* prop, int device);
} cudart_handle_t;
typedef struct cudart_init_resp {
char *err; // If err is non-null handle is invalid
cudart_handle_t ch;
int num_devices;
} cudart_init_resp_t;
void cudart_init(char *cudart_lib_path, cudart_init_resp_t *resp);
void cudart_bootstrap(cudart_handle_t ch, int device_id, mem_info_t *resp);
// TODO - if we keep this library longer term, add cudart_get_free
void cudart_release(cudart_handle_t ch);
#endif // __GPU_INFO_CUDART_H__
#endif // __APPLE__
discover/gpu_info_nvcuda.c deleted 100644 → 0
View file @ 6b50f2b9
#ifndef __APPLE__ // TODO - maybe consider nvidia support on intel macs?
#include <string.h>
#include <inttypes.h>
#include "gpu_info_nvcuda.h"
void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
LOG(resp->ch.verbose, "initializing %s\n", nvcuda_lib_path);
CUresult ret;
resp->err = NULL;
resp->num_devices = 0;
resp->cudaErr = CUDA_SUCCESS;
const int buflen = 256;
char buf[buflen + 1];
int i;
struct lookup {
char *s;
void **p;
} l[] = {
{"cuInit", (void *)&resp->ch.cuInit},
{"cuDriverGetVersion", (void *)&resp->ch.cuDriverGetVersion},
{"cuDeviceGetCount", (void *)&resp->ch.cuDeviceGetCount},
{"cuDeviceGet", (void *)&resp->ch.cuDeviceGet},
{"cuDeviceGetAttribute", (void *)&resp->ch.cuDeviceGetAttribute},
{"cuDeviceGetUuid", (void *)&resp->ch.cuDeviceGetUuid},
{"cuDeviceGetName", (void *)&resp->ch.cuDeviceGetName},
{"cuCtxCreate_v3", (void *)&resp->ch.cuCtxCreate_v3},
{"cuMemGetInfo_v2", (void *)&resp->ch.cuMemGetInfo_v2},
{"cuCtxDestroy", (void *)&resp->ch.cuCtxDestroy},
{NULL, NULL},
};
resp->ch.handle = LOAD_LIBRARY(nvcuda_lib_path, RTLD_LAZY);
if (!resp->ch.handle) {
char *msg = LOAD_ERR();
LOG(resp->ch.verbose, "library %s load err: %s\n", nvcuda_lib_path, msg);
snprintf(buf, buflen,
"Unable to load %s library to query for Nvidia GPUs: %s",
nvcuda_lib_path, msg);
free(msg);
resp->err = strdup(buf);
resp->cudaErr = -1;
return;
}
for (i = 0; l[i].s != NULL; i++) {
*l[i].p = LOAD_SYMBOL(resp->ch.handle, l[i].s);
if (!*(l[i].p)) {
char *msg = LOAD_ERR();
LOG(resp->ch.verbose, "dlerr: %s\n", msg);
UNLOAD_LIBRARY(resp->ch.handle);
resp->ch.handle = NULL;
snprintf(buf, buflen, "symbol lookup for %s failed: %s", l[i].s,
msg);
free(msg);
resp->err = strdup(buf);
resp->cudaErr = -1;
return;
}
LOG(resp->ch.verbose, "dlsym: %s - %p\n", l[i].s, *l[i].p);
}
LOG(resp->ch.verbose, "calling cuInit\n");
ret = (*resp->ch.cuInit)(0);
if (ret != CUDA_SUCCESS) {
LOG(resp->ch.verbose, "cuInit err: %d\n", ret);
UNLOAD_LIBRARY(resp->ch.handle);
resp->ch.handle = NULL;
snprintf(buf, buflen, "cuda driver library init failure: %d", ret);
resp->err = strdup(buf);
resp->cudaErr = ret;
return;
}
int version = 0;
resp->ch.driver_major = 0;
resp->ch.driver_minor = 0;
// Report driver version if we're in verbose mode, ignore errors
LOG(resp->ch.verbose, "calling cuDriverGetVersion\n");
ret = (*resp->ch.cuDriverGetVersion)(&version);
if (ret != CUDA_SUCCESS) {
LOG(resp->ch.verbose, "cuDriverGetVersion failed: %d\n", ret);
} else {
LOG(resp->ch.verbose, "raw version 0x%x\n", version);
resp->ch.driver_major = version / 1000;
resp->ch.driver_minor = (version - (resp->ch.driver_major * 1000)) / 10;
LOG(resp->ch.verbose, "CUDA driver version: %d.%d\n", resp->ch.driver_major, resp->ch.driver_minor);
}
LOG(resp->ch.verbose, "calling cuDeviceGetCount\n");
ret = (*resp->ch.cuDeviceGetCount)(&resp->num_devices);
if (ret != CUDA_SUCCESS) {
LOG(resp->ch.verbose, "cuDeviceGetCount err: %d\n", ret);
UNLOAD_LIBRARY(resp->ch.handle);
resp->ch.handle = NULL;
snprintf(buf, buflen, "unable to get device count: %d", ret);
resp->err = strdup(buf);
resp->cudaErr = ret;
return;
}
LOG(resp->ch.verbose, "device count %d\n", resp->num_devices);
}
const int buflen = 256;
void nvcuda_bootstrap(nvcuda_handle_t h, int i, mem_info_t *resp) {
resp->err = NULL;
nvcudaMemory_t memInfo = {0,0};
CUresult ret;
CUdevice device = -1;
CUcontext ctx = NULL;
char buf[buflen + 1];
CUuuid uuid = {0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
if (h.handle == NULL) {
resp->err = strdup("cuda driver library handle isn't initialized");
return;
}
ret = (*h.cuDeviceGet)(&device, i);
if (ret != CUDA_SUCCESS) {
snprintf(buf, buflen, "cuda driver library device failed to initialize");
resp->err = strdup(buf);
return;
}
int major = 0;
int minor = 0;
ret = (*h.cuDeviceGetAttribute)(&major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, device);
if (ret != CUDA_SUCCESS) {
LOG(h.verbose, "[%d] device major lookup failure: %d\n", i, ret);
} else {
ret = (*h.cuDeviceGetAttribute)(&minor, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, device);
if (ret != CUDA_SUCCESS) {
LOG(h.verbose, "[%d] device minor lookup failure: %d\n", i, ret);
} else {
resp->minor = minor;
resp->major = major;
}
}
ret = (*h.cuDeviceGetUuid)(&uuid, device);
if (ret != CUDA_SUCCESS) {
LOG(h.verbose, "[%d] device uuid lookup failure: %d\n", i, ret);
snprintf(&resp->gpu_id[0], GPU_ID_LEN, "%d", i);
} else {
// GPU-d110a105-ac29-1d54-7b49-9c90440f215b
snprintf(&resp->gpu_id[0], GPU_ID_LEN,
"GPU-%02x%02x%02x%02x-%02x%02x-%02x%02x-%02x%02x-%02x%02x%02x%02x%02x%02x",
uuid.bytes[0],
uuid.bytes[1],
uuid.bytes[2],
uuid.bytes[3],
uuid.bytes[4],
uuid.bytes[5],
uuid.bytes[6],
uuid.bytes[7],
uuid.bytes[8],
uuid.bytes[9],
uuid.bytes[10],
uuid.bytes[11],
uuid.bytes[12],
uuid.bytes[13],
uuid.bytes[14],
uuid.bytes[15]
);
}
ret = (*h.cuDeviceGetName)(&resp->gpu_name[0], GPU_NAME_LEN, device);
if (ret != CUDA_SUCCESS) {
LOG(h.verbose, "[%d] device name lookup failure: %d\n", i, ret);
resp->gpu_name[0] = '\0';
}
// To get memory we have to set (and release) a context
ret = (*h.cuCtxCreate_v3)(&ctx, NULL, 0, 0, device);
if (ret != CUDA_SUCCESS) {
snprintf(buf, buflen, "cuda driver library failed to get device context %d", ret);
resp->err = strdup(buf);
return;
}
ret = (*h.cuMemGetInfo_v2)(&memInfo.free, &memInfo.total);
if (ret != CUDA_SUCCESS) {
snprintf(buf, buflen, "cuda driver library device memory info lookup failure %d", ret);
resp->err = strdup(buf);
// Best effort on failure...
(*h.cuCtxDestroy)(ctx);
return;
}
resp->total = memInfo.total;
resp->free = memInfo.free;
LOG(h.verbose, "[%s] CUDA totalMem %" PRId64 "mb\n", resp->gpu_id, resp->total / 1024 / 1024);
LOG(h.verbose, "[%s] CUDA freeMem %" PRId64 "mb\n", resp->gpu_id, resp->free / 1024 / 1024);
LOG(h.verbose, "[%s] Compute Capability %d.%d\n", resp->gpu_id, resp->major, resp->minor);
ret = (*h.cuCtxDestroy)(ctx);
if (ret != CUDA_SUCCESS) {
LOG(1, "cuda driver library failed to release device context %d", ret);
}
}
void nvcuda_get_free(nvcuda_handle_t h, int i, uint64_t *free, uint64_t *total) {
CUresult ret;
CUcontext ctx = NULL;
CUdevice device = -1;
*free = 0;
*total = 0;
ret = (*h.cuDeviceGet)(&device, i);
if (ret != CUDA_SUCCESS) {
LOG(1, "cuda driver library device failed to initialize");
return;
}
// To get memory we have to set (and release) a context
ret = (*h.cuCtxCreate_v3)(&ctx, NULL, 0, 0, device);
if (ret != CUDA_SUCCESS) {
LOG(1, "cuda driver library failed to get device context %d", ret);
return;
}
ret = (*h.cuMemGetInfo_v2)(free, total);
if (ret != CUDA_SUCCESS) {
LOG(1, "cuda driver library device memory info lookup failure %d", ret);
// Best effort on failure...
(*h.cuCtxDestroy)(ctx);
return;
}
ret = (*h.cuCtxDestroy)(ctx);
if (ret != CUDA_SUCCESS) {
LOG(1, "cuda driver library failed to release device context %d", ret);
}
}
void nvcuda_release(nvcuda_handle_t h) {
LOG(h.verbose, "releasing cuda driver library\n");
UNLOAD_LIBRARY(h.handle);
// TODO and other context release logic?
h.handle = NULL;
}
#endif // __APPLE__
discover/gpu_info_nvcuda.h deleted 100644 → 0
View file @ 6b50f2b9
#ifndef __APPLE__
#ifndef __GPU_INFO_NVCUDA_H__
#define __GPU_INFO_NVCUDA_H__
#include "gpu_info.h"
// Just enough typedef's to dlopen/dlsym for memory information
typedef enum cudaError_enum {
CUDA_SUCCESS = 0,
CUDA_ERROR_INVALID_VALUE = 1,
CUDA_ERROR_OUT_OF_MEMORY = 2,
CUDA_ERROR_NOT_INITIALIZED = 3,
CUDA_ERROR_INSUFFICIENT_DRIVER = 35,
CUDA_ERROR_NO_DEVICE = 100,
CUDA_ERROR_SYSTEM_DRIVER_MISMATCH = 803,
CUDA_ERROR_UNKNOWN = 999,
// Other values omitted for now...
} CUresult;
typedef enum CUdevice_attribute_enum {
CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR = 75,
CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR = 76,
// TODO - not yet wired up but may be useful for Jetson or other
// integrated GPU scenarios with shared memory
CU_DEVICE_ATTRIBUTE_INTEGRATED = 18
} CUdevice_attribute;
typedef void *nvcudaDevice_t; // Opaque is sufficient
typedef struct nvcudaMemory_st {
uint64_t total;
uint64_t free;
} nvcudaMemory_t;
typedef struct nvcudaDriverVersion {
int major;
int minor;
} nvcudaDriverVersion_t;
typedef struct CUuuid_st {
unsigned char bytes[16];
} CUuuid;
typedef int CUdevice;
typedef void* CUcontext;
typedef struct nvcuda_handle {
void *handle;
uint16_t verbose;
int driver_major;
int driver_minor;
CUresult (*cuInit)(unsigned int Flags);
CUresult (*cuDriverGetVersion)(int *driverVersion);
CUresult (*cuDeviceGetCount)(int *);
CUresult (*cuDeviceGet)(CUdevice* device, int ordinal);
CUresult (*cuDeviceGetAttribute)(int* pi, CUdevice_attribute attrib, CUdevice dev);
CUresult (*cuDeviceGetUuid)(CUuuid* uuid, CUdevice dev); // signature compatible with cuDeviceGetUuid_v2
CUresult (*cuDeviceGetName)(char *name, int len, CUdevice dev);
// Context specific aspects
CUresult (*cuCtxCreate_v3)(CUcontext* pctx, void *params, int len, unsigned int flags, CUdevice dev);
CUresult (*cuMemGetInfo_v2)(uint64_t* free, uint64_t* total);
CUresult (*cuCtxDestroy)(CUcontext ctx);
} nvcuda_handle_t;
typedef struct nvcuda_init_resp {
char *err; // If err is non-null handle is invalid
nvcuda_handle_t ch;
int num_devices;
CUresult cudaErr;
} nvcuda_init_resp_t;
void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp);
void nvcuda_bootstrap(nvcuda_handle_t ch, int device_id, mem_info_t *resp);
void nvcuda_get_free(nvcuda_handle_t ch, int device_id, uint64_t *free, uint64_t *total);
void nvcuda_release(nvcuda_handle_t ch);
#endif // __GPU_INFO_NVCUDA_H__
#endif // __APPLE__
discover/gpu_info_nvml.c deleted 100644 → 0
View file @ 6b50f2b9
#ifndef __APPLE__ // TODO - maybe consider nvidia support on intel macs?
#include <string.h>
#include "gpu_info_nvml.h"
void nvml_init(char *nvml_lib_path, nvml_init_resp_t *resp) {
nvmlReturn_t ret;
resp->err = NULL;
const int buflen = 256;
char buf[buflen + 1];
int i;
struct lookup {
char *s;
void **p;
} l[] = {
{"nvmlInit_v2", (void *)&resp->ch.nvmlInit_v2},
{"nvmlShutdown", (void *)&resp->ch.nvmlShutdown},
{"nvmlDeviceGetHandleByUUID", (void *)&resp->ch.nvmlDeviceGetHandleByUUID},
{"nvmlDeviceGetMemoryInfo", (void *)&resp->ch.nvmlDeviceGetMemoryInfo},
{NULL, NULL},
};
resp->ch.handle = LOAD_LIBRARY(nvml_lib_path, RTLD_LAZY);
if (!resp->ch.handle) {
char *msg = LOAD_ERR();
LOG(resp->ch.verbose, "library %s load err: %s\n", nvml_lib_path, msg);
snprintf(buf, buflen,
"Unable to load %s library to query for Nvidia GPUs: %s",
nvml_lib_path, msg);
free(msg);
resp->err = strdup(buf);
return;
}
// TODO once we've squashed the remaining corner cases remove this log
// LOG(resp->ch.verbose, "wiring nvidia management library functions in %s\n", nvml_lib_path);
for (i = 0; l[i].s != NULL; i++) {
// TODO once we've squashed the remaining corner cases remove this log
// LOG(resp->ch.verbose, "dlsym: %s\n", l[i].s);
*l[i].p = LOAD_SYMBOL(resp->ch.handle, l[i].s);
if (!*(l[i].p)) {
resp->ch.handle = NULL;
char *msg = LOAD_ERR();
LOG(resp->ch.verbose, "dlerr: %s\n", msg);
UNLOAD_LIBRARY(resp->ch.handle);
snprintf(buf, buflen, "symbol lookup for %s failed: %s", l[i].s,
msg);
free(msg);
resp->err = strdup(buf);
return;
}
}
ret = (*resp->ch.nvmlInit_v2)();
if (ret != NVML_SUCCESS) {
LOG(resp->ch.verbose, "nvmlInit_v2 err: %d\n", ret);
UNLOAD_LIBRARY(resp->ch.handle);
resp->ch.handle = NULL;
snprintf(buf, buflen, "nvml vram init failure: %d", ret);
resp->err = strdup(buf);
return;
}
}
void nvml_get_free(nvml_handle_t h, char *uuid, uint64_t *free, uint64_t *total, uint64_t *used) {
nvmlDevice_t device;
nvmlMemory_t memInfo = {0};
nvmlReturn_t ret;
ret = (*h.nvmlDeviceGetHandleByUUID)((const char *)(uuid), &device);
if (ret != NVML_SUCCESS) {
LOG(1, "unable to get device handle %s: %d", uuid, ret);
*free = 0;
return;
}
ret = (*h.nvmlDeviceGetMemoryInfo)(device, &memInfo);
if (ret != NVML_SUCCESS) {
LOG(1, "device memory info lookup failure %s: %d", uuid, ret);
*free = 0;
return;
}
*free = memInfo.free;
*total = memInfo.total;
*used = memInfo.used;
}
void nvml_release(nvml_handle_t h) {
LOG(h.verbose, "releasing nvml library\n");
nvmlReturn_t ret;
ret = (*h.nvmlShutdown)();
if (ret != NVML_SUCCESS) {
LOG(1, "error during nvmlShutdown %d", ret);
}
UNLOAD_LIBRARY(h.handle);
h.handle = NULL;
}
#endif // __APPLE__
\ No newline at end of file
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment