This commit is contained in:
Michael Yang
2024-08-01 14:52:15 -07:00
parent 558a54b098
commit b732beba6a
68 changed files with 199 additions and 149 deletions

View File

@@ -3,7 +3,7 @@
package gpu
import (
"fmt"
"errors"
"log/slog"
"os"
"path/filepath"
@@ -95,5 +95,5 @@ func commonAMDValidateLibDir() (string, error) {
}
}
return "", fmt.Errorf("no suitable rocm found, falling back to CPU")
return "", errors.New("no suitable rocm found, falling back to CPU")
}

View File

@@ -1,6 +1,7 @@
package gpu
import (
"errors"
"fmt"
"log/slog"
"syscall"
@@ -76,7 +77,7 @@ func (hl *HipLib) Release() {
func (hl *HipLib) AMDDriverVersion() (driverMajor, driverMinor int, err error) {
if hl.dll == 0 {
return 0, 0, fmt.Errorf("dll has been unloaded")
return 0, 0, errors.New("dll has been unloaded")
}
var version int
status, _, err := syscall.SyscallN(hl.hipDriverGetVersion, uintptr(unsafe.Pointer(&version)))
@@ -110,7 +111,7 @@ func (hl *HipLib) HipGetDeviceCount() int {
func (hl *HipLib) HipSetDevice(device int) error {
if hl.dll == 0 {
return fmt.Errorf("dll has been unloaded")
return errors.New("dll has been unloaded")
}
status, _, err := syscall.SyscallN(hl.hipSetDevice, uintptr(device))
if status != hipSuccess {
@@ -121,7 +122,7 @@ func (hl *HipLib) HipSetDevice(device int) error {
func (hl *HipLib) HipGetDeviceProperties(device int) (*hipDevicePropMinimal, error) {
if hl.dll == 0 {
return nil, fmt.Errorf("dll has been unloaded")
return nil, errors.New("dll has been unloaded")
}
var props hipDevicePropMinimal
status, _, err := syscall.SyscallN(hl.hipGetDeviceProperties, uintptr(unsafe.Pointer(&props)), uintptr(device))
@@ -134,7 +135,7 @@ func (hl *HipLib) HipGetDeviceProperties(device int) (*hipDevicePropMinimal, err
// free, total, err
func (hl *HipLib) HipMemGetInfo() (uint64, uint64, error) {
if hl.dll == 0 {
return 0, 0, fmt.Errorf("dll has been unloaded")
return 0, 0, errors.New("dll has been unloaded")
}
var totalMemory uint64
var freeMemory uint64

View File

@@ -393,7 +393,7 @@ func AMDValidateLibDir() (string, error) {
// 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 "", fmt.Errorf("no suitable rocm found, falling back to CPU")
return "", errors.New("no suitable rocm found, falling back to CPU")
}
func AMDDriverVersion() (driverMajor, driverMinor int, err error) {

View File

@@ -2,7 +2,7 @@ package gpu
import (
"bytes"
"fmt"
"errors"
"log/slog"
"os"
"path/filepath"
@@ -85,7 +85,7 @@ func AMDGetGPUInfo() []RocmGPUInfo {
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
// slog.Info(fmt.Sprintf("[%d] Integrated: %d", i, props.iGPU)) // DOESN'T REPORT CORRECTLY! Always 0
// TODO Why isn't props.iGPU accurate!?
if strings.EqualFold(name, iGPUName) {
slog.Info("unsupported Radeon iGPU detected skipping", "id", i, "name", name, "gfx", gfx)
@@ -161,7 +161,7 @@ func AMDValidateLibDir() (string, error) {
// 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 "", fmt.Errorf("no suitable rocm found, falling back to CPU")
return "", errors.New("no suitable rocm found, falling back to CPU")
}
func (gpus RocmGPUInfoList) RefreshFreeMemory() error {

View File

@@ -42,7 +42,7 @@ func PayloadsDir() (string, error) {
return "", fmt.Errorf("failed to generate tmp dir: %w", err)
}
} else {
err = os.MkdirAll(tmpDir, 0755)
err = os.MkdirAll(tmpDir, 0o755)
if err != nil {
return "", fmt.Errorf("failed to generate tmp dir %s: %w", tmpDir, err)
}
@@ -54,7 +54,7 @@ func PayloadsDir() (string, error) {
if err != nil {
return "", err
}
if _, err := pidFile.Write([]byte(fmt.Sprint(os.Getpid()))); err != nil {
if _, err := pidFile.Write([]byte(strconv.Itoa(os.Getpid()))); err != nil {
return "", err
}

View File

@@ -7,9 +7,9 @@ package gpu
#cgo windows LDFLAGS: -lpthread
#include "gpu_info.h"
*/
import "C"
import (
"fmt"
"log/slog"
@@ -70,7 +70,6 @@ var CudaTegra string = os.Getenv("JETSON_JETPACK")
// Note: gpuMutex must already be held
func initCudaHandles() *cudaHandles {
// TODO - if the ollama build is CPU only, don't do these checks as they're irrelevant and confusing
cHandles := &cudaHandles{}
@@ -211,14 +210,16 @@ func GetGPUInfo() GpuInfoList {
if err != nil {
slog.Warn("error looking up system memory", "error", err)
}
cpus = []CPUInfo{CPUInfo{
GpuInfo: GpuInfo{
memInfo: mem,
Library: "cpu",
Variant: cpuCapability,
ID: "0",
cpus = []CPUInfo{
{
GpuInfo: GpuInfo{
memInfo: mem,
Library: "cpu",
Variant: cpuCapability,
ID: "0",
},
},
}}
}
// Fallback to CPU mode if we're lacking required vector extensions on x86
if cpuCapability < GPURunnerCPUCapability && runtime.GOARCH == "amd64" {

View File

@@ -8,6 +8,7 @@ package gpu
#include "gpu_info_darwin.h"
*/
import "C"
import (
"runtime"

View File

@@ -67,4 +67,4 @@ void cpu_check_ram(mem_info_t *resp);
#include "gpu_info_oneapi.h"
#endif // __GPU_INFO_H__
#endif // __APPLE__
#endif // __APPLE__

View File

@@ -43,10 +43,12 @@ var OneapiGlobs = []string{
"/usr/lib*/libze_intel_gpu.so*",
}
var CudartMgmtName = "libcudart.so*"
var NvcudaMgmtName = "libcuda.so*"
var NvmlMgmtName = "" // not currently wired on linux
var OneapiMgmtName = "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

View File

@@ -40,10 +40,12 @@ var OneapiGlobs = []string{
"c:\\Windows\\System32\\DriverStore\\FileRepository\\*\\ze_intel_gpu64.dll",
}
var CudartMgmtName = "cudart64_*.dll"
var NvcudaMgmtName = "nvcuda.dll"
var NvmlMgmtName = "nvml.dll"
var OneapiMgmtName = "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}