DRY out the runner lifecycle code (#12540)

* DRY out the runner lifecycle code

Now that discovery uses the runners as well, this unifies the runner spawning code
into a single place.  This also unifies GPU discovery types with the newer ml.DeviceInfo

* win: make incremental builds better

Place build artifacts in discrete directories so incremental builds don't have to start fresh

* Adjust sort order to consider iGPUs

* handle cpu inference oom scenarios

* review comments

discover/gpu.go

@@ -1,16 +1,13 @@
 package discover
 
 import (
-	"context"
 	"log/slog"
 	"os"
-	"path/filepath"
 	"regexp"
 	"runtime"
 	"strconv"
 	"strings"
 
-	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/ml"
 )
 
@@ -18,159 +15,28 @@ import (
 // Included to drive logic for reducing Ollama-allocated overhead on L4T/Jetson devices.
 var CudaTegra string = os.Getenv("JETSON_JETPACK")
 
-func GetCPUInfo() GpuInfo {
-	mem, err := GetCPUMem()
+// GetSystemInfo returns the last cached state of the GPUs on the system
+func GetSystemInfo() ml.SystemInfo {
+	memInfo, err := GetCPUMem()
 	if err != nil {
 		slog.Warn("error looking up system memory", "error", err)
 	}
-
-	return GpuInfo{
-		memInfo: mem,
-		DeviceID: ml.DeviceID{
-			Library: "cpu",
-			ID:      "0",
-		},
-	}
-}
-
-func GetGPUInfo(ctx context.Context, runners []FilteredRunnerDiscovery) GpuInfoList {
-	devs := GPUDevices(ctx, runners)
-	return devInfoToInfoList(devs)
-}
-
-func devInfoToInfoList(devs []ml.DeviceInfo) GpuInfoList {
-	resp := []GpuInfo{}
-	// Our current packaging model places ggml-hip in the main directory
-	// but keeps rocm in an isolated directory.  We have to add it to
-	// the [LD_LIBRARY_]PATH so ggml-hip will load properly
-	rocmDir := filepath.Join(LibOllamaPath, "rocm")
-	if _, err := os.Stat(rocmDir); err != nil {
-		rocmDir = ""
+	var threadCount int
+	cpus := GetCPUDetails()
+	for _, c := range cpus {
+		threadCount += c.CoreCount - c.EfficiencyCoreCount
 	}
 
-	for _, dev := range devs {
-		info := GpuInfo{
-			DeviceID: dev.DeviceID,
-			filterID: dev.FilteredID,
-			Name:     dev.Description,
-			memInfo: memInfo{
-				TotalMemory: dev.TotalMemory,
-				FreeMemory:  dev.FreeMemory,
-			},
-			// TODO can we avoid variant
-			DependencyPath: dev.LibraryPath,
-			DriverMajor:    dev.DriverMajor,
-			DriverMinor:    dev.DriverMinor,
-			ComputeMajor:   dev.ComputeMajor,
-			ComputeMinor:   dev.ComputeMinor,
-		}
-		if dev.Library == "CUDA" || dev.Library == "ROCm" {
-			info.MinimumMemory = 457 * format.MebiByte
-		}
-		if dev.Library == "ROCm" && rocmDir != "" {
-			info.DependencyPath = append(info.DependencyPath, rocmDir)
-		}
-		// TODO any special processing of Vulkan devices?
-		resp = append(resp, info)
-	}
-	if len(resp) == 0 {
-		mem, err := GetCPUMem()
-		if err != nil {
-			slog.Warn("error looking up system memory", "error", err)
-		}
-
-		resp = append(resp, GpuInfo{
-			memInfo: mem,
-			DeviceID: ml.DeviceID{
-				Library: "cpu",
-				ID:      "0",
-			},
-		})
-	}
-	return resp
-}
-
-// Given the list of GPUs this instantiation is targeted for,
-// figure out the visible devices environment variable
-//
-// If different libraries are detected, the first one is what we use
-func (l GpuInfoList) GetVisibleDevicesEnv() []string {
-	if len(l) == 0 {
-		return nil
-	}
-	res := []string{}
-	envVar := rocmGetVisibleDevicesEnv(l)
-	if envVar != "" {
-		res = append(res, envVar)
-	}
-	envVar = vkGetVisibleDevicesEnv(l)
-	if envVar != "" {
-		res = append(res, envVar)
-	}
-	return res
-}
-
-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 info.filterID != "" {
-			ids = append(ids, info.filterID)
-		} else {
-			ids = append(ids, info.ID)
-		}
-	}
-	if len(ids) == 0 {
-		return ""
-	}
-	envVar := "ROCR_VISIBLE_DEVICES="
-	if runtime.GOOS != "linux" {
-		envVar = "HIP_VISIBLE_DEVICES="
-	}
-	// 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 envVar + strings.Join(ids, ",")
-}
-
-func vkGetVisibleDevicesEnv(gpuInfo []GpuInfo) string {
-	ids := []string{}
-	for _, info := range gpuInfo {
-		if info.Library != "Vulkan" {
-			continue
-		}
-		if info.filterID != "" {
-			ids = append(ids, info.filterID)
-		} else {
-			ids = append(ids, info.ID)
-		}
-	}
-	if len(ids) == 0 {
-		return ""
-	}
-	envVar := "GGML_VK_VISIBLE_DEVICES="
-	return envVar + strings.Join(ids, ",")
-}
-
-// GetSystemInfo returns the last cached state of the GPUs on the system
-func GetSystemInfo() SystemInfo {
-	deviceMu.Lock()
-	defer deviceMu.Unlock()
-	gpus := devInfoToInfoList(devices)
-	if len(gpus) == 1 && gpus[0].Library == "cpu" {
-		gpus = []GpuInfo{}
+	if threadCount == 0 {
+		// Fall back to Go's num CPU
+		threadCount = runtime.NumCPU()
 	}
 
-	return SystemInfo{
-		System: CPUInfo{
-			CPUs:    GetCPUDetails(),
-			GpuInfo: GetCPUInfo(),
-		},
-		GPUs: gpus,
+	return ml.SystemInfo{
+		ThreadCount: threadCount,
+		TotalMemory: memInfo.TotalMemory,
+		FreeMemory:  memInfo.FreeMemory,
+		FreeSwap:    memInfo.FreeSwap,
 	}
 }