Use correct url for auto updates

fix: offloading on low end GPUs

api/types.go

@@ -3,7 +3,6 @@ package api
 import (
 	"encoding/json"
 	"fmt"
-	"log"
 	"math"
 	"os"
 	"reflect"
@@ -238,44 +237,39 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
 						// when JSON unmarshals numbers, it uses float64, not int
 						field.SetInt(int64(t))
 					default:
-						log.Printf("could not convert model parameter %v of type %T to int, skipped", key, val)
+						return fmt.Errorf("option %q must be of type integer", key)
 					}
 				case reflect.Bool:
 					val, ok := val.(bool)
 					if !ok {
-						log.Printf("could not convert model parameter %v of type %T to bool, skipped", key, val)
-						continue
+						return fmt.Errorf("option %q must be of type boolean", key)
 					}
 					field.SetBool(val)
 				case reflect.Float32:
 					// JSON unmarshals to float64
 					val, ok := val.(float64)
 					if !ok {
-						log.Printf("could not convert model parameter %v of type %T to float32, skipped", key, val)
-						continue
+						return fmt.Errorf("option %q must be of type float32", key)
 					}
 					field.SetFloat(val)
 				case reflect.String:
 					val, ok := val.(string)
 					if !ok {
-						log.Printf("could not convert model parameter %v of type %T to string, skipped", key, val)
-						continue
+						return fmt.Errorf("option %q must be of type string", key)
 					}
 					field.SetString(val)
 				case reflect.Slice:
 					// JSON unmarshals to []interface{}, not []string
 					val, ok := val.([]interface{})
 					if !ok {
-						log.Printf("could not convert model parameter %v of type %T to slice, skipped", key, val)
-						continue
+						return fmt.Errorf("option %q must be of type array", key)
 					}
 					// convert []interface{} to []string
 					slice := make([]string, len(val))
 					for i, item := range val {
 						str, ok := item.(string)
 						if !ok {
-							log.Printf("could not convert model parameter %v of type %T to slice of strings, skipped", key, item)
-							continue
+							return fmt.Errorf("option %q must be of an array of strings", key)
 						}
 						slice[i] = str
 					}

app/src/index.ts

@@ -162,13 +162,56 @@ app.on('before-quit', () => {
   }
 })
 
+const updateURL = `https://ollama.ai/api/update?os=${process.platform}&arch=${
+  process.arch
+}&version=${app.getVersion()}&id=${id()}`
+
+let latest = ''
+async function isNewReleaseAvailable() {
+  try {
+    const response = await fetch(updateURL)
+
+    if (!response.ok) {
+      return false
+    }
+
+    if (response.status === 204) {
+      return false
+    }
+
+    const data = await response.json()
+
+    const url = data?.url
+    if (!url) {
+      return false
+    }
+
+    if (latest === url) {
+      return false
+    }
+
+    latest = url
+
+    return true
+  } catch (error) {
+    logger.error(`update check failed - ${error}`)
+    return false
+  }
+}
+
+async function checkUpdate() {
+  const available = await isNewReleaseAvailable()
+  if (available) {
+    logger.info('checking for update')
+    autoUpdater.checkForUpdates()
+  }
+}
+
 function init() {
   if (app.isPackaged) {
-    autoUpdater.checkForUpdates()
+    checkUpdate()
     setInterval(() => {
-      if (!updateAvailable) {
-        autoUpdater.checkForUpdates()
-      }
+      checkUpdate()
     }, 60 * 60 * 1000)
   }
 
@@ -246,11 +289,7 @@ function id(): string {
   return uuid
 }
 
-autoUpdater.setFeedURL({
-  url: `https://ollama.ai/api/update?os=${process.platform}&arch=${
-    process.arch
-  }&version=${app.getVersion()}&id=${id()}`,
-})
+autoUpdater.setFeedURL({ url: updateURL })
 
 autoUpdater.on('error', e => {
   logger.error(`update check failed - ${e.message}`)

llm/llama.go

@@ -30,42 +30,43 @@ import (
 var llamaCppEmbed embed.FS
 
 type ModelRunner struct {
-	Path string // path to the model runner executable
+	Path        string // path to the model runner executable
+	Accelerated bool
 }
 
 func chooseRunners(workDir, runnerType string) []ModelRunner {
 	buildPath := path.Join("llama.cpp", runnerType, "build")
-	var runners []string
+	var runners []ModelRunner
 
 	// set the runners based on the OS
 	// IMPORTANT: the order of the runners in the array is the priority order
 	switch runtime.GOOS {
 	case "darwin":
-		runners = []string{
-			path.Join(buildPath, "metal", "bin", "ollama-runner"),
-			path.Join(buildPath, "cpu", "bin", "ollama-runner"),
+		runners = []ModelRunner{
+			{Path: path.Join(buildPath, "metal", "bin", "ollama-runner")},
+			{Path: path.Join(buildPath, "cpu", "bin", "ollama-runner")},
 		}
 	case "linux":
-		runners = []string{
-			path.Join(buildPath, "cuda", "bin", "ollama-runner"),
-			path.Join(buildPath, "cpu", "bin", "ollama-runner"),
+		runners = []ModelRunner{
+			{Path: path.Join(buildPath, "cuda", "bin", "ollama-runner"), Accelerated: true},
+			{Path: path.Join(buildPath, "cpu", "bin", "ollama-runner")},
 		}
 	case "windows":
 		// TODO: select windows GPU runner here when available
-		runners = []string{
-			path.Join(buildPath, "cpu", "bin", "Release", "ollama-runner.exe"),
+		runners = []ModelRunner{
+			{Path: path.Join(buildPath, "cpu", "bin", "Release", "ollama-runner.exe")},
 		}
 	default:
 		log.Printf("unknown OS, running on CPU: %s", runtime.GOOS)
-		runners = []string{
-			path.Join(buildPath, "cpu", "bin", "ollama-runner"),
+		runners = []ModelRunner{
+			{Path: path.Join(buildPath, "cpu", "bin", "ollama-runner")},
 		}
 	}
 
 	runnerAvailable := false // if no runner files are found in the embed, this flag will cause a fast fail
 	for _, r := range runners {
 		// find all the files in the runner's bin directory
-		files, err := fs.Glob(llamaCppEmbed, path.Join(path.Dir(r), "*"))
+		files, err := fs.Glob(llamaCppEmbed, path.Join(path.Dir(r.Path), "*"))
 		if err != nil {
 			// this is expected, ollama may be compiled without all runners packed in
 			log.Printf("%s runner not found: %v", r, err)
@@ -115,7 +116,10 @@ func chooseRunners(workDir, runnerType string) []ModelRunner {
 	localRunnersByPriority := []ModelRunner{}
 	for _, r := range runners {
 		// clean the ModelRunner paths so that they match the OS we are running on
-		localRunnersByPriority = append(localRunnersByPriority, ModelRunner{Path: filepath.Clean(path.Join(workDir, r))})
+		localRunnersByPriority = append(localRunnersByPriority, ModelRunner{
+			Path:        filepath.Clean(path.Join(workDir, r.Path)),
+			Accelerated: r.Accelerated,
+		})
 	}
 
 	return localRunnersByPriority
@@ -215,6 +219,11 @@ func CheckVRAM() (int64, error) {
 		free += vram
 	}
 
+	if free*1024*1024 < 2*1000*1000*1000 {
+		log.Printf("less than 2 GB VRAM available, falling back to CPU only")
+		free = 0
+	}
+
 	return free, nil
 }
 
@@ -238,8 +247,8 @@ func NumGPU(numLayer, fileSizeBytes int64, opts api.Options) int {
 		// TODO: this is a rough heuristic, better would be to calculate this based on number of layers and context size
 		bytesPerLayer := fileSizeBytes / numLayer
 
-		// max number of layers we can fit in VRAM, subtract 5% to prevent consuming all available VRAM and running out of memory
-		layers := int(freeVramBytes/bytesPerLayer) * 95 / 100
+		// max number of layers we can fit in VRAM, subtract 8% to prevent consuming all available VRAM and running out of memory
+		layers := int(freeVramBytes/bytesPerLayer) * 92 / 100
 		log.Printf("%d MiB VRAM available, loading up to %d GPU layers", vramMib, layers)
 
 		return layers
@@ -261,8 +270,7 @@ func NewStatusWriter() *StatusWriter {
 
 func (w *StatusWriter) Write(b []byte) (int, error) {
 	if _, after, ok := bytes.Cut(b, []byte("error:")); ok {
-		err := fmt.Errorf("llama runner: %s", after)
-		w.ErrCh <- err
+		w.ErrCh <- fmt.Errorf("llama runner: %s", bytes.TrimSpace(after))
 	}
 	return os.Stderr.Write(b)
 }
@@ -277,16 +285,20 @@ func newLlama(model string, adapters []string, runners []ModelRunner, numLayers
 		return nil, errors.New("ollama supports only one lora adapter, but multiple were provided")
 	}
 
+	numGPU := NumGPU(numLayers, fileInfo.Size(), opts)
 	params := []string{
 		"--model", model,
 		"--ctx-size", fmt.Sprintf("%d", opts.NumCtx),
 		"--rope-freq-base", fmt.Sprintf("%f", opts.RopeFrequencyBase),
 		"--rope-freq-scale", fmt.Sprintf("%f", opts.RopeFrequencyScale),
 		"--batch-size", fmt.Sprintf("%d", opts.NumBatch),
-		"--n-gpu-layers", fmt.Sprintf("%d", NumGPU(numLayers, fileInfo.Size(), opts)),
 		"--embedding",
 	}
 
+	if numGPU > 0 {
+		params = append(params, "--n-gpu-layers", fmt.Sprintf("%d", numGPU))
+	}
+
 	if opts.NumGQA > 0 {
 		params = append(params, "--gqa", fmt.Sprintf("%d", opts.NumGQA))
 	}
@@ -317,6 +329,11 @@ func newLlama(model string, adapters []string, runners []ModelRunner, numLayers
 
 	// start the llama.cpp server with a retry in case the port is already in use
 	for _, runner := range runners {
+		if runner.Accelerated && numGPU == 0 {
+			log.Printf("skipping accelerated runner because num_gpu=0")
+			continue
+		}
+
 		if _, err := os.Stat(runner.Path); err != nil {
 			log.Printf("llama runner not found: %v", err)
 			continue