remove tmp directories created by previous servers (#559)

* remove tmp directories created by previous servers

* clean up on server stop

* Update routes.go

* Update server/routes.go

Co-authored-by: Jeffrey Morgan <jmorganca@gmail.com>

* create top-level temp ollama dir

* check file exists before creating

---------

Co-authored-by: Jeffrey Morgan <jmorganca@gmail.com>
Co-authored-by: Michael Yang <mxyng@pm.me>

llm/ggml.go

@@ -4,7 +4,6 @@ import (
 	"encoding/binary"
 	"errors"
 	"io"
-	"sync"
 )
 
 type GGML struct {
@@ -165,18 +164,6 @@ func (c *containerLORA) Decode(r io.Reader) (model, error) {
 	return nil, nil
 }
 
-var (
-	ggmlInit    sync.Once
-	ggmlRunners []ModelRunner // a slice of ModelRunners ordered by priority
-)
-
-func ggmlRunner() []ModelRunner {
-	ggmlInit.Do(func() {
-		ggmlRunners = chooseRunners("ggml")
-	})
-	return ggmlRunners
-}
-
 const (
 	// Magic constant for `ggml` files (unversioned).
 	FILE_MAGIC_GGML = 0x67676d6c

llm/gguf.go

@@ -6,7 +6,6 @@ import (
 	"errors"
 	"fmt"
 	"io"
-	"sync"
 )
 
 type containerGGUF struct {
@@ -368,16 +367,3 @@ func (llm *ggufModel) readArray(r io.Reader) (arr []any, err error) {
 
 	return
 }
-
-var (
-	ggufInit    sync.Once
-	ggufRunners []ModelRunner // a slice of ModelRunners ordered by priority
-)
-
-func ggufRunner() []ModelRunner {
-	ggufInit.Do(func() {
-		ggufRunners = chooseRunners("gguf")
-	})
-
-	return ggufRunners
-}

llm/llama.go

@@ -32,7 +32,7 @@ type ModelRunner struct {
 	Path string // path to the model runner executable
 }
 
-func chooseRunners(runnerType string) []ModelRunner {
+func chooseRunners(workDir, runnerType string) []ModelRunner {
 	buildPath := path.Join("llama.cpp", runnerType, "build")
 	var runners []string
 
@@ -61,11 +61,6 @@ func chooseRunners(runnerType string) []ModelRunner {
 		}
 	}
 
-	// copy the files locally to run the llama.cpp server
-	tmpDir, err := os.MkdirTemp("", "llama-*")
-	if err != nil {
-		log.Fatalf("load llama runner: failed to create temp dir: %v", err)
-	}
 	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
@@ -85,11 +80,17 @@ func chooseRunners(runnerType string) []ModelRunner {
 			defer srcFile.Close()
 
 			// create the directory in case it does not exist
-			destPath := filepath.Join(tmpDir, filepath.Dir(f))
+			destPath := filepath.Join(workDir, filepath.Dir(f))
 			if err := os.MkdirAll(destPath, 0o755); err != nil {
 				log.Fatalf("create runner temp dir %s: %v", filepath.Dir(f), err)
 			}
-			destFile, err := os.OpenFile(filepath.Join(destPath, filepath.Base(f)), os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o755)
+
+			destFile := filepath.Join(destPath, filepath.Base(f))
+
+			_, 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 {
 					log.Fatalf("write llama runner %s: %v", f, err)
 				}
@@ -98,6 +99,9 @@ func chooseRunners(runnerType string) []ModelRunner {
 				if _, err := io.Copy(destFile, srcFile); err != nil {
 					log.Fatalf("copy llama runner %s: %v", f, err)
 				}
+			case err != nil:
+				log.Fatalf("stat llama runner %s: %v", f, err)
+			}
 		}
 	}
 	if !runnerAvailable {
@@ -107,7 +111,7 @@ func chooseRunners(runnerType string) []ModelRunner {
 	// return the runners to try in priority order
 	localRunnersByPriority := []ModelRunner{}
 	for _, r := range runners {
-		localRunnersByPriority = append(localRunnersByPriority, ModelRunner{Path: path.Join(tmpDir, r)})
+		localRunnersByPriority = append(localRunnersByPriority, ModelRunner{Path: path.Join(workDir, r)})
 	}
 
 	return localRunnersByPriority

llm/llm.go

@@ -21,7 +21,7 @@ type LLM interface {
 	Ping(context.Context) error
 }
 
-func New(model string, adapters []string, opts api.Options) (LLM, error) {
+func New(workDir, model string, adapters []string, opts api.Options) (LLM, error) {
 	if _, err := os.Stat(model); err != nil {
 		return nil, err
 	}
@@ -91,9 +91,9 @@ func New(model string, adapters []string, opts api.Options) (LLM, error) {
 	switch ggml.Name() {
 	case "gguf":
 		opts.NumGQA = 0 // TODO: remove this when llama.cpp runners differ enough to need separate newLlama functions
-		return newLlama(model, adapters, ggufRunner(), opts)
+		return newLlama(model, adapters, chooseRunners(workDir, "gguf"), opts)
 	case "ggml", "ggmf", "ggjt", "ggla":
-		return newLlama(model, adapters, ggmlRunner(), opts)
+		return newLlama(model, adapters, chooseRunners(workDir, "ggml"), opts)
 	default:
 		return nil, fmt.Errorf("unknown ggml type: %s", ggml.ModelFamily())
 	}

server/images.go

@@ -267,7 +267,7 @@ func filenameWithPath(path, f string) (string, error) {
 	return f, nil
 }
 
-func CreateModel(ctx context.Context, name string, path string, fn func(resp api.ProgressResponse)) error {
+func CreateModel(ctx context.Context, workDir, name string, path string, fn func(resp api.ProgressResponse)) error {
 	mp := ParseModelPath(name)
 
 	var manifest *ManifestV2
@@ -524,7 +524,7 @@ func CreateModel(ctx context.Context, name string, path string, fn func(resp api
 	}
 
 	// generate the embedding layers
-	embeddingLayers, err := embeddingLayers(embed)
+	embeddingLayers, err := embeddingLayers(workDir, embed)
 	if err != nil {
 		return err
 	}
@@ -581,7 +581,7 @@ type EmbeddingParams struct {
 }
 
 // embeddingLayers loads the associated LLM and generates the embeddings to be stored from an input file
-func embeddingLayers(e EmbeddingParams) ([]*LayerReader, error) {
+func embeddingLayers(workDir string, e EmbeddingParams) ([]*LayerReader, error) {
 	layers := []*LayerReader{}
 	if len(e.files) > 0 {
 		// check if the model is a file path or a model name
@@ -594,7 +594,7 @@ func embeddingLayers(e EmbeddingParams) ([]*LayerReader, error) {
 			model = &Model{ModelPath: e.model}
 		}
 
-		if err := load(context.Background(), model, e.opts, defaultSessionDuration); err != nil {
+		if err := load(context.Background(), workDir, model, e.opts, defaultSessionDuration); err != nil {
 			return nil, fmt.Errorf("load model to generate embeddings: %v", err)
 		}
 

server/routes.go

@@ -58,7 +58,7 @@ var loaded struct {
 var defaultSessionDuration = 5 * time.Minute
 
 // load a model into memory if it is not already loaded, it is up to the caller to lock loaded.mu before calling this function
-func load(ctx context.Context, model *Model, reqOpts map[string]interface{}, sessionDuration time.Duration) error {
+func load(ctx context.Context, workDir string, model *Model, reqOpts map[string]interface{}, sessionDuration time.Duration) error {
 	opts := api.DefaultOptions()
 	if err := opts.FromMap(model.Options); err != nil {
 		log.Printf("could not load model options: %v", err)
@@ -94,7 +94,7 @@ func load(ctx context.Context, model *Model, reqOpts map[string]interface{}, ses
 			loaded.Embeddings = model.Embeddings
 		}
 
-		llmModel, err := llm.New(model.ModelPath, model.AdapterPaths, opts)
+		llmModel, err := llm.New(workDir, model.ModelPath, model.AdapterPaths, opts)
 		if err != nil {
 			return err
 		}
@@ -130,6 +130,7 @@ func load(ctx context.Context, model *Model, reqOpts map[string]interface{}, ses
 			llmModel.SetOptions(opts)
 		}
 	}
+
 	loaded.expireAt = time.Now().Add(sessionDuration)
 
 	if loaded.expireTimer == nil {
@@ -150,6 +151,7 @@ func load(ctx context.Context, model *Model, reqOpts map[string]interface{}, ses
 			loaded.digest = ""
 		})
 	}
+
 	loaded.expireTimer.Reset(sessionDuration)
 	return nil
 }
@@ -172,8 +174,11 @@ func GenerateHandler(c *gin.Context) {
 		return
 	}
 
-	sessionDuration := defaultSessionDuration // TODO: set this duration from the request if specified
-	if err := load(c.Request.Context(), model, req.Options, sessionDuration); err != nil {
+	workDir := c.GetString("workDir")
+
+	// TODO: set this duration from the request if specified
+	sessionDuration := defaultSessionDuration
+	if err := load(c.Request.Context(), workDir, model, req.Options, sessionDuration); err != nil {
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 		return
 	}
@@ -245,7 +250,9 @@ func EmbeddingHandler(c *gin.Context) {
 		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
 		return
 	}
-	if err := load(c.Request.Context(), model, req.Options, 5*time.Minute); err != nil {
+
+	workDir := c.GetString("workDir")
+	if err := load(c.Request.Context(), workDir, model, req.Options, 5*time.Minute); err != nil {
 		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
 		return
 	}
@@ -335,6 +342,8 @@ func CreateModelHandler(c *gin.Context) {
 		return
 	}
 
+	workDir := c.GetString("workDir")
+
 	ch := make(chan any)
 	go func() {
 		defer close(ch)
@@ -345,7 +354,7 @@ func CreateModelHandler(c *gin.Context) {
 		ctx, cancel := context.WithCancel(c.Request.Context())
 		defer cancel()
 
-		if err := CreateModel(ctx, req.Name, req.Path, fn); err != nil {
+		if err := CreateModel(ctx, workDir, req.Name, req.Path, fn); err != nil {
 			ch <- gin.H{"error": err.Error()}
 		}
 	}()
@@ -519,8 +528,20 @@ func Serve(ln net.Listener, allowOrigins []string) error {
 		)
 	}
 
+	workDir, err := os.MkdirTemp("", "ollama")
+	if err != nil {
+		return err
+	}
+	defer os.RemoveAll(workDir)
+
 	r := gin.Default()
-	r.Use(cors.New(config))
+	r.Use(
+		cors.New(config),
+		func(c *gin.Context) {
+			c.Set("workDir", workDir)
+			c.Next()
+		},
+	)
 
 	r.GET("/", func(c *gin.Context) {
 		c.String(http.StatusOK, "Ollama is running")
@@ -546,12 +567,10 @@ func Serve(ln net.Listener, allowOrigins []string) error {
 
 	// listen for a ctrl+c and stop any loaded llm
 	signals := make(chan os.Signal, 1)
-	signal.Notify(signals, syscall.SIGINT)
+	signal.Notify(signals, syscall.SIGINT, syscall.SIGTERM)
 	go func() {
 		<-signals
-		if loaded.llm != nil {
-			loaded.llm.Close()
-		}
+		os.RemoveAll(workDir)
 		os.Exit(0)
 	}()