Update amd_linux.go

remove old gpu limits due to the new rocm supports

.github/workflows/release.yaml

@@ -28,6 +28,7 @@ jobs:
           security unlock-keychain -p password build.keychain
           security import certificate.p12 -k build.keychain -P $MACOS_SIGNING_KEY_PASSWORD -T /usr/bin/codesign
           security set-key-partition-list -S apple-tool:,apple:,codesign: -s -k password build.keychain
+          security set-keychain-settings -lut 3600 build.keychain
       - uses: actions/setup-go@v5
         with:
           go-version-file: go.mod

README.md

@@ -69,15 +69,17 @@ Here are some example models that can be downloaded:
 | ------------------ | ---------- | ----- | ------------------------------ |
 | Llama 3            | 8B         | 4.7GB | `ollama run llama3`            |
 | Llama 3            | 70B        | 40GB  | `ollama run llama3:70b`        |
-| Phi-3              | 3.8B       | 2.3GB | `ollama run phi3`              |
+| Phi 3 Mini         | 3.8B       | 2.3GB | `ollama run phi3`              |
+| Phi 3 Medium       | 14B        | 7.9GB | `ollama run phi3:medium`       |
+| Gemma              | 2B         | 1.4GB | `ollama run gemma:2b`          |
+| Gemma              | 7B         | 4.8GB | `ollama run gemma:7b`          |
 | Mistral            | 7B         | 4.1GB | `ollama run mistral`           |
+| Moondream 2        | 1.4B       | 829MB | `ollama run moondream`         |
 | Neural Chat        | 7B         | 4.1GB | `ollama run neural-chat`       |
 | Starling           | 7B         | 4.1GB | `ollama run starling-lm`       |
 | Code Llama         | 7B         | 3.8GB | `ollama run codellama`         |
 | Llama 2 Uncensored | 7B         | 3.8GB | `ollama run llama2-uncensored` |
 | LLaVA              | 7B         | 4.5GB | `ollama run llava`             |
-| Gemma              | 2B         | 1.4GB | `ollama run gemma:2b`          |
-| Gemma              | 7B         | 4.8GB | `ollama run gemma:7b`          |
 | Solar              | 10.7B      | 6.1GB | `ollama run solar`             |
 
 > Note: You should have at least 8 GB of RAM available to run the 7B models, 16 GB to run the 13B models, and 32 GB to run the 33B models.
@@ -210,25 +212,7 @@ ollama list
 
 ## Building
 
-Install `cmake` and `go`:
+See the [developer guide](https://github.com/ollama/ollama/blob/main/docs/development.md)
-
-```
-brew install cmake go
-```
-
-Then generate dependencies:
-
-```
-go generate ./...
-```
-
-Then build the binary:
-
-```
-go build .
-```
-
-More detailed instructions can be found in the [developer guide](https://github.com/ollama/ollama/blob/main/docs/development.md)
 
 ### Running local builds
 
@@ -377,7 +361,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Testcontainers](https://testcontainers.com/modules/ollama/)
 - [Portkey](https://portkey.ai/docs/welcome/integration-guides/ollama)
 - [PromptingTools.jl](https://github.com/svilupp/PromptingTools.jl) with an [example](https://svilupp.github.io/PromptingTools.jl/dev/examples/working_with_ollama)
-- [LlamaScript](https://github.com/WolfTheDeveloper/llamascript)
+- [LlamaScript](https://github.com/Project-Llama/llamascript)
 ### Mobile
 
 - [Enchanted](https://github.com/AugustDev/enchanted)
@@ -409,6 +393,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [AI Telegram Bot](https://github.com/tusharhero/aitelegrambot) (Telegram bot using Ollama in backend)
 - [AI ST Completion](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (Sublime Text 4 AI assistant plugin with Ollama support)
 - [Discord-Ollama Chat Bot](https://github.com/kevinthedang/discord-ollama) (Generalized TypeScript Discord Bot w/ Tuning Documentation)
+- [Discord AI chat/moderation bot](https://github.com/rapmd73/Companion) Chat/moderation bot written in python. Uses Ollama to create personalities.
 
 ### Supported backends 
 - [llama.cpp](https://github.com/ggerganov/llama.cpp) project founded by Georgi Gerganov.

api/client.go

@@ -354,6 +354,15 @@ func (c *Client) List(ctx context.Context) (*ListResponse, error) {
 	return &lr, nil
 }
 
+// List running models.
+func (c *Client) ListRunning(ctx context.Context) (*ListResponse, error) {
+	var lr ListResponse
+	if err := c.do(ctx, http.MethodGet, "/api/ps", nil, &lr); err != nil {
+		return nil, err
+	}
+	return &lr, nil
+}
+
 // Copy copies a model - creating a model with another name from an existing
 // model.
 func (c *Client) Copy(ctx context.Context, req *CopyRequest) error {

api/types.go

@@ -289,10 +289,12 @@ type ListResponse struct {
 type ModelResponse struct {
 	Name       string       `json:"name"`
 	Model      string       `json:"model"`
-	ModifiedAt time.Time    `json:"modified_at"`
+	ModifiedAt time.Time    `json:"modified_at,omitempty"`
 	Size       int64        `json:"size"`
 	Digest     string       `json:"digest"`
 	Details    ModelDetails `json:"details,omitempty"`
+	ExpiresAt  time.Time    `json:"expires_at,omitempty"`
+	SizeVRAM   int64        `json:"size_vram,omitempty"`
 }
 
 type TokenResponse struct {

cmd/cmd.go

@@ -12,6 +12,7 @@ import (
 	"fmt"
 	"io"
 	"log"
+	"math"
 	"net"
 	"net/http"
 	"os"
@@ -24,7 +25,7 @@ import (
 	"time"
 
 	"github.com/containerd/console"
-
+	"github.com/mattn/go-runewidth"
 	"github.com/olekukonko/tablewriter"
 	"github.com/spf13/cobra"
 	"golang.org/x/crypto/ssh"
@@ -34,6 +35,7 @@ import (
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/auth"
 	"github.com/ollama/ollama/format"
+	"github.com/ollama/ollama/parser"
 	"github.com/ollama/ollama/progress"
 	"github.com/ollama/ollama/server"
 	"github.com/ollama/ollama/types/errtypes"
@@ -62,7 +64,7 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	}
 	defer f.Close()
 
-	modelfile, err := model.ParseFile(f)
+	modelfile, err := parser.ParseFile(f)
 	if err != nil {
 		return err
 	}
@@ -206,7 +208,7 @@ func tempZipFiles(path string) (string, error) {
 		// pytorch files might also be unresolved git lfs references; skip if they are
 		// covers pytorch_model-x-of-y.bin, pytorch_model.fp32-x-of-y.bin, pytorch_model.bin
 		files = append(files, pt...)
-	} else if pt, _ := glob(filepath.Join(path, "consolidated*.pth"), "application/octet-stream"); len(pt) > 0 {
+	} else if pt, _ := glob(filepath.Join(path, "consolidated*.pth"), "application/zip"); len(pt) > 0 {
 		// pytorch files might also be unresolved git lfs references; skip if they are
 		// covers consolidated.x.pth, consolidated.pth
 		files = append(files, pt...)
@@ -324,6 +326,18 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	}
 	opts.Format = format
 
+	keepAlive, err := cmd.Flags().GetString("keepalive")
+	if err != nil {
+		return err
+	}
+	if keepAlive != "" {
+		d, err := time.ParseDuration(keepAlive)
+		if err != nil {
+			return err
+		}
+		opts.KeepAlive = &api.Duration{Duration: d}
+	}
+
 	prompts := args[1:]
 	// prepend stdin to the prompt if provided
 	if !term.IsTerminal(int(os.Stdin.Fd())) {
@@ -496,6 +510,52 @@ func ListHandler(cmd *cobra.Command, args []string) error {
 	return nil
 }
 
+func ListRunningHandler(cmd *cobra.Command, args []string) error {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return err
+	}
+
+	models, err := client.ListRunning(cmd.Context())
+	if err != nil {
+		return err
+	}
+
+	var data [][]string
+
+	for _, m := range models.Models {
+		if len(args) == 0 || strings.HasPrefix(m.Name, args[0]) {
+			var procStr string
+			switch {
+			case m.SizeVRAM == 0:
+				procStr = "100% CPU"
+			case m.SizeVRAM == m.Size:
+				procStr = "100% GPU"
+			case m.SizeVRAM > m.Size || m.Size == 0:
+				procStr = "Unknown"
+			default:
+				sizeCPU := m.Size - m.SizeVRAM
+				cpuPercent := math.Round(float64(sizeCPU) / float64(m.Size) * 100)
+				procStr = fmt.Sprintf("%d%%/%d%% CPU/GPU", int(cpuPercent), int(100-cpuPercent))
+			}
+			data = append(data, []string{m.Name, m.Digest[:12], format.HumanBytes(m.Size), procStr, format.HumanTime(m.ExpiresAt, "Never")})
+		}
+	}
+
+	table := tablewriter.NewWriter(os.Stdout)
+	table.SetHeader([]string{"NAME", "ID", "SIZE", "PROCESSOR", "UNTIL"})
+	table.SetHeaderAlignment(tablewriter.ALIGN_LEFT)
+	table.SetAlignment(tablewriter.ALIGN_LEFT)
+	table.SetHeaderLine(false)
+	table.SetBorder(false)
+	table.SetNoWhiteSpace(true)
+	table.SetTablePadding("\t")
+	table.AppendBulk(data)
+	table.Render()
+
+	return nil
+}
+
 func DeleteHandler(cmd *cobra.Command, args []string) error {
 	client, err := api.ClientFromEnvironment()
 	if err != nil {
@@ -672,6 +732,7 @@ type runOptions struct {
 	Images      []api.ImageData
 	Options     map[string]interface{}
 	MultiModal  bool
+	KeepAlive   *api.Duration
 }
 
 type displayResponseState struct {
@@ -684,7 +745,8 @@ func displayResponse(content string, wordWrap bool, state *displayResponseState)
 	if wordWrap && termWidth >= 10 {
 		for _, ch := range content {
 			if state.lineLength+1 > termWidth-5 {
-				if len(state.wordBuffer) > termWidth-10 {
+
+				if runewidth.StringWidth(state.wordBuffer) > termWidth-10 {
 					fmt.Printf("%s%c", state.wordBuffer, ch)
 					state.wordBuffer = ""
 					state.lineLength = 0
@@ -692,12 +754,18 @@ func displayResponse(content string, wordWrap bool, state *displayResponseState)
 				}
 
 				// backtrack the length of the last word and clear to the end of the line
-				fmt.Printf("\x1b[%dD\x1b[K\n", len(state.wordBuffer))
+				fmt.Printf("\x1b[%dD\x1b[K\n", runewidth.StringWidth(state.wordBuffer))
 				fmt.Printf("%s%c", state.wordBuffer, ch)
-				state.lineLength = len(state.wordBuffer) + 1
+				chWidth := runewidth.RuneWidth(ch)
+
+				state.lineLength = runewidth.StringWidth(state.wordBuffer) + chWidth
 			} else {
 				fmt.Print(string(ch))
-				state.lineLength += 1
+				state.lineLength += runewidth.RuneWidth(ch)
+				if runewidth.RuneWidth(ch) >= 2 {
+					state.wordBuffer = ""
+					continue
+				}
 
 				switch ch {
 				case ' ':
@@ -766,6 +834,10 @@ func chat(cmd *cobra.Command, opts runOptions) (*api.Message, error) {
 		Options:  opts.Options,
 	}
 
+	if opts.KeepAlive != nil {
+		req.KeepAlive = opts.KeepAlive
+	}
+
 	if err := client.Chat(cancelCtx, req, fn); err != nil {
 		if errors.Is(err, context.Canceled) {
 			return nil, nil
@@ -841,14 +913,15 @@ func generate(cmd *cobra.Command, opts runOptions) error {
 	}
 
 	request := api.GenerateRequest{
-		Model:    opts.Model,
+		Model:     opts.Model,
-		Prompt:   opts.Prompt,
+		Prompt:    opts.Prompt,
-		Context:  generateContext,
+		Context:   generateContext,
-		Images:   opts.Images,
+		Images:    opts.Images,
-		Format:   opts.Format,
+		Format:    opts.Format,
-		System:   opts.System,
+		System:    opts.System,
-		Template: opts.Template,
+		Template:  opts.Template,
-		Options:  opts.Options,
+		Options:   opts.Options,
+		KeepAlive: opts.KeepAlive,
 	}
 
 	if err := client.Generate(ctx, &request, fn); err != nil {
@@ -1006,12 +1079,24 @@ func versionHandler(cmd *cobra.Command, _ []string) {
 	}
 }
 
-func appendHostEnvDocs(cmd *cobra.Command) {
+type EnvironmentVar struct {
-	const hostEnvDocs = `
+	Name        string
+	Description string
+}
+
+func appendEnvDocs(cmd *cobra.Command, envs []EnvironmentVar) {
+	if len(envs) == 0 {
+		return
+	}
+
+	envUsage := `
 Environment Variables:
-      OLLAMA_HOST        The host:port or base URL of the Ollama server (e.g. http://localhost:11434)
 `
-	cmd.SetUsageTemplate(cmd.UsageTemplate() + hostEnvDocs)
+	for _, e := range envs {
+		envUsage += fmt.Sprintf("      %-16s   %s\n", e.Name, e.Description)
+	}
+
+	cmd.SetUsageTemplate(cmd.UsageTemplate() + envUsage)
 }
 
 func NewCLI() *cobra.Command {
@@ -1075,6 +1160,7 @@ func NewCLI() *cobra.Command {
 		RunE:    RunHandler,
 	}
 
+	runCmd.Flags().String("keepalive", "", "Duration to keep a model loaded (e.g. 5m)")
 	runCmd.Flags().Bool("verbose", false, "Show timings for response")
 	runCmd.Flags().Bool("insecure", false, "Use an insecure registry")
 	runCmd.Flags().Bool("nowordwrap", false, "Don't wrap words to the next line automatically")
@@ -1090,9 +1176,9 @@ func NewCLI() *cobra.Command {
 Environment Variables:
 
     OLLAMA_HOST         The host:port to bind to (default "127.0.0.1:11434")
-    OLLAMA_ORIGINS      A comma separated list of allowed origins.
+    OLLAMA_ORIGINS      A comma separated list of allowed origins
-    OLLAMA_MODELS       The path to the models directory (default is "~/.ollama/models")
+    OLLAMA_MODELS       The path to the models directory (default "~/.ollama/models")
-    OLLAMA_KEEP_ALIVE   The duration that models stay loaded in memory (default is "5m")
+    OLLAMA_KEEP_ALIVE   The duration that models stay loaded in memory (default "5m")
     OLLAMA_DEBUG        Set to 1 to enable additional debug logging
 `)
 
@@ -1123,6 +1209,14 @@ Environment Variables:
 		PreRunE: checkServerHeartbeat,
 		RunE:    ListHandler,
 	}
+
+	psCmd := &cobra.Command{
+		Use:     "ps",
+		Short:   "List running models",
+		PreRunE: checkServerHeartbeat,
+		RunE:    ListRunningHandler,
+	}
+
 	copyCmd := &cobra.Command{
 		Use:     "cp SOURCE DESTINATION",
 		Short:   "Copy a model",
@@ -1139,6 +1233,10 @@ Environment Variables:
 		RunE:    DeleteHandler,
 	}
 
+	ollamaHostEnv := EnvironmentVar{"OLLAMA_HOST", "The host:port or base URL of the Ollama server (e.g. http://localhost:11434)"}
+	ollamaNoHistoryEnv := EnvironmentVar{"OLLAMA_NOHISTORY", "Disable readline history"}
+	envs := []EnvironmentVar{ollamaHostEnv}
+
 	for _, cmd := range []*cobra.Command{
 		createCmd,
 		showCmd,
@@ -1146,10 +1244,16 @@ Environment Variables:
 		pullCmd,
 		pushCmd,
 		listCmd,
+		psCmd,
 		copyCmd,
 		deleteCmd,
 	} {
-		appendHostEnvDocs(cmd)
+		switch cmd {
+		case runCmd:
+			appendEnvDocs(cmd, []EnvironmentVar{ollamaHostEnv, ollamaNoHistoryEnv})
+		default:
+			appendEnvDocs(cmd, envs)
+		}
 	}
 
 	rootCmd.AddCommand(
@@ -1160,6 +1264,7 @@ Environment Variables:
 		pullCmd,
 		pushCmd,
 		listCmd,
+		psCmd,
 		copyCmd,
 		deleteCmd,
 	)

cmd/interactive.go

@@ -17,6 +17,7 @@ import (
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/progress"
 	"github.com/ollama/ollama/readline"
+	"github.com/ollama/ollama/types/errtypes"
 )
 
 type MultilineState int
@@ -56,6 +57,11 @@ func loadModel(cmd *cobra.Command, opts *runOptions) error {
 		Model:    opts.Model,
 		Messages: []api.Message{},
 	}
+
+	if opts.KeepAlive != nil {
+		chatReq.KeepAlive = opts.KeepAlive
+	}
+
 	err = client.Chat(cmd.Context(), chatReq, func(resp api.ChatResponse) error {
 		p.StopAndClear()
 		if len(opts.Messages) > 0 {
@@ -132,6 +138,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 		fmt.Fprintln(os.Stderr, "   Alt + f            Move forward (right) one word")
 		fmt.Fprintln(os.Stderr, "  Ctrl + k            Delete the sentence after the cursor")
 		fmt.Fprintln(os.Stderr, "  Ctrl + u            Delete the sentence before the cursor")
+		fmt.Fprintln(os.Stderr, "  Ctrl + w            Delete the word before the cursor")
 		fmt.Fprintln(os.Stderr, "")
 		fmt.Fprintln(os.Stderr, "  Ctrl + l            Clear the screen")
 		fmt.Fprintln(os.Stderr, "  Ctrl + c            Stop the model from responding")
@@ -176,6 +183,10 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 		return err
 	}
 
+	if os.Getenv("OLLAMA_NOHISTORY") != "" {
+		scanner.HistoryDisable()
+	}
+
 	fmt.Print(readline.StartBracketedPaste)
 	defer fmt.Printf(readline.EndBracketedPaste)
 
@@ -276,13 +287,20 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 			fn := func(resp api.ProgressResponse) error { return nil }
 			err = client.Create(cmd.Context(), req, fn)
 			if err != nil {
-				fmt.Println("error: couldn't save model")
+				if strings.Contains(err.Error(), errtypes.InvalidModelNameErrMsg) {
+					fmt.Printf("error: The model name '%s' is invalid\n", args[1])
+					continue
+				}
 				return err
 			}
 			fmt.Printf("Created new model '%s'\n", args[1])
 			continue
 		case strings.HasPrefix(line, "/clear"):
 			opts.Messages = []api.Message{}
+			if opts.System != "" {
+				newMessage := api.Message{Role: "system", Content: opts.System}
+				opts.Messages = append(opts.Messages, newMessage)
+			}
 			fmt.Println("Cleared session context")
 			continue
 		case strings.HasPrefix(line, "/set"):

convert/convert.go

@@ -18,6 +18,16 @@ import (
 	"github.com/ollama/ollama/llm"
 )
 
+const (
+	_ int32 = iota
+	tokenTypeNormal
+	tokenTypeUnknown
+	tokenTypeControl
+	tokenTypeUserDefined
+	tokenTypeUnused
+	tokenTypeByte
+)
+
 type Params struct {
 	Architectures     []string `json:"architectures"`
 	VocabSize         int      `json:"vocab_size"`
@@ -37,6 +47,8 @@ type Params struct {
 	Experts     int `json:"num_local_experts"`
 	ExpertsUsed int `json:"num_experts_per_tok"`
 
+	PreTokenizer string
+
 	ByteOrder
 }
 
@@ -74,10 +86,9 @@ func GetModelFormat(dirname string) (ModelFormat, error) {
 	}
 
 	for _, fn := range files {
-		slog.Debug(fmt.Sprintf("file = %s", fn))
 		if strings.HasSuffix(fn, ".safetensors") {
 			return &SafetensorFormat{}, nil
-		} else if strings.HasSuffix(fn, ".bin") {
+		} else if strings.HasSuffix(fn, ".bin") || strings.HasSuffix(fn, ".pth") {
 			slog.Debug("model is torch")
 			return &TorchFormat{}, nil
 		}
@@ -92,6 +103,7 @@ type Vocab struct {
 	Tokens []string
 	Scores []float32
 	Types  []int32
+	Merges []string
 }
 
 func LoadSentencePieceTokens(dirpath string, params *Params) (*Vocab, error) {
@@ -170,7 +182,7 @@ func LoadSentencePieceTokens(dirpath string, params *Params) (*Vocab, error) {
 		}
 		v.Tokens = append(v.Tokens, t.key)
 		v.Scores = append(v.Scores, -1000.0)
-		v.Types = append(v.Types, int32(llm.GGUFTokenUserDefined))
+		v.Types = append(v.Types, tokenTypeUserDefined)
 	}
 	slog.Info(fmt.Sprintf("vocab size w/ extra tokens: %d", len(v.Tokens)))
 
@@ -180,7 +192,7 @@ func LoadSentencePieceTokens(dirpath string, params *Params) (*Vocab, error) {
 		for cnt := 0; cnt < missingTokens; cnt++ {
 			v.Tokens = append(v.Tokens, fmt.Sprintf("<dummy%05d>", cnt+1))
 			v.Scores = append(v.Scores, -1)
-			v.Types = append(v.Types, int32(llm.GGUFTokenUserDefined))
+			v.Types = append(v.Types, tokenTypeUserDefined)
 		}
 	}
 

convert/convert_test.go

@@ -0,0 +1,103 @@
+//go:build slow
+
+package convert
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+
+	"github.com/ollama/ollama/llm"
+)
+
+func convertFull(t *testing.T, p string) (llm.KV, llm.Tensors) {
+	t.Helper()
+
+	mf, err := GetModelFormat(p)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	params, err := mf.GetParams(p)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	arch, err := mf.GetModelArch("", p, params)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	if err := arch.LoadVocab(); err != nil {
+		t.Fatal(err)
+	}
+
+	if err := arch.GetTensors(); err != nil {
+		t.Fatal(err)
+	}
+
+	f, err := os.CreateTemp(t.TempDir(), "f16")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer f.Close()
+
+	if err := arch.WriteGGUF(f); err != nil {
+		t.Fatal(err)
+	}
+
+	r, err := os.Open(f.Name())
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer r.Close()
+
+	m, _, err := llm.DecodeGGML(r)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	return m.KV(), m.Tensors()
+}
+
+func TestConvertFull(t *testing.T) {
+	cases := []struct {
+		path    string
+		arch    string
+		tensors int
+		layers  int
+	}{
+		{"Meta-Llama-3-8B-Instruct", "llama", 291, 35},
+		{"Mistral-7B-Instruct-v0.2", "llama", 291, 35},
+		{"Mixtral-8x7B-Instruct-v0.1", "llama", 291, 35},
+		{"gemma-2b-it", "gemma", 164, 20},
+	}
+
+	for _, tt := range cases {
+		t.Run(tt.path, func(t *testing.T) {
+			p := filepath.Join("testdata", tt.path)
+			if _, err := os.Stat(p); err != nil {
+				t.Skipf("%s not found", p)
+			}
+
+			kv, tensors := convertFull(t, p)
+
+			if kv.Architecture() != tt.arch {
+				t.Fatalf("expected llama, got %s", kv.Architecture())
+			}
+
+			if kv.FileType().String() != "F16" {
+				t.Fatalf("expected F16, got %s", kv.FileType())
+			}
+
+			if len(tensors) != tt.tensors {
+				t.Fatalf("expected %d tensors, got %d", tt.tensors, len(tensors))
+			}
+
+			layers := tensors.Layers()
+			if len(layers) != tt.layers {
+				t.Fatalf("expected %d layers, got %d", tt.layers, len(layers))
+			}
+		})
+	}
+}

convert/gemma.go

@@ -1,14 +1,11 @@
 package convert
 
 import (
-	"encoding/binary"
 	"fmt"
 	"io"
 	"log/slog"
-	"os"
 	"strings"
 
-	"github.com/d4l3k/go-bfloat16"
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
 
@@ -19,49 +16,27 @@ type GemmaModel struct {
 	ModelData
 }
 
-func gemmaLayerHandler(w io.Writer, r safetensorWriterTo, f *os.File) error {
-	slog.Debug(fmt.Sprintf("converting '%s'", r.t.Name))
-
-	data := make([]byte, r.end-r.start)
-	if err := binary.Read(f, r.bo, data); err != nil {
-		return err
-	}
-
-	tDataF32 := bfloat16.DecodeFloat32(data)
-
-	var err error
-	tDataF32, err = addOnes(tDataF32, int(r.t.Shape[0]))
-	if err != nil {
-		return err
-	}
-
-	if err := binary.Write(w, r.bo, tDataF32); err != nil {
-		return err
-	}
-	return nil
-}
-
 func addOnes(data []float32, vectorSize int) ([]float32, error) {
 	n := tensor.New(tensor.WithShape(vectorSize), tensor.WithBacking(data))
 	ones := tensor.Ones(tensor.Float32, vectorSize)
 
-	var err error
+	n, err := n.Add(ones)
-	n, err = n.Add(ones)
 	if err != nil {
-		return []float32{}, err
+		return nil, err
 	}
 
-	newN, err := native.SelectF32(n, 0)
+	ts, err := native.SelectF32(n, 0)
 	if err != nil {
-		return []float32{}, err
+		return nil, err
 	}
 
-	var fullTensor []float32
+	var f32s []float32
-	for _, v := range newN {
+	for _, t := range ts {
-		fullTensor = append(fullTensor, v...)
+		f32s = append(f32s, t...)
 	}
 
-	return fullTensor, nil
+
+	return f32s, nil
 }
 
 func (m *GemmaModel) GetTensors() error {
@@ -71,12 +46,10 @@ func (m *GemmaModel) GetTensors() error {
 	}
 
 	slog.Debug(fmt.Sprintf("Total tensors: %d", len(t)))
-
-	m.Tensors = []llm.Tensor{}
 	for _, l := range t {
 		if strings.HasSuffix(l.Name, "norm.weight") {
 			wt := l.WriterTo.(safetensorWriterTo)
-			wt.handler = gemmaLayerHandler
+			wt.repacker = m.Repack
 			l.WriterTo = wt
 		}
 		m.Tensors = append(m.Tensors, l)
@@ -94,6 +67,10 @@ func (m *GemmaModel) LoadVocab() error {
 	return nil
 }
 
+func (m *GemmaModel) Repack(_ string, data []float32, shape []uint64) ([]float32, error) {
+	return addOnes(data, int(shape[0]))
+}
+
 func (m *GemmaModel) WriteGGUF(ws io.WriteSeeker) error {
 	kv := llm.KV{
 		"general.architecture":                   "gemma",

convert/llama.go

@@ -1,17 +1,17 @@
 package convert
 
 import (
-	"encoding/binary"
+	"cmp"
+	"errors"
 	"fmt"
 	"io"
-	"log/slog"
+	"os"
+	"path/filepath"
 	"regexp"
 	"strings"
 
-	"github.com/nlpodyssey/gopickle/pytorch"
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
-	"github.com/x448/float16"
 
 	"github.com/ollama/ollama/llm"
 )
@@ -20,81 +20,12 @@ type LlamaModel struct {
 	ModelData
 }
 
-func llamaLayerHandler(w io.Writer, r torchWriterTo) error {
-	slog.Debug(fmt.Sprintf("repacking layer '%s'", r.t.Name))
-
-	data := r.storage.(*pytorch.HalfStorage).Data
-	tData := make([]uint16, len(data))
-	for cnt, v := range data {
-		tData[cnt] = uint16(float16.Fromfloat32(v))
-	}
-
-	var err error
-	var heads uint32
-	if strings.Contains(r.t.Name, "attn_q") {
-		heads = uint32(r.params.AttentionHeads)
-	} else if strings.Contains(r.t.Name, "attn_k") {
-		heads = uint32(r.params.KeyValHeads)
-		if heads == 0 {
-			heads = uint32(r.params.AttentionHeads)
-		}
-	} else {
-		return fmt.Errorf("unknown layer type")
-	}
-
-	slog.Debug(fmt.Sprintf("heads = %d", heads))
-
-	tData, err = llamaRepack(tData, int(heads), r.t.Shape)
-	if err != nil {
-		return err
-	}
-
-	if err = binary.Write(w, r.bo, tData); err != nil {
-		return err
-	}
-	return nil
-}
-
-func llamaRepack(data []uint16, heads int, shape []uint64) ([]uint16, error) {
-	n := tensor.New(tensor.WithShape(int(shape[0]), int(shape[1])), tensor.WithBacking(data))
-	origShape := n.Shape().Clone()
-
-	// reshape the tensor and swap axes 1 and 2 to unpack the layer for gguf
-	if err := n.Reshape(heads, 2, origShape[0]/heads/2, origShape[1]); err != nil {
-		return nil, err
-	}
-
-	if err := n.T(0, 2, 1, 3); err != nil {
-		return nil, err
-	}
-
-	if err := n.Reshape(origShape...); err != nil {
-		return nil, err
-	}
-
-	if err := n.Transpose(); err != nil {
-		return nil, err
-	}
-	newN, err := native.SelectU16(n, 1)
-	if err != nil {
-		return nil, err
-	}
-
-	var fullTensor []uint16
-	for _, v := range newN {
-		fullTensor = append(fullTensor, v...)
-	}
-	return fullTensor, nil
-}
-
 func (m *LlamaModel) GetTensors() error {
 	t, err := m.Format.GetTensors(m.Path, m.Params)
 	if err != nil {
 		return err
 	}
 
-	m.Tensors = []llm.Tensor{}
-
 	pattern := `^blk\.[0-9]+\.attn_(?P<layer>q|k)\.weight$`
 	re, err := regexp.Compile(pattern)
 	if err != nil {
@@ -104,10 +35,16 @@ func (m *LlamaModel) GetTensors() error {
 	for _, l := range t {
 		matches := re.FindAllStringSubmatch(l.Name, -1)
 		if len(matches) > 0 {
-			slog.Debug(fmt.Sprintf("setting handler for: %s", l.Name))
+			switch m.Format.(type) {
-			wt := l.WriterTo.(torchWriterTo)
+			case *TorchFormat:
-			wt.handler = llamaLayerHandler
+				wt := l.WriterTo.(torchWriterTo)
-			l.WriterTo = wt
+				wt.repacker = m.Repack
+				l.WriterTo = wt
+			case *SafetensorFormat:
+				wt := l.WriterTo.(safetensorWriterTo)
+				wt.repacker = m.Repack
+				l.WriterTo = wt
+			}
 		}
 		m.Tensors = append(m.Tensors, l)
 	}
@@ -115,19 +52,22 @@ func (m *LlamaModel) GetTensors() error {
 	return nil
 }
 
-func (m *LlamaModel) LoadVocab() error {
+func (m *LlamaModel) LoadVocab() (err error) {
-	var v *Vocab
+	pre, ts, merges, err := parseTokens(filepath.Join(m.Path, "tokenizer.json"))
-	var err error
+	if errors.Is(err, os.ErrNotExist) {
-
+		return nil
-	slog.Debug("loading vocab")
+	} else if err != nil {
-	v, err = LoadSentencePieceTokens(m.Path, m.Params)
-	if err != nil {
 		return err
 	}
 
-	slog.Debug("vocab loaded")
+	m.Vocab = &Vocab{}
+	for _, t := range ts {
+		m.Vocab.Tokens = append(m.Vocab.Tokens, t.Content)
+		m.Vocab.Types = append(m.Vocab.Types, t.Type())
+	}
 
-	m.Vocab = v
+	m.Vocab.Merges = merges
+	m.Params.PreTokenizer = pre
 	return nil
 }
 
@@ -140,23 +80,79 @@ func (m *LlamaModel) WriteGGUF(ws io.WriteSeeker) error {
 		"llama.embedding_length":                 uint32(m.Params.HiddenSize),
 		"llama.block_count":                      uint32(m.Params.HiddenLayers),
 		"llama.feed_forward_length":              uint32(m.Params.IntermediateSize),
+		"llama.rope.freq_base":                   float32(m.Params.RopeFrequencyBase),
 		"llama.rope.dimension_count":             uint32(m.Params.HiddenSize / m.Params.AttentionHeads),
 		"llama.attention.head_count":             uint32(m.Params.AttentionHeads),
 		"llama.attention.head_count_kv":          uint32(m.Params.KeyValHeads),
 		"llama.attention.layer_norm_rms_epsilon": float32(m.Params.NormEPS),
 		"general.file_type":                      uint32(1),
-		"tokenizer.ggml.model":                   "llama",
+		"tokenizer.ggml.model":                   "gpt2",
 
+		"tokenizer.ggml.pre":        m.Params.PreTokenizer,
 		"tokenizer.ggml.tokens":     m.Vocab.Tokens,
-		"tokenizer.ggml.scores":     m.Vocab.Scores,
 		"tokenizer.ggml.token_type": m.Vocab.Types,
 
 		"tokenizer.ggml.bos_token_id":     uint32(m.Params.BoSTokenID),
 		"tokenizer.ggml.eos_token_id":     uint32(m.Params.EoSTokenID),
 		"tokenizer.ggml.unknown_token_id": uint32(0),
-		"tokenizer.ggml.add_bos_token":    true,
+	}
-		"tokenizer.ggml.add_eos_token":    false,
+
+	if len(m.Vocab.Merges) > 0 {
+		kv["tokenizer.ggml.merges"] = m.Vocab.Merges
+	} else {
+		kv["tokenizer.ggml.scores"] = m.Vocab.Scores
 	}
 
 	return llm.NewGGUFV3(m.Params.ByteOrder).Encode(ws, kv, m.Tensors)
 }
+
+func (m *LlamaModel) Repack(name string, data []float32, shape []uint64) ([]float32, error) {
+	return llamaRepack(name, m.Params, data, shape)
+}
+
+func llamaRepack(name string, params *Params, data []float32, shape []uint64) ([]float32, error) {
+	var dims []int
+	for _, dim := range shape {
+		if dim != 0 {
+			dims = append(dims, int(dim))
+		}
+	}
+
+	var heads int
+	if strings.HasSuffix(name, "attn_q.weight") {
+		heads = params.AttentionHeads
+	} else if strings.HasSuffix(name, "attn_k.weight") {
+		heads = cmp.Or(params.KeyValHeads, params.AttentionHeads)
+	} else {
+		return nil, fmt.Errorf("unknown tensor name: %s", name)
+	}
+
+	n := tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+	if err := n.Reshape(append([]int{heads, 2, dims[0] / heads / 2}, dims[1:]...)...); err != nil {
+		return nil, err
+	}
+
+	if err := n.T(0, 2, 1, 3); err != nil {
+		return nil, err
+	}
+
+	if err := n.Reshape(dims...); err != nil {
+		return nil, err
+	}
+
+	if err := n.Transpose(); err != nil {
+		return nil, err
+	}
+
+	ts, err := native.SelectF32(n, 1)
+	if err != nil {
+		return nil, err
+	}
+
+	var f32s []float32
+	for _, t := range ts {
+		f32s = append(f32s, t...)
+	}
+
+	return f32s, nil
+}

convert/mistral.go

@@ -1,17 +1,8 @@
 package convert
 
 import (
-	"encoding/binary"
-	"fmt"
 	"io"
-	"os"
 	"regexp"
-	"strings"
-
-	"github.com/d4l3k/go-bfloat16"
-	"github.com/pdevine/tensor"
-	"github.com/pdevine/tensor/native"
-	"github.com/x448/float16"
 
 	"github.com/ollama/ollama/llm"
 )
@@ -20,90 +11,12 @@ type MistralModel struct {
 	ModelData
 }
 
-func mistralLayerHandler(w io.Writer, r safetensorWriterTo, f *os.File) error {
-	layerSize := r.end - r.start
-
-	var err error
-	tData := make([]uint16, layerSize/2)
-	if err = binary.Read(f, r.bo, tData); err != nil {
-		return err
-	}
-
-	var heads uint32
-	if strings.Contains(r.t.Name, "attn_q") {
-		heads = uint32(r.params.AttentionHeads)
-	} else if strings.Contains(r.t.Name, "attn_k") {
-		heads = uint32(r.params.KeyValHeads)
-		if heads == 0 {
-			heads = uint32(r.params.AttentionHeads)
-		}
-	} else {
-		return fmt.Errorf("unknown layer type")
-	}
-
-	tData, err = repack(tData, int(heads), r.t.Shape)
-	if err != nil {
-		return err
-	}
-
-	var buf []byte
-	for _, n := range tData {
-		buf = r.bo.AppendUint16(buf, n)
-	}
-
-	tempBuf := make([]uint16, len(tData))
-	tDataF32 := bfloat16.DecodeFloat32(buf)
-	for cnt, v := range tDataF32 {
-		tDataF16 := float16.Fromfloat32(v)
-		tempBuf[cnt] = uint16(tDataF16)
-	}
-
-	if err = binary.Write(w, r.bo, tempBuf); err != nil {
-		return err
-	}
-	return nil
-}
-
-func repack(data []uint16, heads int, shape []uint64) ([]uint16, error) {
-	n := tensor.New(tensor.WithShape(int(shape[0]), int(shape[1])), tensor.WithBacking(data))
-	origShape := n.Shape().Clone()
-
-	// reshape the tensor and swap axes 1 and 2 to unpack the layer for gguf
-	if err := n.Reshape(heads, 2, origShape[0]/heads/2, origShape[1]); err != nil {
-		return nil, err
-	}
-
-	if err := n.T(0, 2, 1, 3); err != nil {
-		return nil, err
-	}
-
-	if err := n.Reshape(origShape...); err != nil {
-		return nil, err
-	}
-
-	if err := n.Transpose(); err != nil {
-		return nil, err
-	}
-	newN, err := native.SelectU16(n, 1)
-	if err != nil {
-		return nil, err
-	}
-
-	var fullTensor []uint16
-	for _, v := range newN {
-		fullTensor = append(fullTensor, v...)
-	}
-	return fullTensor, nil
-}
-
 func (m *MistralModel) GetTensors() error {
 	t, err := m.Format.GetTensors(m.Path, m.Params)
 	if err != nil {
 		return err
 	}
 
-	m.Tensors = []llm.Tensor{}
-
 	pattern := `^blk\.[0-9]+\.attn_(?P<layer>q|k)\.weight$`
 	re, err := regexp.Compile(pattern)
 	if err != nil {
@@ -114,7 +27,7 @@ func (m *MistralModel) GetTensors() error {
 		matches := re.FindAllStringSubmatch(l.Name, -1)
 		if len(matches) > 0 {
 			wt := l.WriterTo.(safetensorWriterTo)
-			wt.handler = mistralLayerHandler
+			wt.repacker = m.Repack
 			l.WriterTo = wt
 		}
 		m.Tensors = append(m.Tensors, l)
@@ -160,3 +73,7 @@ func (m *MistralModel) WriteGGUF(ws io.WriteSeeker) error {
 
 	return llm.NewGGUFV3(m.Params.ByteOrder).Encode(ws, kv, m.Tensors)
 }
+
+func (m *MistralModel) Repack(name string, data []float32, shape []uint64) ([]float32, error) {
+	return llamaRepack(name, m.Params, data, shape)
+}

convert/mixtral.go

@@ -17,8 +17,6 @@ func (m *MixtralModel) GetTensors() error {
 		return err
 	}
 
-	m.Tensors = []llm.Tensor{}
-
 	pattern := `^blk\.[0-9]+\.attn_(?P<layer>q|k)\.weight$`
 	re, err := regexp.Compile(pattern)
 	if err != nil {
@@ -29,7 +27,7 @@ func (m *MixtralModel) GetTensors() error {
 		matches := re.FindAllStringSubmatch(l.Name, -1)
 		if len(matches) > 0 {
 			wt := l.WriterTo.(safetensorWriterTo)
-			wt.handler = mistralLayerHandler
+			wt.repacker = m.Repack
 			l.WriterTo = wt
 		}
 		m.Tensors = append(m.Tensors, l)
@@ -83,3 +81,7 @@ func (m *MixtralModel) WriteGGUF(ws io.WriteSeeker) error {
 
 	return llm.NewGGUFV3(m.Params.ByteOrder).Encode(ws, kv, m.Tensors)
 }
+
+func (m *MixtralModel) Repack(name string, data []float32, shape []uint64) ([]float32, error) {
+	return llamaRepack(name, m.Params, data, shape)
+}

convert/safetensors.go

@@ -6,14 +6,13 @@ import (
 	"encoding/json"
 	"fmt"
 	"io"
-	"log/slog"
 	"os"
 	"path/filepath"
 	"regexp"
 	"slices"
+	"strings"
 
 	"github.com/d4l3k/go-bfloat16"
-	"github.com/mitchellh/mapstructure"
 	"github.com/x448/float16"
 
 	"github.com/ollama/ollama/llm"
@@ -26,39 +25,38 @@ type safetensorWriterTo struct {
 	bo     ByteOrder
 
 	filename string
+	dtype    string
 
-	start, end, padding uint64
+	offset, size int64
-	handler             func(w io.Writer, r safetensorWriterTo, f *os.File) error
+	repacker     func(string, []float32, []uint64) ([]float32, error)
 }
 
-type tensorMetaData struct {
+type safetensorMetadata struct {
-	Type    string `mapstructure:"dtype"`
+	Type    string   `json:"dtype"`
-	Shape   []int  `mapstructure:"shape"`
+	Shape   []uint64 `json:"shape"`
-	Offsets []int  `mapstructure:"data_offsets"`
+	Offsets []int64  `json:"data_offsets"`
 }
 
 type SafetensorFormat struct{}
 
 func (m *SafetensorFormat) GetTensors(dirpath string, params *Params) ([]llm.Tensor, error) {
-	slog.Debug("getting tensor data")
 	var tensors []llm.Tensor
-	files, err := filepath.Glob(filepath.Join(dirpath, "/model-*.safetensors"))
+	matches, err := filepath.Glob(filepath.Join(dirpath, "*.safetensors"))
 	if err != nil {
 		return nil, err
 	}
 
 	var offset uint64
-	for _, f := range files {
+	for _, f := range matches {
 		var t []llm.Tensor
 		var err error
 		t, offset, err = m.readTensors(f, offset, params)
 		if err != nil {
-			slog.Error(err.Error())
 			return nil, err
 		}
+
 		tensors = append(tensors, t...)
 	}
-	slog.Debug(fmt.Sprintf("all tensors = %d", len(tensors)))
 	return tensors, nil
 }
 
@@ -69,70 +67,57 @@ func (m *SafetensorFormat) readTensors(fn string, offset uint64, params *Params)
 	}
 	defer f.Close()
 
-	var jsonSize uint64
+	var n int64
-	if err := binary.Read(f, binary.LittleEndian, &jsonSize); err != nil {
+	if err := binary.Read(f, binary.LittleEndian, &n); err != nil {
 		return nil, 0, err
 	}
 
-	buf := make([]byte, jsonSize)
+	b := bytes.NewBuffer(make([]byte, 0, n))
-	_, err = io.ReadFull(f, buf)
+	if _, err = io.CopyN(b, f, n); err != nil {
-	if err != nil {
 		return nil, 0, err
 	}
 
-	d := json.NewDecoder(bytes.NewBuffer(buf))
+	var headers map[string]safetensorMetadata
-	d.UseNumber()
+	if err := json.NewDecoder(b).Decode(&headers); err != nil {
-	var parsed map[string]interface{}
-	if err = d.Decode(&parsed); err != nil {
 		return nil, 0, err
 	}
 
 	var keys []string
-	for k := range parsed {
+	for key := range headers {
-		keys = append(keys, k)
+		if !strings.HasSuffix(key, "self_attn.rotary_embd.inv_freq") {
+			keys = append(keys, key)
+		}
 	}
 
 	slices.Sort(keys)
-	slog.Info("converting layers")
 
 	var tensors []llm.Tensor
-	for _, k := range keys {
+	for _, key := range keys {
-		vals := parsed[k].(map[string]interface{})
+		value := headers[key]
-		var data tensorMetaData
-		if err = mapstructure.Decode(vals, &data); err != nil {
-			slog.Error("couldn't decode properly")
-			return nil, 0, err
-		}
 
-		var size uint64
 		var kind uint32
-		switch len(data.Shape) {
+		switch len(value.Shape) {
 		case 0:
-			// metadata
+			// valuedata
 			continue
-		case 1:
-			// convert to float32
-			kind = 0
-			size = uint64(data.Shape[0] * 4)
 		case 2:
-			// convert to float16
 			kind = 1
-			size = uint64(data.Shape[0] * data.Shape[1] * 2)
 		}
 
-		ggufName, err := m.GetLayerName(k)
+		name, err := m.GetLayerName(key)
 		if err != nil {
-			slog.Error(err.Error())
 			return nil, 0, err
 		}
 
-		shape := []uint64{0, 0, 0, 0}
+		shape := make([]uint64, len(value.Shape))
-		for i := range data.Shape {
+		copy(shape, value.Shape)
-			shape[i] = uint64(data.Shape[i])
+
+		pad := func(s int64) int64 {
+			return 8 + n + s
 		}
 
 		t := llm.Tensor{
-			Name:   ggufName,
+			Name:   name,
 			Kind:   kind,
 			Offset: offset,
 			Shape:  shape[:],
@@ -143,18 +128,15 @@ func (m *SafetensorFormat) readTensors(fn string, offset uint64, params *Params)
 			params:   params,
 			bo:       params.ByteOrder,
 			filename: fn,
-			start:    uint64(data.Offsets[0]),
+			dtype:    value.Type,
-			end:      uint64(data.Offsets[1]),
+			offset:   pad(value.Offsets[0]),
-			padding:  8 + jsonSize,
+			size:     pad(value.Offsets[1]) - pad(value.Offsets[0]),
 		}
 
-		offset += size
+		offset += t.Size()
 		tensors = append(tensors, t)
 	}
 
-	slog.Debug(fmt.Sprintf("total tensors for file = %d", len(tensors)))
-	slog.Debug(fmt.Sprintf("offset = %d", offset))
-
 	return tensors, offset, nil
 }
 
@@ -167,9 +149,7 @@ func (m *SafetensorFormat) GetParams(dirpath string) (*Params, error) {
 
 	var params Params
 
-	d := json.NewDecoder(f)
+	if err := json.NewDecoder(f).Decode(&params); err != nil {
-	err = d.Decode(&params)
-	if err != nil {
 		return nil, err
 	}
 
@@ -224,55 +204,58 @@ func (r safetensorWriterTo) WriteTo(w io.Writer) (n int64, err error) {
 	}
 	defer f.Close()
 
-	if _, err = f.Seek(int64(r.padding+r.start), 0); err != nil {
+	if _, err = f.Seek(r.offset, io.SeekStart); err != nil {
 		return 0, err
 	}
 
-	// use the handler if one is present
+	var f32s []float32
-	if r.handler != nil {
+	switch r.dtype {
-		return 0, r.handler(w, r, f)
+	case "F32":
-	}
+		f32s = make([]float32, r.size/4)
-
+		if err = binary.Read(f, r.bo, f32s); err != nil {
-	remaining := r.end - r.start
+			return 0, err
-
+		}
-	bufSize := uint64(10240)
+	case "F16":
-	var finished bool
+		u16s := make([]uint16, r.size/2)
-	for {
+		if err = binary.Read(f, r.bo, u16s); err != nil {
-		data := make([]byte, min(bufSize, remaining))
-
-		b, err := io.ReadFull(f, data)
-		remaining -= uint64(b)
-
-		if err == io.EOF || remaining <= 0 {
-			finished = true
-		} else if err != nil {
 			return 0, err
 		}
 
-		// convert bfloat16 -> ieee float32
+		for _, b := range u16s {
-		tDataF32 := bfloat16.DecodeFloat32(data)
+			f32s = append(f32s, float16.Frombits(b).Float32())
-
-		switch r.t.Kind {
-		case 0:
-			if err := binary.Write(w, r.bo, tDataF32); err != nil {
-				return 0, err
-			}
-		case 1:
-			// convert float32 -> float16
-			tempBuf := make([]uint16, len(data)/2)
-			for cnt, v := range tDataF32 {
-				tDataF16 := float16.Fromfloat32(v)
-				tempBuf[cnt] = uint16(tDataF16)
-			}
-			if err := binary.Write(w, r.bo, tempBuf); err != nil {
-				return 0, err
-			}
 		}
-		if finished {
+
-			break
+	case "BF16":
+		u8s := make([]uint8, r.size)
+		if err = binary.Read(f, r.bo, u8s); err != nil {
+			return 0, err
+		}
+
+		f32s = bfloat16.DecodeFloat32(u8s)
+	default:
+		return 0, fmt.Errorf("unknown data type: %s", r.dtype)
+	}
+
+	if r.repacker != nil {
+		f32s, err = r.repacker(r.t.Name, f32s, r.t.Shape)
+		if err != nil {
+			return 0, err
 		}
 	}
-	return 0, nil
+
+	switch r.t.Kind {
+	case 0:
+		return 0, binary.Write(w, r.bo, f32s)
+	case 1:
+		f16s := make([]uint16, len(f32s))
+		for i := range f32s {
+			f16s[i] = float16.Fromfloat32(f32s[i]).Bits()
+		}
+
+		return 0, binary.Write(w, r.bo, f16s)
+	default:
+		return 0, fmt.Errorf("unknown storage type: %d", r.t.Kind)
+	}
 }
 
 func (m *SafetensorFormat) GetModelArch(name, dirPath string, params *Params) (ModelArch, error) {
@@ -281,6 +264,15 @@ func (m *SafetensorFormat) GetModelArch(name, dirPath string, params *Params) (M
 		return nil, fmt.Errorf("No architecture specified to convert")
 	case 1:
 		switch params.Architectures[0] {
+		case "LlamaForCausalLM":
+			return &LlamaModel{
+				ModelData{
+					Name:   name,
+					Path:   dirPath,
+					Params: params,
+					Format: m,
+				},
+			}, nil
 		case "MistralForCausalLM":
 			return &MistralModel{
 				ModelData{

convert/tokenizer.go

@@ -0,0 +1,109 @@
+package convert
+
+import (
+	"cmp"
+	"crypto/sha256"
+	"encoding/json"
+	"fmt"
+	"log/slog"
+	"os"
+	"slices"
+
+	"golang.org/x/exp/maps"
+)
+
+type Tokenizer struct {
+	Version     string         `json:"version"`
+	AddedTokens []Token        `json:"added_tokens"`
+	Model       TokenizerModel `json:"model"`
+
+	PreTokenizer struct {
+		PreTokenizers []struct {
+			Type    string `json:"type"`
+			Pattern struct {
+				Regex string `json:"Regex"`
+			} `json:"pattern"`
+		} `json:"pretokenizers"`
+	} `json:"pre_tokenizer"`
+}
+
+type TokenizerModel struct {
+	Type   string         `json:"type"`
+	Vocab  map[string]int `json:"vocab"`
+	Merges []string       `json:"merges"`
+	Tokens []Token
+}
+
+type Token struct {
+	ID          int    `json:"id"`
+	Content     string `json:"content"`
+	Special     bool   `json:"special"`
+	UserDefined bool
+}
+
+func (t *Token) Type() int32 {
+	switch {
+	case t.Special:
+		return tokenTypeControl
+	case t.UserDefined:
+		return tokenTypeUserDefined
+	default:
+		return tokenTypeNormal
+	}
+}
+
+func (t *Tokenizer) maxID() int {
+	return max(
+		slices.Max(maps.Values(t.Model.Vocab)),
+		slices.MaxFunc(t.AddedTokens, func(a, b Token) int {
+			return cmp.Compare(a.ID, b.ID)
+		}).ID,
+	)
+}
+
+func parseTokens(dirpath string) (pre string, tokens []Token, merges []string, err error) {
+	f, err := os.Open(dirpath)
+	if err != nil {
+		panic(err)
+	}
+	defer f.Close()
+
+	var t Tokenizer
+	if err := json.NewDecoder(f).Decode(&t); err != nil {
+		return "", nil, nil, err
+	}
+
+	tokens = make([]Token, t.maxID()+1)
+	for k, v := range t.Model.Vocab {
+		tokens[v] = Token{ID: v, Content: k, Special: false, UserDefined: false}
+	}
+
+	for _, v := range t.AddedTokens {
+		v.UserDefined = true
+		tokens[v.ID] = v
+	}
+
+	sha256sum := sha256.New()
+	for _, pt := range t.PreTokenizer.PreTokenizers {
+		switch pt.Type {
+		case "Split":
+			if pt.Pattern.Regex != "" {
+				sha256sum.Write([]byte(pt.Pattern.Regex))
+			}
+		}
+	}
+
+	switch digest := fmt.Sprintf("%x", sha256sum.Sum(nil)); digest {
+	case "d98f9631be1e9607a9848c26c1f9eac1aa9fc21ac6ba82a2fc0741af9780a48f":
+		pre = "llama-bpe"
+	case "03df5c5863ad70781dcfdef491ead25140f895fe8010964be0daefe27be32b02":
+		pre = "deepseek-llm"
+	case "21cde974d587f0d54dc8d56b183cc1e6239600172035c68fbd6d4b9f8da0576e":
+		pre = "deepseek-coder"
+	default:
+		slog.Warn("unknown pretokenizer, using default", "digest", digest)
+		pre = "default"
+	}
+
+	return pre, tokens, t.Model.Merges, nil
+}

convert/torch.go

@@ -24,8 +24,8 @@ type torchWriterTo struct {
 	params *Params
 	bo     ByteOrder
 
-	storage pytorch.StorageInterface
+	storage  pytorch.StorageInterface
-	handler func(w io.Writer, r torchWriterTo) error
+	repacker func(string, []float32, []uint64) ([]float32, error)
 }
 
 type TorchFormat struct{}
@@ -33,14 +33,14 @@ type TorchFormat struct{}
 func (tf *TorchFormat) GetTensors(dirpath string, params *Params) ([]llm.Tensor, error) {
 	slog.Debug("getting torch tensors")
 
-	files, err := filepath.Glob(filepath.Join(dirpath, "pytorch_model-*.bin"))
+	var files []string
-	if err != nil {
+	if pt, _ := filepath.Glob(filepath.Join(dirpath, "consolidated*.pth")); len(pt) > 0 {
-		slog.Error("didn't find any torch files")
+		files = append(files, pt...)
-		return nil, err
+	} else if pt, _ := filepath.Glob(filepath.Join(dirpath, "pytorch_model*.pth")); len(pt) > 0 {
+		files = append(files, pt...)
 	}
 
 	var offset uint64
-
 	var tensors []llm.Tensor
 	for _, fn := range files {
 		m, err := pytorch.Load(fn)
@@ -77,7 +77,7 @@ func (tf *TorchFormat) GetTensors(dirpath string, params *Params) ([]llm.Tensor,
 				slog.Error(err.Error())
 				return nil, err
 			}
-			slog.Debug(fmt.Sprintf("finding name for '%s' -> '%s'", k.(string), ggufName))
+			slog.Debug(fmt.Sprintf("'%35s': '%30s' %10d [%#v]", k.(string), ggufName, size, tshape))
 
 			shape := []uint64{0, 0, 0, 0}
 			for i := range tshape {
@@ -120,7 +120,7 @@ func getAltParams(dirpath string) (*Params, error) {
 		AttentionHeads int     `json:"n_heads"`
 		KeyValHeads    int     `json:"n_kv_heads"`
 		HiddenLayers   int     `json:"n_layers"`
-		RopeTheta      int     `json:"rope_theta"`
+		RopeTheta      float64 `json:"rope_theta"`
 		NormEPS        float64 `json:"norm_eps"`
 	}
 
@@ -133,6 +133,7 @@ func getAltParams(dirpath string) (*Params, error) {
 	}
 
 	params := &Params{
+		Architectures:  []string{"LlamaForCausalLM"},
 		HiddenSize:     tparams.HiddenSize,
 		AttentionHeads: tparams.AttentionHeads,
 		KeyValHeads:    tparams.KeyValHeads,
@@ -229,37 +230,38 @@ func (m *TorchFormat) GetLayerName(n string) (string, error) {
 }
 
 func (r torchWriterTo) WriteTo(w io.Writer) (n int64, err error) {
-	// use the handler if one is present
+	var f32s []float32
-	if r.handler != nil {
+	switch s := r.storage.(type) {
-		return 0, r.handler(w, r)
+	case *pytorch.FloatStorage:
+		f32s = s.Data
+	case *pytorch.HalfStorage:
+		f32s = s.Data
+	case *pytorch.BFloat16Storage:
+		f32s = s.Data
+	default:
+		return 0, fmt.Errorf("unknown data type: %T", s)
 	}
 
-	switch r.storage.(type) {
+	if r.repacker != nil {
-	case *pytorch.FloatStorage:
+		f32s, err = r.repacker(r.t.Name, f32s, r.t.Shape)
-		slog.Warn(fmt.Sprintf("unexpected storage found for layer '%s'; skipping", r.t.Name))
+		if err != nil {
-		return 0, nil
+			return 0, err
-	case *pytorch.HalfStorage:
-		switch r.t.Kind {
-		case 0:
-			data := r.storage.(*pytorch.HalfStorage).Data
-			slog.Debug(fmt.Sprintf("%35s F32 (%d)", r.t.Name, len(data)))
-			if err := binary.Write(w, r.bo, data); err != nil {
-				return 0, err
-			}
-		case 1:
-			data := r.storage.(*pytorch.HalfStorage).Data
-			tData := make([]uint16, len(data))
-			for cnt, v := range data {
-				tData[cnt] = uint16(float16.Fromfloat32(v))
-			}
-			slog.Debug(fmt.Sprintf("%35s F16 (%d)", r.t.Name, len(tData)))
-			if err := binary.Write(w, r.bo, tData); err != nil {
-				return 0, err
-			}
 		}
 	}
 
-	return 0, nil
+	switch r.t.Kind {
+	case 0:
+		return 0, binary.Write(w, r.bo, f32s)
+	case 1:
+		f16s := make([]uint16, len(f32s))
+		for i := range f32s {
+			f16s[i] = float16.Fromfloat32(f32s[i]).Bits()
+		}
+
+		return 0, binary.Write(w, r.bo, f16s)
+	default:
+		return 0, fmt.Errorf("unknown storage type: %d", r.t.Kind)
+	}
 }
 
 func (m *TorchFormat) GetModelArch(name, dirPath string, params *Params) (ModelArch, error) {

docs/api.md

@@ -797,9 +797,9 @@ curl http://localhost:11434/api/show -d '{
 
 ```json
 {
-  "modelfile": "# Modelfile generated by \"ollama show\"\n# To build a new Modelfile based on this one, replace the FROM line with:\n# FROM llava:latest\n\nFROM /Users/matt/.ollama/models/blobs/sha256:200765e1283640ffbd013184bf496e261032fa75b99498a9613be4e94d63ad52\nTEMPLATE \"\"\"{{ .System }}\nUSER: {{ .Prompt }}\nASSSISTANT: \"\"\"\nPARAMETER num_ctx 4096\nPARAMETER stop \"\u003c/s\u003e\"\nPARAMETER stop \"USER:\"\nPARAMETER stop \"ASSSISTANT:\"",
+  "modelfile": "# Modelfile generated by \"ollama show\"\n# To build a new Modelfile based on this one, replace the FROM line with:\n# FROM llava:latest\n\nFROM /Users/matt/.ollama/models/blobs/sha256:200765e1283640ffbd013184bf496e261032fa75b99498a9613be4e94d63ad52\nTEMPLATE \"\"\"{{ .System }}\nUSER: {{ .Prompt }}\nASSISTANT: \"\"\"\nPARAMETER num_ctx 4096\nPARAMETER stop \"\u003c/s\u003e\"\nPARAMETER stop \"USER:\"\nPARAMETER stop \"ASSISTANT:\"",
-  "parameters": "num_ctx                        4096\nstop                           \u003c/s\u003e\nstop                           USER:\nstop                           ASSSISTANT:",
+  "parameters": "num_ctx                        4096\nstop                           \u003c/s\u003e\nstop                           USER:\nstop                           ASSISTANT:",
-  "template": "{{ .System }}\nUSER: {{ .Prompt }}\nASSSISTANT: ",
+  "template": "{{ .System }}\nUSER: {{ .Prompt }}\nASSISTANT: ",
   "details": {
     "format": "gguf",
     "family": "llama",

docs/development.md

@@ -6,6 +6,8 @@ Install required tools:
 - go version 1.22 or higher
 - gcc version 11.4.0 or higher
 
+### MacOS
+
 ```bash
 brew install go cmake gcc
 ```

docs/faq.md

@@ -6,7 +6,7 @@ Ollama on macOS and Windows will automatically download updates. Click on the ta
 
 On Linux, re-run the install script:
 
-```
+```shell
 curl -fsSL https://ollama.com/install.sh | sh
 ```
 
@@ -30,7 +30,7 @@ To change this when using `ollama run`, use `/set parameter`:
 
 When using the API, specify the `num_ctx` parameter:
 
-```
+```shell
 curl http://localhost:11434/api/generate -d '{
   "model": "llama3",
   "prompt": "Why is the sky blue?",
@@ -40,6 +40,21 @@ curl http://localhost:11434/api/generate -d '{
 }'
 ```
 
+## How can I tell if my model was loaded onto the GPU?
+
+Use the `ollama ps` command to see what models are currently loaded into memory.
+
+```shell
+ollama ps
+NAME      	ID          	SIZE 	PROCESSOR	UNTIL
+llama3:70b	bcfb190ca3a7	42 GB	100% GPU 	4 minutes from now
+```
+
+The `Processor` column will show which memory the model was loaded in to:
+* `100% GPU` means the model was loaded entirely into the GPU
+* `100% CPU` means the model was loaded entirely in system memory
+* `48%/52% CPU/GPU` means the model was loaded partially onto both the GPU and into system memory
+
 ## How do I configure Ollama server?
 
 Ollama server can be configured with environment variables.
@@ -80,81 +95,19 @@ If Ollama is run as a systemd service, environment variables should be set using
 
 ### Setting environment variables on Windows
 
-On windows, Ollama inherits your user and system environment variables.
+On Windows, Ollama inherits your user and system environment variables.
 
-1. First Quit Ollama by clicking on it in the task bar
+1. First Quit Ollama by clicking on it in the task bar.
 
-2. Edit system environment variables from the control panel
+2. Start the Settings (Windows 11) or Control Panel (Windows 10) application and search for _environment variables_.
 
-3. Edit or create New variable(s) for your user account for `OLLAMA_HOST`, `OLLAMA_MODELS`, etc.
+3. Click on _Edit environment variables for your account_.
 
-4. Click OK/Apply to save
+4. Edit or create a new variable for your user account for `OLLAMA_HOST`, `OLLAMA_MODELS`, etc.
 
-5. Run `ollama` from a new terminal window
+5. Click OK/Apply to save.
 
-
+6. Start the Ollama application from the Windows Start menu.
-## How can I expose Ollama on my network?
-
-Ollama binds 127.0.0.1 port 11434 by default. Change the bind address with the `OLLAMA_HOST` environment variable.
-
-Refer to the section [above](#how-do-i-configure-ollama-server) for how to set environment variables on your platform.
-
-## How can I use Ollama with a proxy server?
-
-Ollama runs an HTTP server and can be exposed using a proxy server such as Nginx. To do so, configure the proxy to forward requests and optionally set required headers (if not exposing Ollama on the network). For example, with Nginx:
-
-```
-server {
-    listen 80;
-    server_name example.com;  # Replace with your domain or IP
-    location / {
-        proxy_pass http://localhost:11434;
-        proxy_set_header Host localhost:11434;
-    }
-}
-```
-
-## How can I use Ollama with ngrok?
-
-Ollama can be accessed using a range of tools for tunneling tools. For example with Ngrok:
-
-```
-ngrok http 11434 --host-header="localhost:11434"
-```
-
-## How can I use Ollama with Cloudflare Tunnel?
-
-To use Ollama with Cloudflare Tunnel, use the `--url` and `--http-host-header` flags:
-
-```
-cloudflared tunnel --url http://localhost:11434 --http-host-header="localhost:11434"
-```
-
-## How can I allow additional web origins to access Ollama?
-
-Ollama allows cross-origin requests from `127.0.0.1` and `0.0.0.0` by default. Additional origins can be configured with `OLLAMA_ORIGINS`.
-
-Refer to the section [above](#how-do-i-configure-ollama-server) for how to set environment variables on your platform.
-
-## Where are models stored?
-
-- macOS: `~/.ollama/models`
-- Linux: `/usr/share/ollama/.ollama/models`
-- Windows: `C:\Users\%username%\.ollama\models`
-
-### How do I set them to a different location?
-
-If a different directory needs to be used, set the environment variable `OLLAMA_MODELS` to the chosen directory.
-
-Refer to the section [above](#how-do-i-configure-ollama-server) for how to set environment variables on your platform.
-
-## Does Ollama send my prompts and answers back to ollama.com?
-
-No. Ollama runs locally, and conversation data does not leave your machine.
-
-## How can I use Ollama in Visual Studio Code?
-
-There is already a large collection of plugins available for VSCode as well as other editors that leverage Ollama. See the list of [extensions & plugins](https://github.com/ollama/ollama#extensions--plugins) at the bottom of the main repository readme.
 
 ## How do I use Ollama behind a proxy?
 
@@ -181,6 +134,69 @@ docker build -t ollama-with-ca .
 docker run -d -e HTTPS_PROXY=https://my.proxy.example.com -p 11434:11434 ollama-with-ca
 ```
 
+## Does Ollama send my prompts and answers back to ollama.com?
+
+No. Ollama runs locally, and conversation data does not leave your machine.
+
+## How can I expose Ollama on my network?
+
+Ollama binds 127.0.0.1 port 11434 by default. Change the bind address with the `OLLAMA_HOST` environment variable.
+
+Refer to the section [above](#how-do-i-configure-ollama-server) for how to set environment variables on your platform.
+
+## How can I use Ollama with a proxy server?
+
+Ollama runs an HTTP server and can be exposed using a proxy server such as Nginx. To do so, configure the proxy to forward requests and optionally set required headers (if not exposing Ollama on the network). For example, with Nginx:
+
+```
+server {
+    listen 80;
+    server_name example.com;  # Replace with your domain or IP
+    location / {
+        proxy_pass http://localhost:11434;
+        proxy_set_header Host localhost:11434;
+    }
+}
+```
+
+## How can I use Ollama with ngrok?
+
+Ollama can be accessed using a range of tools for tunneling tools. For example with Ngrok:
+
+```shell
+ngrok http 11434 --host-header="localhost:11434"
+```
+
+## How can I use Ollama with Cloudflare Tunnel?
+
+To use Ollama with Cloudflare Tunnel, use the `--url` and `--http-host-header` flags:
+
+```shell
+cloudflared tunnel --url http://localhost:11434 --http-host-header="localhost:11434"
+```
+
+## How can I allow additional web origins to access Ollama?
+
+Ollama allows cross-origin requests from `127.0.0.1` and `0.0.0.0` by default. Additional origins can be configured with `OLLAMA_ORIGINS`.
+
+Refer to the section [above](#how-do-i-configure-ollama-server) for how to set environment variables on your platform.
+
+## Where are models stored?
+
+- macOS: `~/.ollama/models`
+- Linux: `/usr/share/ollama/.ollama/models`
+- Windows: `C:\Users\%username%\.ollama\models`
+
+### How do I set them to a different location?
+
+If a different directory needs to be used, set the environment variable `OLLAMA_MODELS` to the chosen directory.
+
+Refer to the section [above](#how-do-i-configure-ollama-server) for how to set environment variables on your platform.
+
+## How can I use Ollama in Visual Studio Code?
+
+There is already a large collection of plugins available for VSCode as well as other editors that leverage Ollama. See the list of [extensions & plugins](https://github.com/ollama/ollama#extensions--plugins) at the bottom of the main repository readme.
+
 ## How do I use Ollama with GPU acceleration in Docker?
 
 The Ollama Docker container can be configured with GPU acceleration in Linux or Windows (with WSL2). This requires the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia-container-toolkit). See [ollama/ollama](https://hub.docker.com/r/ollama/ollama) for more details.
@@ -195,7 +211,7 @@ Open `Control Panel > Networking and Internet > View network status and tasks` a
 Click on `Configure` and open the `Advanced` tab. Search through each of the properties until you find `Large Send Offload Version 2 (IPv4)` and `Large Send Offload Version 2 (IPv6)`. *Disable* both of these
 properties.
 
-## How can I pre-load a model to get faster response times?
+## How can I preload a model into Ollama to get faster response times?
 
 If you are using the API you can preload a model by sending the Ollama server an empty request. This works with both the `/api/generate` and `/api/chat` API endpoints.
 
@@ -209,6 +225,11 @@ To use the chat completions endpoint, use:
 curl http://localhost:11434/api/chat -d '{"model": "mistral"}'
 ```
 
+To preload a model using the CLI, use the command:
+```shell
+ollama run llama3 ""
+```
+
 ## How do I keep a model loaded in memory or make it unload immediately?
 
 By default models are kept in memory for 5 minutes before being unloaded. This allows for quicker response times if you are making numerous requests to the LLM. You may, however, want to free up the memory before the 5 minutes have elapsed or keep the model loaded indefinitely. Use the `keep_alive` parameter with either the `/api/generate` and `/api/chat` API endpoints to control how long the model is left in memory.
@@ -233,8 +254,6 @@ Alternatively, you can change the amount of time all models are loaded into memo
 
 If you wish to override the `OLLAMA_KEEP_ALIVE` setting, use the `keep_alive` API parameter with the `/api/generate` or `/api/chat` API endpoints.
 
-## How do I manage the maximum number of requests the server can queue
+## How do I manage the maximum number of requests the Ollama server can queue?
 
-If too many requests are sent to the server, it will respond with a 503 error
+If too many requests are sent to the server, it will respond with a 503 error indicating the server is overloaded.  You can adjust how many requests may be queue by setting `OLLAMA_MAX_QUEUE`.
-indicating the server is overloaded.  You can adjust how many requests may be
-queue by setting `OLLAMA_MAX_QUEUE`

docs/troubleshooting.md

@@ -1,104 +1,86 @@
-# How to troubleshoot issues
+# How to troubleshoot issues
-
+
-Sometimes Ollama may not perform as expected. One of the best ways to figure out what happened is to take a look at the logs. Find the logs on **Mac** by running the command:
+Sometimes Ollama may not perform as expected. One of the best ways to figure out what happened is to take a look at the logs. Find the logs on **Mac** by running the command:
-
+
-```shell
+```shell
-cat ~/.ollama/logs/server.log
+cat ~/.ollama/logs/server.log
-```
+```
-
+
-On **Linux** systems with systemd, the logs can be found with this command:
+On **Linux** systems with systemd, the logs can be found with this command:
-
+
-```shell
+```shell
-journalctl -u ollama
+journalctl -u ollama
-```
+```
-
+
-When you run Ollama in a **container**, the logs go to stdout/stderr in the container:
+When you run Ollama in a **container**, the logs go to stdout/stderr in the container:
-
+
-```shell
+```shell
-docker logs <container-name>
+docker logs <container-name>
-```
+```
-(Use `docker ps` to find the container name)
+(Use `docker ps` to find the container name)
-
+
-If manually running `ollama serve` in a terminal, the logs will be on that terminal.
+If manually running `ollama serve` in a terminal, the logs will be on that terminal.
-
+
-When you run Ollama on **Windows**, there are a few different locations.  You can view them in the explorer window by hitting `<cmd>+R` and type in:
+When you run Ollama on **Windows**, there are a few different locations. You can view them in the explorer window by hitting `<cmd>+R` and type in:
-- `explorer %LOCALAPPDATA%\Ollama` to view logs
+- `explorer %LOCALAPPDATA%\Ollama` to view logs
-- `explorer %LOCALAPPDATA%\Programs\Ollama` to browse the binaries (The installer adds this to your user PATH)
+- `explorer %LOCALAPPDATA%\Programs\Ollama` to browse the binaries (The installer adds this to your user PATH)
-- `explorer %HOMEPATH%\.ollama` to browse where models and configuration is stored
+- `explorer %HOMEPATH%\.ollama` to browse where models and configuration is stored
-- `explorer %TEMP%` where temporary executable files are stored in one or more `ollama*` directories
+- `explorer %TEMP%` where temporary executable files are stored in one or more `ollama*` directories
-
+
-To enable additional debug logging to help troubleshoot problems, first **Quit the running app from the tray menu** then in a powershell terminal
+To enable additional debug logging to help troubleshoot problems, first **Quit the running app from the tray menu** then in a powershell terminal
-```powershell
+```powershell
-$env:OLLAMA_DEBUG="1"
+$env:OLLAMA_DEBUG="1"
-& "ollama app.exe"
+& "ollama app.exe"
-```
+```
-
+
-Join the [Discord](https://discord.gg/ollama) for help interpreting the logs.
+Join the [Discord](https://discord.gg/ollama) for help interpreting the logs.
-
+
-## LLM libraries
+## LLM libraries
-
+
-Ollama includes multiple LLM libraries compiled for different GPUs and CPU
+Ollama includes multiple LLM libraries compiled for different GPUs and CPU vector features. Ollama tries to pick the best one based on the capabilities of your system. If this autodetection has problems, or you run into other problems (e.g. crashes in your GPU) you can workaround this by forcing a specific LLM library. `cpu_avx2` will perform the best, followed by `cpu_avx` an the slowest but most compatible is `cpu`. Rosetta emulation under MacOS will work with the `cpu` library. 
-vector features.  Ollama tries to pick the best one based on the capabilities of
+
-your system.  If this autodetection has problems, or you run into other problems
+In the server log, you will see a message that looks something like this (varies from release to release):
-(e.g. crashes in your GPU) you can workaround this by forcing a specific LLM
+
-library.  `cpu_avx2` will perform the best, followed by `cpu_avx` an the slowest
+```
-but most compatible is `cpu`.  Rosetta emulation under MacOS will work with the
+Dynamic LLM libraries [rocm_v6 cpu cpu_avx cpu_avx2 cuda_v11 rocm_v5]
-`cpu` library. 
+```
-
+
-In the server log, you will see a message that looks something like this (varies
+**Experimental LLM Library Override**
-from release to release):
+
-
+You can set OLLAMA_LLM_LIBRARY to any of the available LLM libraries to bypass autodetection, so for example, if you have a CUDA card, but want to force the CPU LLM library with AVX2 vector support, use:
-```
+
-Dynamic LLM libraries [rocm_v6 cpu cpu_avx cpu_avx2 cuda_v11 rocm_v5]
+```
-```
+OLLAMA_LLM_LIBRARY="cpu_avx2" ollama serve
-
+```
-**Experimental LLM Library Override**
+
-
+You can see what features your CPU has with the following.
-You can set OLLAMA_LLM_LIBRARY to any of the available LLM libraries to bypass
+```
-autodetection, so for example, if you have a CUDA card, but want to force the
+cat /proc/cpuinfo| grep flags | head -1
-CPU LLM library with AVX2 vector support, use:
+```
-
+
-```
+## Installing older or pre-release versions on Linux
-OLLAMA_LLM_LIBRARY="cpu_avx2" ollama serve
+
-```
+If you run into problems on Linux and want to install an older version, or you'd like to try out a pre-release before it's officially released, you can tell the install script which version to install.
-
+
-You can see what features your CPU has with the following.  
+```sh
-```
+curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION="0.1.29" sh
-cat /proc/cpuinfo| grep flags  | head -1
+```
-```
+
-
+## Linux tmp noexec 
-## Installing older or pre-release versions on Linux
+
-
+If your system is configured with the "noexec" flag where Ollama stores its temporary executable files, you can specify an alternate location by setting OLLAMA_TMPDIR to a location writable by the user ollama runs as. For example OLLAMA_TMPDIR=/usr/share/ollama/
-If you run into problems on Linux and want to install an older version, or you'd
+
-like to try out a pre-release before it's officially released, you can tell the
+## Container fails to run on NVIDIA GPU
-install script which version to install.
+
-
+Make sure you've set up the container runtime first as described in [docker.md](./docker.md)
-```sh
+
-curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION="0.1.29" sh
+Sometimes the container runtime can have difficulties initializing the GPU. When you check the server logs, this can show up as various error codes, such as "3" (not initialized), "46" (device unavailable), "100" (no device), "999" (unknown), or others. The following troubleshooting techniques may help resolve the problem
-```
+
-
+- Is the uvm driver not loaded? `sudo nvidia-modprobe -u`
-## Linux tmp noexec 
+- Try reloading the nvidia_uvm driver - `sudo rmmod nvidia_uvm` then `sudo modprobe nvidia_uvm`
-
+- Try rebooting
-If your system is configured with the "noexec" flag where Ollama stores its
+- Make sure you're running the latest nvidia drivers
-temporary executable files, you can specify an alternate location by setting
+
-OLLAMA_TMPDIR to a location writable by the user ollama runs as.  For example
+If none of those resolve the problem, gather additional information and file an issue:
-OLLAMA_TMPDIR=/usr/share/ollama/
+- Set `CUDA_ERROR_LEVEL=50` and try again to get more diagnostic logs
-
+- Check dmesg for any errors `sudo dmesg | grep -i nvrm` and `sudo dmesg | grep -i nvidia`
-## Container fails to run on NVIDIA GPU
-
-Make sure you've set up the conatiner runtime first as described in [docker.md](./docker.md)
-
-Sometimes the container runtime can have difficulties initializing the GPU.
-When you check the server logs, this can show up as various error codes, such
-as "3" (not initialized), "46" (device unavailable), "100" (no device), "999"
-(unknown), or others.  The following troubleshooting techniques may help resolve
-the problem
-
-- Is the uvm driver not loaded? `sudo nvidia-modprobe -u`
-- Try reloading the nvidia_uvm driver - `sudo rmmod nvidia_uvm` then `sudo modprobe nvidia_uvm`
-- Try rebooting
-- Make sure you're running the latest nvidia drivers
-
-If none of those resolve the problem, gather additional information and file an issue:
-- Set `CUDA_ERROR_LEVEL=50` and try again to get more diagnostic logs
-- Check dmesg for any errors `sudo dmesg | grep -i nvrm` and `sudo dmesg | grep -i nvidia`

docs/windows.md

@@ -33,7 +33,7 @@ Here's a quick example showing API access from `powershell`
 ## Troubleshooting
 
 While we're in preview, `OLLAMA_DEBUG` is always enabled, which adds
-a "view logs" menu item to the app, and increses logging for the GUI app and
+a "view logs" menu item to the app, and increases logging for the GUI app and
 server.
 
 Ollama on Windows stores files in a few different locations.  You can view them in

format/time.go

@@ -60,7 +60,9 @@ func humanTime(t time.Time, zeroValue string) string {
 	}
 
 	delta := time.Since(t)
-	if delta < 0 {
+	if int(delta.Hours())/24/365 < -20 {
+		return "Forever"
+	} else if delta < 0 {
 		return humanDuration(-delta) + " from now"
 	}
 

format/time_test.go

@@ -32,4 +32,14 @@ func TestHumanTime(t *testing.T) {
 		v := now.Add(800 * time.Millisecond)
 		assertEqual(t, HumanTime(v, ""), "Less than a second from now")
 	})
+
+	t.Run("time way in the future", func(t *testing.T) {
+		v := now.Add(24 * time.Hour * 365 * 200)
+		assertEqual(t, HumanTime(v, ""), "Forever")
+	})
+
+	t.Run("time way in the future lowercase", func(t *testing.T) {
+		v := now.Add(24 * time.Hour * 365 * 200)
+		assertEqual(t, HumanTimeLower(v, ""), "forever")
+	})
 }

go.mod

@@ -4,12 +4,10 @@ go 1.22.0
 
 require (
 	github.com/containerd/console v1.0.3
-	github.com/d4l3k/go-bfloat16 v0.0.0-20211005043715-690c3bdd05f1
 	github.com/emirpasic/gods v1.18.1
 	github.com/gin-gonic/gin v1.10.0
 	github.com/golang/protobuf v1.5.4 // indirect
 	github.com/google/uuid v1.1.2
-	github.com/mitchellh/mapstructure v1.5.0
 	github.com/olekukonko/tablewriter v0.0.5
 	github.com/spf13/cobra v1.7.0
 	github.com/stretchr/testify v1.9.0
@@ -18,6 +16,8 @@ require (
 )
 
 require (
+	github.com/d4l3k/go-bfloat16 v0.0.0-20211005043715-690c3bdd05f1
+	github.com/mattn/go-runewidth v0.0.14
 	github.com/nlpodyssey/gopickle v0.3.0
 	github.com/pdevine/tensor v0.0.0-20240510204454-f88f4562727c
 )
@@ -33,7 +33,6 @@ require (
 	github.com/gogo/protobuf v1.3.2 // indirect
 	github.com/google/flatbuffers v24.3.25+incompatible // indirect
 	github.com/kr/text v0.2.0 // indirect
-	github.com/mattn/go-runewidth v0.0.14 // indirect
 	github.com/pkg/errors v0.9.1 // indirect
 	github.com/pmezard/go-difflib v1.0.0 // indirect
 	github.com/rivo/uniseg v0.2.0 // indirect

go.sum

@@ -135,8 +135,6 @@ github.com/mattn/go-isatty v0.0.20/go.mod h1:W+V8PltTTMOvKvAeJH7IuucS94S2C6jfK/D
 github.com/mattn/go-runewidth v0.0.9/go.mod h1:H031xJmbD/WCDINGzjvQ9THkh0rPKHF+m2gUSrubnMI=
 github.com/mattn/go-runewidth v0.0.14 h1:+xnbZSEeDbOIg5/mE6JF0w6n9duR1l3/WmbinWVwUuU=
 github.com/mattn/go-runewidth v0.0.14/go.mod h1:Jdepj2loyihRzMpdS35Xk/zdY8IAYHsh153qUoGf23w=
-github.com/mitchellh/mapstructure v1.5.0 h1:jeMsZIYE/09sWLaz43PL7Gy6RuMjD2eJVyuac5Z2hdY=
-github.com/mitchellh/mapstructure v1.5.0/go.mod h1:bFUtVrKA4DC2yAKiSyO/QUcy7e+RRV2QTWOzhPopBRo=
 github.com/modern-go/concurrent v0.0.0-20180228061459-e0a39a4cb421/go.mod h1:6dJC0mAP4ikYIbvyc7fijjWJddQyLn8Ig3JB5CqoB9Q=
 github.com/modern-go/concurrent v0.0.0-20180306012644-bacd9c7ef1dd h1:TRLaZ9cD/w8PVh93nsPXa1VrQ6jlwL5oN8l14QlcNfg=
 github.com/modern-go/concurrent v0.0.0-20180306012644-bacd9c7ef1dd/go.mod h1:6dJC0mAP4ikYIbvyc7fijjWJddQyLn8Ig3JB5CqoB9Q=

gpu/amd_linux.go

@@ -159,10 +159,10 @@ func AMDGetGPUInfo() []GpuInfo {
 			return []GpuInfo{}
 		}
 
-		if int(major) < RocmComputeMin {
+		//if int(major) < RocmComputeMin {
-			slog.Warn(fmt.Sprintf("amdgpu too old gfx%d%x%x", major, minor, patch), "gpu", gpuID)
+		//	slog.Warn(fmt.Sprintf("amdgpu too old gfx%d%x%x", major, minor, patch), "gpu", gpuID)
-			continue
+		//	continue
-		}
+		//}
 
 		// Look up the memory for the current node
 		totalMemory := uint64(0)

gpu/amd_windows.go

@@ -86,10 +86,10 @@ func AMDGetGPUInfo() []GpuInfo {
 		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!?
-		if strings.EqualFold(name, iGPUName) {
+		//if strings.EqualFold(name, iGPUName) {
-			slog.Info("unsupported Radeon iGPU detected skipping", "id", i, "name", name, "gfx", gfx)
+		//	slog.Info("unsupported Radeon iGPU detected skipping", "id", i, "name", name, "gfx", gfx)
-			continue
+		//	continue
-		}
+		//}
 		if gfxOverride == "" {
 			if !slices.Contains[[]string, string](supported, gfx) {
 				slog.Warn("amdgpu is not supported", "gpu", i, "gpu_type", gfx, "library", libDir, "supported_types", supported)
@@ -108,10 +108,10 @@ func AMDGetGPUInfo() []GpuInfo {
 		}
 
 		// iGPU detection, remove this check once we can support an iGPU variant of the rocm library
-		if totalMemory < IGPUMemLimit {
+		//if totalMemory < IGPUMemLimit {
-			slog.Info("amdgpu appears to be an iGPU, skipping", "gpu", i, "total", format.HumanBytes2(totalMemory))
+		//	slog.Info("amdgpu appears to be an iGPU, skipping", "gpu", i, "total", format.HumanBytes2(totalMemory))
-			continue
+		//	continue
-		}
+		//}
 
 		// TODO revisit this once ROCm v6 is available on windows.
 		// v5.7 only reports VRAM used by this process, so it's completely wrong and unusable

integration/max_queue_test.go

@@ -19,6 +19,11 @@ import (
 )
 
 func TestMaxQueue(t *testing.T) {
+	if os.Getenv("OLLAMA_TEST_EXISTING") != "" {
+		t.Skip("Max Queue test requires spawing a local server so we can adjust the queue size")
+		return
+	}
+
 	// Note: This test can be quite slow when running in CPU mode, so keep the threadCount low unless your on GPU
 	// Also note that by default Darwin can't sustain > ~128 connections without adjusting limits
 	threadCount := 32
@@ -109,9 +114,9 @@ func TestMaxQueue(t *testing.T) {
 	slog.Info("generate done, waiting for embeds")
 	embedwg.Wait()
 
+	slog.Info("embeds completed", "success", succesCount, "busy", busyCount, "reset", resetByPeerCount, "canceled", canceledCount)
 	require.Equal(t, resetByPeerCount, 0, "Connections reset by peer, have you updated your fd and socket limits?")
 	require.True(t, busyCount > 0, "no requests hit busy error but some should have")
 	require.True(t, canceledCount == 0, "no requests should have been canceled due to timeout")
 
-	slog.Info("embeds completed", "success", succesCount, "busy", busyCount, "reset", resetByPeerCount, "canceled", canceledCount)
 }

llm/ext_server/server.cpp

@@ -334,6 +334,7 @@ struct server_metrics {
 struct llama_server_context
 {
     llama_model *model = nullptr;
+    float modelProgress = 0.0;
     llama_context *ctx = nullptr;
 
     clip_ctx *clp_ctx = nullptr;
@@ -737,7 +738,7 @@ struct llama_server_context
                     sampler_names.emplace_back(sampler_name);
                 }
             }
-            slot->sparams.samplers_sequence = sampler_types_from_names(sampler_names, false);
+            slot->sparams.samplers_sequence = llama_sampling_types_from_names(sampler_names, false);
         }
         else
         {
@@ -1095,7 +1096,7 @@ struct llama_server_context
         std::vector<std::string> samplers_sequence;
         for (const auto &sampler_type : slot.sparams.samplers_sequence)
         {
-            samplers_sequence.emplace_back(sampler_type_to_name_string(sampler_type));
+            samplers_sequence.emplace_back(llama_sampling_type_to_str(sampler_type));
         }
 
         return json {
@@ -2104,6 +2105,7 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     printf("  --embedding               enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled");
     printf("  -np N, --parallel N       number of slots for process requests (default: %d)\n", params.n_parallel);
     printf("  -cb, --cont-batching      enable continuous batching (a.k.a dynamic batching) (default: disabled)\n");
+    printf("  -fa, --flash-attn         enable Flash Attention (default: %s)\n", params.flash_attn ? "enabled" : "disabled");
     printf("  -spf FNAME, --system-prompt-file FNAME\n");
     printf("                            set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications.\n");
     printf("  -ctk TYPE, --cache-type-k TYPE\n");
@@ -2501,7 +2503,8 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
         {
             params.use_mmap = false;
         }
-        else if (arg == "--numa") {
+        else if (arg == "--numa")
+        {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
@@ -2521,6 +2524,10 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
         {
             params.cont_batching = true;
         }
+        else if (arg == "-fa" || arg == "--flash-attn")
+        {
+            params.flash_attn = true;
+        }
         else if (arg == "-np" || arg == "--parallel")
         {
             if (++i >= argc)
@@ -2529,7 +2536,8 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
                 break;
             }
             params.n_parallel = std::stoi(argv[i]);
-        } else if (arg == "-n" || arg == "--n-predict")
+        }
+        else if (arg == "-n" || arg == "--n-predict")
         {
             if (++i >= argc)
             {
@@ -2537,7 +2545,8 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
                 break;
             }
             params.n_predict = std::stoi(argv[i]);
-        } else if (arg == "-spf" || arg == "--system-prompt-file")
+        }
+        else if (arg == "-spf" || arg == "--system-prompt-file")
         {
             if (++i >= argc)
             {
@@ -2771,6 +2780,12 @@ inline void signal_handler(int signal) {
     shutdown_handler(signal);
 }
 
+static bool update_load_progress(float progress, void *data)
+{
+    ((llama_server_context*)data)->modelProgress = progress;
+    return true;
+}
+
 #if defined(_WIN32)
 char* wchar_to_char(const wchar_t* wstr) {
     if (wstr == nullptr) return nullptr;
@@ -2876,7 +2891,9 @@ int main(int argc, char **argv) {
                 break;
             }
             case SERVER_STATE_LOADING_MODEL:
-                res.set_content(R"({"status": "loading model"})", "application/json");
+                char buf[128];
+                snprintf(&buf[0], 128, R"({"status": "loading model", "progress": %0.2f})", llama.modelProgress);
+                res.set_content(buf, "application/json");
                 res.status = 503; // HTTP Service Unavailable
                 break;
             case SERVER_STATE_ERROR:
@@ -3071,6 +3088,9 @@ int main(int argc, char **argv) {
             });
 
     // load the model
+    params.progress_callback = update_load_progress;
+    params.progress_callback_user_data = (void*)&llama;
+
     if (!llama.load_model(params))
     {
         state.store(SERVER_STATE_ERROR);

llm/filetype.go

@@ -27,8 +27,16 @@ const (
 	fileTypeIQ2_XXS
 	fileTypeIQ2_XS
 	fileTypeQ2_K_S
-	fileTypeQ3_K_XS
+	fileTypeIQ3_XS
 	fileTypeIQ3_XXS
+	fileTypeIQ1_S
+	fileTypeIQ4_NL
+	fileTypeIQ3_S
+	fileTypeIQ2_S
+	fileTypeIQ4_XS
+	fileTypeIQ2_M
+	fileTypeIQ1_M
+	fileTypeBF16
 
 	fileTypeUnknown
 )
@@ -75,10 +83,26 @@ func ParseFileType(s string) (fileType, error) {
 		return fileTypeIQ2_XS, nil
 	case "Q2_K_S":
 		return fileTypeQ2_K_S, nil
-	case "Q3_K_XS":
+	case "IQ3_XS":
-		return fileTypeQ3_K_XS, nil
+		return fileTypeIQ3_XS, nil
 	case "IQ3_XXS":
 		return fileTypeIQ3_XXS, nil
+	case "IQ1_S":
+		return fileTypeIQ1_S, nil
+	case "IQ4_NL":
+		return fileTypeIQ4_NL, nil
+	case "IQ3_S":
+		return fileTypeIQ3_S, nil
+	case "IQ2_S":
+		return fileTypeIQ2_S, nil
+	case "IQ4_XS":
+		return fileTypeIQ4_XS, nil
+	case "IQ2_M":
+		return fileTypeIQ2_M, nil
+	case "IQ1_M":
+		return fileTypeIQ1_M, nil
+	case "BF16":
+		return fileTypeBF16, nil
 	default:
 		return fileTypeUnknown, fmt.Errorf("unknown fileType: %s", s)
 	}
@@ -126,10 +150,26 @@ func (t fileType) String() string {
 		return "IQ2_XS"
 	case fileTypeQ2_K_S:
 		return "Q2_K_S"
-	case fileTypeQ3_K_XS:
+	case fileTypeIQ3_XS:
-		return "Q3_K_XS"
+		return "IQ3_XS"
 	case fileTypeIQ3_XXS:
 		return "IQ3_XXS"
+	case fileTypeIQ1_S:
+		return "IQ1_S"
+	case fileTypeIQ4_NL:
+		return "IQ4_NL"
+	case fileTypeIQ3_S:
+		return "IQ3_S"
+	case fileTypeIQ2_S:
+		return "IQ2_S"
+	case fileTypeIQ4_XS:
+		return "IQ4_XS"
+	case fileTypeIQ2_M:
+		return "IQ2_M"
+	case fileTypeIQ1_M:
+		return "IQ1_M"
+	case fileTypeBF16:
+		return "BF16"
 	default:
 		return "unknown"
 	}

llm/generate/gen_linux.sh

@@ -165,7 +165,7 @@ if [ -z "${CUDART_LIB_DIR}" ]; then
     CUDART_LIB_DIR="${CUDA_LIB_DIR}"
 fi
 
-if [ -d "${CUDA_LIB_DIR}" ]; then
+if [ -z "${OLLAMA_SKIP_CUDA_GENERATE}" -a -d "${CUDA_LIB_DIR}" ]; then
     echo "CUDA libraries detected - building dynamic CUDA library"
     init_vars
     CUDA_MAJOR=$(ls "${CUDA_LIB_DIR}"/libcudart.so.* | head -1 | cut -f3 -d. || true)
@@ -227,7 +227,7 @@ if [ -z "${CLBlast_DIR}" ]; then
     fi
 fi
 
-if [ -d "${ROCM_PATH}" ]; then
+if [ -z "${OLLAMA_SKIP_ROCM_GENERATE}" -a -d "${ROCM_PATH}" ]; then
     echo "ROCm libraries detected - building dynamic ROCm library"
     if [ -f ${ROCM_PATH}/lib/librocblas.so.*.*.????? ]; then
         ROCM_VARIANT=_v$(ls ${ROCM_PATH}/lib/librocblas.so.*.*.????? | cut -f5 -d. || true)

llm/ggla.go

@@ -119,7 +119,7 @@ func (llm *ggla) decode(rs io.ReadSeeker) error {
 
 		t.Offset = uint64(offset)
 
-		if _, err := rs.Seek(int64(t.size()), io.SeekCurrent); err != nil {
+		if _, err := rs.Seek(int64(t.Size()), io.SeekCurrent); err != nil {
 			return err
 		}
 

llm/ggml.go

@@ -106,7 +106,7 @@ type Layer map[string]*Tensor
 
 func (l Layer) size() (size uint64) {
 	for _, t := range l {
-		size += t.size()
+		size += t.Size()
 	}
 
 	return size
@@ -124,12 +124,12 @@ type Tensor struct {
 }
 
 func (t Tensor) blockSize() uint64 {
-	switch {
+	switch t.Kind {
-	case t.Kind < 2:
+	case 0, 1, 24, 25, 26, 27, 28, 31: // F32, F16, I8, I16, I32, I64, F64, BF16
 		return 1
-	case t.Kind < 10:
+	case 2, 3, 8, 9, 20: // Q4_0, Q4_1, Q8_0, Q8_1, IQ4_NL
 		return 32
-	default:
+	default: // All others
 		return 256
 	}
 }
@@ -171,7 +171,29 @@ func (t Tensor) typeSize() uint64 {
 	case 17: // IQ2_XS
 		return 2 + 2*blockSize/8 + blockSize/32
 	case 18: // IQ3_XXS
-		return 2 + 3*blockSize/8
+		return 2 + blockSize/4 + blockSize/8
+	case 19: // IQ1_S
+		return 2 + blockSize/8 + blockSize/16
+	case 20: // IQ4_NL
+		return 2 + blockSize/2
+	case 21: // IQ3_S
+		return 2 + blockSize/4 + blockSize/8 + blockSize/32 + 4
+	case 22: // IQ2_S
+		return 2 + blockSize/4 + blockSize/16
+	case 23: // IQ4_XS
+		return 2 + 2 + blockSize/2 + blockSize/64
+	case 24: // I8
+		return 1
+	case 25: // I16
+		return 2
+	case 26: // I32
+		return 4
+	case 27: // I64
+		return 8
+	case 28: // F64
+		return 8
+	case 29: // IQ1_M
+		return blockSize/8 + blockSize/16 + blockSize/32
 	default:
 		return 0
 	}
@@ -185,7 +207,7 @@ func (t Tensor) parameters() uint64 {
 	return count
 }
 
-func (t Tensor) size() uint64 {
+func (t Tensor) Size() uint64 {
 	return t.parameters() * t.typeSize() / t.blockSize()
 }
 
@@ -288,7 +310,7 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 			// mixtral 8x22b
 			ff := uint64(llm.KV()["llama.feed_forward_length"].(uint32))
 			partialOffload = max(
-				3*ffnGateExpsWeight.size()+4*batch*(2*ff+headsKV+embedding+context+embedding/heads*headsKV),
+				3*ffnGateExpsWeight.Size()+4*batch*(2*ff+headsKV+embedding+context+embedding/heads*headsKV),
 				4*(context*batch*heads+context*embedding/heads*headsKV+batch*1024+embedding/heads*headsKV*batch),
 			)
 		} else if ffnGateWeight, ok := layers["blk.0"]["ffn_gate.0.weight"]; ok {

llm/gguf.go

@@ -62,16 +62,6 @@ func (c *containerGGUF) Decode(rs io.ReadSeeker) (model, error) {
 	return model, nil
 }
 
-const (
-	_ uint32 = iota
-	GGUFTokenNormal
-	GGUFTokenUnknown
-	GGUFTokenControl
-	GGUFTokenUserDefined
-	GGUFTokenUnused
-	GGUFTokenByte
-)
-
 const (
 	ggufTypeUint8 uint32 = iota
 	ggufTypeInt8
@@ -251,11 +241,11 @@ func (llm *gguf) Decode(rs io.ReadSeeker) error {
 	}
 
 	for _, tensor := range llm.tensors {
-		if _, err := rs.Seek(int64(tensor.size()), io.SeekCurrent); err != nil {
+		if _, err := rs.Seek(int64(tensor.Size()), io.SeekCurrent); err != nil {
 			return err
 		}
 
-		padding := llm.padding(int64(tensor.size()), int64(alignment))
+		padding := llm.padding(int64(tensor.Size()), int64(alignment))
 		if _, err := rs.Seek(padding, io.SeekCurrent); err != nil {
 			return err
 		}
@@ -480,9 +470,11 @@ var ggufKVOrder = map[string][]string{
 		"gemma.attention.key_length",
 		"gemma.attention.value_length",
 		"general.file_type",
+		"tokenizer.ggml.pre",
 		"tokenizer.ggml.model",
 		"tokenizer.ggml.tokens",
 		"tokenizer.ggml.scores",
+		"tokenizer.ggml.merges",
 		"tokenizer.ggml.token_type",
 		"tokenizer.ggml.bos_token_id",
 		"tokenizer.ggml.eos_token_id",

llm/memory.go

@@ -53,6 +53,12 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 		opts.NumCtx = max(opts.NumCtx, 2048)
 	}
 
+	layers := ggml.Tensors().Layers()
+	// add one layer worth of memory as a buffer
+	if blk0, ok := layers["blk.0"]; ok {
+		memoryMinimum += blk0.size()
+	}
+
 	// fp16 k,v = (1 (k) + 1 (v)) * sizeof(float16) * n_ctx * n_layer * n_embd / n_head * n_head_kv
 	var kv uint64 = 2 * 2 * uint64(opts.NumCtx) * ggml.KV().BlockCount() * ggml.KV().EmbeddingLength() / ggml.KV().HeadCount() * ggml.KV().HeadCountKV()
 
@@ -73,13 +79,11 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 		graphPartialOffload = graphFullOffload
 	}
 
-	layers := ggml.Tensors().Layers()
-
 	// memoryRequiredTotal represents the memory required for full GPU offloading (all layers)
-	memoryRequiredTotal := memoryMinimum + graphFullOffload + layers["blk.0"].size()
+	memoryRequiredTotal := memoryMinimum + graphFullOffload
 
 	// memoryRequiredPartial represents the memory required for partial GPU offloading (n > 0, n < layers)
-	memoryRequiredPartial := memoryMinimum + graphPartialOffload + layers["blk.0"].size()
+	memoryRequiredPartial := memoryMinimum + graphPartialOffload
 
 	var memoryLayerOutput uint64
 	if layer, ok := layers["output_norm"]; ok {
@@ -100,15 +104,17 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 
 	var layerCount int
 	for i := 0; i < int(ggml.KV().BlockCount()); i++ {
-		memoryLayer := layers[fmt.Sprintf("blk.%d", i)].size()
+		if blk, ok := layers[fmt.Sprintf("blk.%d", i)]; ok {
+			memoryLayer := blk.size()
 
-		// KV is proportional to the number of layers
+			// KV is proportional to the number of layers
-		memoryLayer += kv / ggml.KV().BlockCount()
+			memoryLayer += kv / ggml.KV().BlockCount()
 
-		memoryRequiredTotal += memoryLayer
+			memoryRequiredTotal += memoryLayer
-		if memoryAvailable > memoryRequiredPartial+memoryLayer {
+			if (opts.NumGPU >= 0 && layerCount+1 <= opts.NumGPU) || (opts.NumGPU < 0 && memoryAvailable > memoryRequiredPartial+memoryLayer) {
-			memoryRequiredPartial += memoryLayer
+				memoryRequiredPartial += memoryLayer
-			layerCount++
+				layerCount++
+			}
 		}
 	}
 
@@ -117,7 +123,7 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 		memoryRequiredTotal += memoryLayerOutput
 	}
 
-	if memoryAvailable > memoryRequiredTotal {
+	if (opts.NumGPU >= 0 && layerCount+1 <= opts.NumGPU) || (opts.NumGPU < 0 && memoryAvailable > memoryRequiredTotal) {
 		layerCount = int(ggml.KV().BlockCount()) + 1
 		memoryRequiredPartial = memoryRequiredTotal
 	}
@@ -128,10 +134,10 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 		"offload to gpu",
 		slog.Group(
 			"layers",
-			// actual number of layers offloaded
+			// requested number of layers to offload
-			"real", opts.NumGPU,
+			"requested", opts.NumGPU,
 			// estimated number of layers that can be offloaded
-			"estimate", layerCount,
+			"real", layerCount,
 		),
 		slog.Group(
 			"memory",

llm/patches/01-load-progress.diff

@@ -0,0 +1,31 @@
+diff --git a/common/common.cpp b/common/common.cpp
+index ba1ecf0e..cead57cc 100644
+--- a/common/common.cpp
++++ b/common/common.cpp
+@@ -1836,6 +1836,8 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
+     mparams.use_mmap        = params.use_mmap;
+     mparams.use_mlock       = params.use_mlock;
+     mparams.check_tensors   = params.check_tensors;
++    mparams.progress_callback = params.progress_callback;
++    mparams.progress_callback_user_data = params.progress_callback_user_data;
+     if (params.kv_overrides.empty()) {
+         mparams.kv_overrides = NULL;
+     } else {
+diff --git a/common/common.h b/common/common.h
+index d80344f2..71e84834 100644
+--- a/common/common.h
++++ b/common/common.h
+@@ -174,6 +174,13 @@ struct gpt_params {
+     // multimodal models (see examples/llava)
+     std::string mmproj = "";        // path to multimodal projector
+     std::vector<std::string> image; // path to image file(s)
++
++    // Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
++    // If the provided progress_callback returns true, model loading continues.
++    // If it returns false, model loading is immediately aborted.
++    llama_progress_callback progress_callback = NULL;
++    // context pointer passed to the progress callback
++    void * progress_callback_user_data;
+ };
+ 
+ void gpt_params_handle_model_default(gpt_params & params);

llm/patches/03-load_exception.diff

@@ -1,8 +1,17 @@
+From 544a2d2e646d39e878d87dfbb3398a356bc560ab Mon Sep 17 00:00:00 2001
+From: Michael Yang <mxyng@pm.me>
+Date: Thu, 23 May 2024 11:18:45 -0700
+Subject: [PATCH] throw exception on load errors
+
+---
+ llama.cpp | 25 ++++++++++++++++---------
+ 1 file changed, 16 insertions(+), 9 deletions(-)
+
 diff --git a/llama.cpp b/llama.cpp
-index 4225f955..7b762f86 100644
+index 15c66077..8ba90b6a 100644
 --- a/llama.cpp
 +++ b/llama.cpp
-@@ -4756,7 +4756,7 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
+@@ -6346,7 +6346,7 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
          }
      } catch (const std::exception & err) {
          LLAMA_LOG_ERROR("%s: error loading model: %s\n", __func__, err.what());
@@ -11,10 +20,10 @@ index 4225f955..7b762f86 100644
      }
  
      return 0;
-@@ -12102,16 +12102,22 @@ struct llama_model * llama_load_model_from_file(
+@@ -15600,16 +15600,23 @@ struct llama_model * llama_load_model_from_file(
-         };
+         }
+         model->rpc_servers.push_back(servers);
      }
- 
 -    int status = llama_model_load(path_model, *model, params);
 -    GGML_ASSERT(status <= 0);
 -    if (status < 0) {
@@ -22,6 +31,7 @@ index 4225f955..7b762f86 100644
 -            LLAMA_LOG_ERROR("%s: failed to load model\n", __func__);
 -        } else if (status == -2) {
 -            LLAMA_LOG_INFO("%s: cancelled model load\n", __func__);
++
 +    try {
 +        int status = llama_model_load(path_model, *model, params);
 +        GGML_ASSERT(status <= 0);
@@ -42,3 +52,6 @@ index 4225f955..7b762f86 100644
      }
  
      return model;
+-- 
+2.45.1
+

llm/patches/05-clip-fix.diff

@@ -1,24 +0,0 @@
-diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp
-index e3c9bcd4..b43f892d 100644
---- a/examples/llava/clip.cpp
-+++ b/examples/llava/clip.cpp
-@@ -573,14 +573,16 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
-     struct ggml_tensor * embeddings = inp;
-     if (ctx->has_class_embedding) {
-         embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size);
-+    }
-+    ggml_set_name(embeddings, "embeddings");
-+    ggml_set_input(embeddings);
-+
-+    if (ctx->has_class_embedding) {
-         embeddings = ggml_acc(ctx0, embeddings, model.class_embedding,
-                 embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], 0);
-         embeddings = ggml_acc(ctx0, embeddings, inp,
-                 embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]);
-     }
--    ggml_set_name(embeddings, "embeddings");
--    ggml_set_input(embeddings);
--
- 
-     struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_positions);
-     ggml_set_name(positions, "positions");

llm/patches/05-default-pretokenizer.diff

@@ -0,0 +1,35 @@
+From d02a06f3f45a09255ace8684a66590e06ce44605 Mon Sep 17 00:00:00 2001
+From: Michael Yang <mxyng@pm.me>
+Date: Thu, 23 May 2024 11:33:20 -0700
+Subject: [PATCH] default pretokenizer on unrecognized type
+
+---
+ llama.cpp | 5 +----
+ 1 file changed, 1 insertion(+), 4 deletions(-)
+
+diff --git a/llama.cpp b/llama.cpp
+index 15c66077..af1aede3 100644
+--- a/llama.cpp
++++ b/llama.cpp
+@@ -4504,9 +4504,6 @@ static void llm_load_vocab(
+                 LLAMA_LOG_WARN("%s: ************************************        \n", __func__);
+                 LLAMA_LOG_WARN("%s:                                             \n", __func__);
+                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
+-            } else if (
+-                    tokenizer_pre == "default") {
+-                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
+             } else if (
+                     tokenizer_pre == "llama3"   ||
+                     tokenizer_pre == "llama-v3" ||
+@@ -4553,7 +4550,7 @@ static void llm_load_vocab(
+                 tokenizer_pre == "dbrx") {
+                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DBRX;
+             } else {
+-                throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
++                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
+             }
+         } else {
+             vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
+-- 
+2.45.1
+

llm/server.go

@@ -38,6 +38,7 @@ type LlamaServer interface {
 	Detokenize(ctx context.Context, tokens []int) (string, error)
 	Close() error
 	EstimatedVRAM() uint64
+	EstimatedTotal() uint64
 }
 
 // llmServer is an instance of the llama.cpp server
@@ -54,6 +55,7 @@ type llmServer struct {
 	totalLayers    uint64
 	gpuCount       int
 	loadDuration   time.Duration // Record how long it took the model to load
+	loadProgress   float32
 
 	sem *semaphore.Weighted
 }
@@ -88,6 +90,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 
 		cpuRunner = serverForCpu()
 		gpuCount = 0
+		_, _, estimatedTotal = EstimateGPULayers(gpus, ggml, projectors, opts)
 	} else {
 		if gpus[0].Library == "metal" {
 			memInfo, err := gpu.GetCPUMem()
@@ -198,6 +201,23 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		params = append(params, "--numa")
 	}
 
+	flashAttnEnabled := envconfig.FlashAttention
+
+	// partial offloading does not support flash attention
+	if uint64(opts.NumGPU) < ggml.KV().BlockCount()+1 {
+		flashAttnEnabled = false
+	}
+
+	// only cuda (compute capability 7+) and metal support flash attention
+	for _, g := range gpus {
+		if g.Library != "metal" && (g.Library != "cuda" || g.DriverMajor < 7) {
+			flashAttnEnabled = false
+		}
+	}
+	if flashAttnEnabled {
+		params = append(params, "--flash-attn")
+	}
+
 	numParallel := envconfig.NumParallel
 
 	// TODO (jmorganca): multimodal models don't support parallel yet
@@ -316,8 +336,22 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		}
 
 		slog.Info("starting llama server", "cmd", s.cmd.String())
-		// Log at debug as the environment is inherited and might contain sensitive information
+		if envconfig.Debug {
-		slog.Debug("subprocess", "environment", s.cmd.Env)
+			filteredEnv := []string{}
+			for _, ev := range s.cmd.Env {
+				if strings.HasPrefix(ev, "CUDA_") ||
+					strings.HasPrefix(ev, "ROCM_") ||
+					strings.HasPrefix(ev, "HIP_") ||
+					strings.HasPrefix(ev, "HSA_") ||
+					strings.HasPrefix(ev, "GGML_") ||
+					strings.HasPrefix(ev, "PATH=") ||
+					strings.HasPrefix(ev, "LD_LIBRARY_PATH=") {
+					filteredEnv = append(filteredEnv, ev)
+				}
+			}
+			// Log at debug as the environment is inherited and might contain sensitive information
+			slog.Debug("subprocess", "environment", filteredEnv)
+		}
 
 		if err = s.cmd.Start(); err != nil {
 			// Detect permission denied and augment them essage about noexec
@@ -392,10 +426,11 @@ func (s ServerStatus) ToString() string {
 }
 
 type ServerStatusResp struct {
-	Status          string `json:"status"`
+	Status          string  `json:"status"`
-	SlotsIdle       int    `json:"slots_idle"`
+	SlotsIdle       int     `json:"slots_idle"`
-	SlotsProcessing int    `json:"slots_processing"`
+	SlotsProcessing int     `json:"slots_processing"`
-	Error           string `json:"error"`
+	Error           string  `json:"error"`
+	Progress        float32 `json:"progress"`
 }
 
 func (s *llmServer) getServerStatus(ctx context.Context) (ServerStatus, error) {
@@ -443,6 +478,7 @@ func (s *llmServer) getServerStatus(ctx context.Context) (ServerStatus, error) {
 	case "no slot available":
 		return ServerStatusNoSlotsAvailable, nil
 	case "loading model":
+		s.loadProgress = status.Progress
 		return ServerStatusLoadingModel, nil
 	default:
 		return ServerStatusError, fmt.Errorf("server error: %+v", status)
@@ -483,7 +519,8 @@ func (s *llmServer) Ping(ctx context.Context) error {
 
 func (s *llmServer) WaitUntilRunning(ctx context.Context) error {
 	start := time.Now()
-	expiresAt := time.Now().Add(10 * time.Minute) // be generous with timeout, large models can take a while to load
+	stallDuration := 60 * time.Second
+	stallTimer := time.Now().Add(stallDuration) // give up if we stall for
 
 	slog.Info("waiting for llama runner to start responding")
 	var lastStatus ServerStatus = -1
@@ -501,13 +538,13 @@ func (s *llmServer) WaitUntilRunning(ctx context.Context) error {
 			return fmt.Errorf("llama runner process has terminated: %v %s", err, msg)
 		default:
 		}
-		if time.Now().After(expiresAt) {
+		if time.Now().After(stallTimer) {
 			// timeout
 			msg := ""
 			if s.status != nil && s.status.LastErrMsg != "" {
 				msg = s.status.LastErrMsg
 			}
-			return fmt.Errorf("timed out waiting for llama runner to start: %s", msg)
+			return fmt.Errorf("timed out waiting for llama runner to start - progress %0.2f - %s", s.loadProgress, msg)
 		}
 		if s.cmd.ProcessState != nil {
 			msg := ""
@@ -518,6 +555,7 @@ func (s *llmServer) WaitUntilRunning(ctx context.Context) error {
 		}
 		ctx, cancel := context.WithTimeout(ctx, 200*time.Millisecond)
 		defer cancel()
+		priorProgress := s.loadProgress
 		status, _ := s.getServerStatus(ctx)
 		if lastStatus != status && status != ServerStatusReady {
 			// Only log on status changes
@@ -530,6 +568,11 @@ func (s *llmServer) WaitUntilRunning(ctx context.Context) error {
 			return nil
 		default:
 			lastStatus = status
+			// Reset the timer as long as we're making forward progress on the load
+			if priorProgress != s.loadProgress {
+				slog.Debug(fmt.Sprintf("model load progress %0.2f", s.loadProgress))
+				stallTimer = time.Now().Add(stallDuration)
+			}
 			time.Sleep(time.Millisecond * 250)
 			continue
 		}
@@ -955,6 +998,10 @@ func (s *llmServer) EstimatedVRAM() uint64 {
 	return s.estimatedVRAM
 }
 
+func (s *llmServer) EstimatedTotal() uint64 {
+	return s.estimatedTotal
+}
+
 func parseDurationMs(ms float64) time.Duration {
 	dur, err := time.ParseDuration(fmt.Sprintf("%fms", ms))
 	if err != nil {

macapp/src/index.ts

@@ -162,7 +162,7 @@ app.on('before-quit', () => {
   }
 })
 
-const updateURL = `https://ollama.ai/api/update?os=${process.platform}&arch=${
+const updateURL = `https://ollama.com/api/update?os=${process.platform}&arch=${
   process.arch
 }&version=${app.getVersion()}&id=${id()}`
 

parser/parser.go

@@ -1,4 +1,4 @@
-package model
+package parser
 
 import (
 	"bufio"
@@ -8,6 +8,7 @@ import (
 	"io"
 	"strconv"
 	"strings"
+	"unicode"
 )
 
 type File struct {
@@ -68,6 +69,11 @@ func ParseFile(r io.Reader) (*File, error) {
 	var b bytes.Buffer
 	var role string
 
+	var lineCount int
+	var linePos int
+
+	var utf16 bool
+
 	var f File
 
 	br := bufio.NewReader(r)
@@ -79,6 +85,17 @@ func ParseFile(r io.Reader) (*File, error) {
 			return nil, err
 		}
 
+		// the utf16 byte order mark will be read as "unreadable" by ReadRune()
+		if isUnreadable(r) && lineCount == 0 && linePos == 0 {
+			utf16 = true
+			continue
+		}
+
+		// skip the second byte if we're reading utf16
+		if utf16 && r == 0 {
+			continue
+		}
+
 		next, r, err := parseRuneForState(r, curr)
 		if errors.Is(err, io.ErrUnexpectedEOF) {
 			return nil, fmt.Errorf("%w: %s", err, b.String())
@@ -86,6 +103,13 @@ func ParseFile(r io.Reader) (*File, error) {
 			return nil, err
 		}
 
+		if isNewline(r) {
+			lineCount++
+			linePos = 0
+		} else {
+			linePos++
+		}
+
 		// process the state transition, some transitions need to be intercepted and redirected
 		if next != curr {
 			switch curr {
@@ -285,6 +309,10 @@ func isNewline(r rune) bool {
 	return r == '\r' || r == '\n'
 }
 
+func isUnreadable(r rune) bool {
+	return r == unicode.ReplacementChar
+}
+
 func isValidMessageRole(role string) bool {
 	return role == "system" || role == "user" || role == "assistant"
 }

parser/parser_test.go

@@ -1,11 +1,13 @@
-package model
+package parser
 
 import (
 	"bytes"
+	"encoding/binary"
 	"fmt"
 	"io"
 	"strings"
 	"testing"
+	"unicode/utf16"
 
 	"github.com/stretchr/testify/assert"
 )
@@ -509,3 +511,37 @@ SYSTEM ""
 	}
 
 }
+
+func TestParseFileUTF16ParseFile(t *testing.T) {
+	data := `FROM bob
+PARAMETER param1 1
+PARAMETER param2 4096
+SYSTEM You are a utf16 file.
+`
+	// simulate a utf16 le file
+	utf16File := utf16.Encode(append([]rune{'\ufffe'}, []rune(data)...))
+	buf := new(bytes.Buffer)
+	err := binary.Write(buf, binary.LittleEndian, utf16File)
+	assert.NoError(t, err)
+
+	actual, err := ParseFile(buf)
+	assert.NoError(t, err)
+
+	expected := []Command{
+		{Name: "model", Args: "bob"},
+		{Name: "param1", Args: "1"},
+		{Name: "param2", Args: "4096"},
+		{Name: "system", Args: "You are a utf16 file."},
+	}
+
+	assert.Equal(t, expected, actual.Commands)
+
+	// simulate a utf16 be file
+	buf = new(bytes.Buffer)
+	err = binary.Write(buf, binary.BigEndian, utf16File)
+	assert.NoError(t, err)
+
+	actual, err = ParseFile(buf)
+	assert.NoError(t, err)
+	assert.Equal(t, expected, actual.Commands)
+}

server/download.go

@@ -221,7 +221,7 @@ func (b *blobDownload) downloadChunk(ctx context.Context, requestURL *url.URL, w
 		}
 		defer resp.Body.Close()
 
-		n, err := io.Copy(w, io.TeeReader(resp.Body, part))
+		n, err := io.CopyN(w, io.TeeReader(resp.Body, part), part.Size)
 		if err != nil && !errors.Is(err, context.Canceled) && !errors.Is(err, io.ErrUnexpectedEOF) {
 			// rollback progress
 			b.Completed.Add(-n)

server/envconfig/config.go

@@ -31,6 +31,8 @@ var (
 	RunnersDir string
 	// Set via OLLAMA_TMPDIR in the environment
 	TmpDir string
+	// Experimental flash attention
+	FlashAttention bool
 )
 
 func AsMap() map[string]string {
@@ -45,6 +47,7 @@ func AsMap() map[string]string {
 		"OLLAMA_NUM_PARALLEL":      fmt.Sprintf("%v", NumParallel),
 		"OLLAMA_RUNNERS_DIR":       fmt.Sprintf("%v", RunnersDir),
 		"OLLAMA_TMPDIR":            fmt.Sprintf("%v", TmpDir),
+		"OLLAMA_FLASH_ATTENTION":   fmt.Sprintf("%v", FlashAttention),
 	}
 }
 
@@ -78,6 +81,13 @@ func LoadConfig() {
 		}
 	}
 
+	if fa := clean("OLLAMA_FLASH_ATTENTION"); fa != "" {
+		d, err := strconv.ParseBool(fa)
+		if err == nil {
+			FlashAttention = d
+		}
+	}
+
 	RunnersDir = clean("OLLAMA_RUNNERS_DIR")
 	if runtime.GOOS == "windows" && RunnersDir == "" {
 		// On Windows we do not carry the payloads inside the main executable

server/envconfig/config_test.go

@@ -17,4 +17,7 @@ func TestConfig(t *testing.T) {
 	t.Setenv("OLLAMA_DEBUG", "1")
 	LoadConfig()
 	require.True(t, Debug)
+	t.Setenv("OLLAMA_FLASH_ATTENTION", "1")
+	LoadConfig()
+	require.True(t, FlashAttention)
 }

server/images.go

@@ -27,6 +27,7 @@ import (
 	"github.com/ollama/ollama/auth"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/parser"
 	"github.com/ollama/ollama/server/envconfig"
 	"github.com/ollama/ollama/types/errtypes"
 	"github.com/ollama/ollama/types/model"
@@ -61,36 +62,36 @@ func (m *Model) IsEmbedding() bool {
 }
 
 func (m *Model) String() string {
-	var modelfile model.File
+	var modelfile parser.File
 
-	modelfile.Commands = append(modelfile.Commands, model.Command{
+	modelfile.Commands = append(modelfile.Commands, parser.Command{
 		Name: "model",
 		Args: m.ModelPath,
 	})
 
 	for _, adapter := range m.AdapterPaths {
-		modelfile.Commands = append(modelfile.Commands, model.Command{
+		modelfile.Commands = append(modelfile.Commands, parser.Command{
 			Name: "adapter",
 			Args: adapter,
 		})
 	}
 
 	for _, projector := range m.ProjectorPaths {
-		modelfile.Commands = append(modelfile.Commands, model.Command{
+		modelfile.Commands = append(modelfile.Commands, parser.Command{
 			Name: "model",
 			Args: projector,
 		})
 	}
 
 	if m.Template != "" {
-		modelfile.Commands = append(modelfile.Commands, model.Command{
+		modelfile.Commands = append(modelfile.Commands, parser.Command{
 			Name: "template",
 			Args: m.Template,
 		})
 	}
 
 	if m.System != "" {
-		modelfile.Commands = append(modelfile.Commands, model.Command{
+		modelfile.Commands = append(modelfile.Commands, parser.Command{
 			Name: "system",
 			Args: m.System,
 		})
@@ -100,13 +101,13 @@ func (m *Model) String() string {
 		switch v := v.(type) {
 		case []any:
 			for _, s := range v {
-				modelfile.Commands = append(modelfile.Commands, model.Command{
+				modelfile.Commands = append(modelfile.Commands, parser.Command{
 					Name: k,
 					Args: fmt.Sprintf("%v", s),
 				})
 			}
 		default:
-			modelfile.Commands = append(modelfile.Commands, model.Command{
+			modelfile.Commands = append(modelfile.Commands, parser.Command{
 				Name: k,
 				Args: fmt.Sprintf("%v", v),
 			})
@@ -114,14 +115,14 @@ func (m *Model) String() string {
 	}
 
 	for _, license := range m.License {
-		modelfile.Commands = append(modelfile.Commands, model.Command{
+		modelfile.Commands = append(modelfile.Commands, parser.Command{
 			Name: "license",
 			Args: license,
 		})
 	}
 
 	for _, msg := range m.Messages {
-		modelfile.Commands = append(modelfile.Commands, model.Command{
+		modelfile.Commands = append(modelfile.Commands, parser.Command{
 			Name: "message",
 			Args: fmt.Sprintf("%s %s", msg.Role, msg.Content),
 		})
@@ -314,7 +315,7 @@ func realpath(rel, from string) string {
 	return abspath
 }
 
-func CreateModel(ctx context.Context, name, modelFileDir, quantization string, modelfile *model.File, fn func(resp api.ProgressResponse)) (err error) {
+func CreateModel(ctx context.Context, name, modelFileDir, quantization string, modelfile *parser.File, fn func(resp api.ProgressResponse)) (err error) {
 	config := ConfigV2{
 		OS:           "linux",
 		Architecture: "amd64",
@@ -339,7 +340,24 @@ func CreateModel(ctx context.Context, name, modelFileDir, quantization string, m
 					return err
 				}
 			} else if strings.HasPrefix(c.Args, "@") {
-				blobpath, err := GetBlobsPath(strings.TrimPrefix(c.Args, "@"))
+				digest := strings.TrimPrefix(c.Args, "@")
+				if ib, ok := intermediateBlobs[digest]; ok {
+					p, err := GetBlobsPath(ib)
+					if err != nil {
+						return err
+					}
+
+					if _, err := os.Stat(p); errors.Is(err, os.ErrNotExist) {
+						// pass
+					} else if err != nil {
+						return err
+					} else {
+						fn(api.ProgressResponse{Status: fmt.Sprintf("using cached layer %s", ib)})
+						digest = ib
+					}
+				}
+
+				blobpath, err := GetBlobsPath(digest)
 				if err != nil {
 					return err
 				}
@@ -350,14 +368,14 @@ func CreateModel(ctx context.Context, name, modelFileDir, quantization string, m
 				}
 				defer blob.Close()
 
-				baseLayers, err = parseFromFile(ctx, blob, fn)
+				baseLayers, err = parseFromFile(ctx, blob, digest, fn)
 				if err != nil {
 					return err
 				}
 			} else if file, err := os.Open(realpath(modelFileDir, c.Args)); err == nil {
 				defer file.Close()
 
-				baseLayers, err = parseFromFile(ctx, file, fn)
+				baseLayers, err = parseFromFile(ctx, file, "", fn)
 				if err != nil {
 					return err
 				}
@@ -397,10 +415,17 @@ func CreateModel(ctx context.Context, name, modelFileDir, quantization string, m
 							return err
 						}
 
-						baseLayer.Layer, err = NewLayer(temp, baseLayer.Layer.MediaType)
+						layers, err := parseFromFile(ctx, temp, "", fn)
 						if err != nil {
 							return err
 						}
+
+						if len(layers) != 1 {
+							return errors.New("quantization failed")
+						}
+
+						baseLayer.Layer = layers[0].Layer
+						baseLayer.GGML = layers[0].GGML
 					}
 				}
 

server/layer.go

@@ -80,7 +80,7 @@ func NewLayerFromLayer(digest, mediatype, from string) (*Layer, error) {
 	}, nil
 }
 
-func (l *Layer) Open() (io.ReadCloser, error) {
+func (l *Layer) Open() (io.ReadSeekCloser, error) {
 	blob, err := GetBlobsPath(l.Digest)
 	if err != nil {
 		return nil, err

server/model.go

@@ -17,6 +17,8 @@ import (
 	"github.com/ollama/ollama/types/model"
 )
 
+var intermediateBlobs map[string]string = make(map[string]string)
+
 type layerWithGGML struct {
 	*Layer
 	*llm.GGML
@@ -76,7 +78,7 @@ func parseFromModel(ctx context.Context, name model.Name, fn func(api.ProgressRe
 	return layers, nil
 }
 
-func parseFromZipFile(_ context.Context, file *os.File, fn func(api.ProgressResponse)) (layers []*layerWithGGML, err error) {
+func parseFromZipFile(_ context.Context, file *os.File, digest string, fn func(api.ProgressResponse)) (layers []*layerWithGGML, err error) {
 	stat, err := file.Stat()
 	if err != nil {
 		return nil, err
@@ -165,16 +167,11 @@ func parseFromZipFile(_ context.Context, file *os.File, fn func(api.ProgressResp
 	}
 
 	layer, err := NewLayer(temp, "application/vnd.ollama.image.model")
-	if err != nil {
-		return nil, fmt.Errorf("aaa: %w", err)
-	}
-
-	blobpath, err := GetBlobsPath(layer.Digest)
 	if err != nil {
 		return nil, err
 	}
 
-	bin, err := os.Open(blobpath)
+	bin, err := layer.Open()
 	if err != nil {
 		return nil, err
 	}
@@ -185,16 +182,13 @@ func parseFromZipFile(_ context.Context, file *os.File, fn func(api.ProgressResp
 		return nil, err
 	}
 
-	layer, err = NewLayerFromLayer(layer.Digest, layer.MediaType, "")
-	if err != nil {
-		return nil, err
-	}
-
 	layers = append(layers, &layerWithGGML{layer, ggml})
+
+	intermediateBlobs[digest] = layer.Digest
 	return layers, nil
 }
 
-func parseFromFile(ctx context.Context, file *os.File, fn func(api.ProgressResponse)) (layers []*layerWithGGML, err error) {
+func parseFromFile(ctx context.Context, file *os.File, digest string, fn func(api.ProgressResponse)) (layers []*layerWithGGML, err error) {
 	sr := io.NewSectionReader(file, 0, 512)
 	contentType, err := detectContentType(sr)
 	if err != nil {
@@ -205,7 +199,7 @@ func parseFromFile(ctx context.Context, file *os.File, fn func(api.ProgressRespo
 	case "gguf", "ggla":
 		// noop
 	case "application/zip":
-		return parseFromZipFile(ctx, file, fn)
+		return parseFromZipFile(ctx, file, digest, fn)
 	default:
 		return nil, fmt.Errorf("unsupported content type: %s", contentType)
 	}

server/routes.go

@@ -29,7 +29,9 @@ import (
 	"github.com/ollama/ollama/gpu"
 	"github.com/ollama/ollama/llm"
 	"github.com/ollama/ollama/openai"
+	"github.com/ollama/ollama/parser"
 	"github.com/ollama/ollama/server/envconfig"
+	"github.com/ollama/ollama/types/errtypes"
 	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/version"
 )
@@ -517,7 +519,7 @@ func (s *Server) CreateModelHandler(c *gin.Context) {
 
 	name := model.ParseName(cmp.Or(req.Model, req.Name))
 	if !name.IsValid() {
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "invalid model name"})
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": errtypes.InvalidModelNameErrMsg})
 		return
 	}
 
@@ -538,7 +540,7 @@ func (s *Server) CreateModelHandler(c *gin.Context) {
 		r = f
 	}
 
-	modelfile, err := model.ParseFile(r)
+	modelfile, err := parser.ParseFile(r)
 	if err != nil {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
 		return
@@ -708,7 +710,7 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 	}
 
 	var sb strings.Builder
-	fmt.Fprintln(&sb, "# Modelfile generate by \"ollama show\"")
+	fmt.Fprintln(&sb, "# Modelfile generated by \"ollama show\"")
 	fmt.Fprintln(&sb, "# To build a new Modelfile based on this, replace FROM with:")
 	fmt.Fprintf(&sb, "# FROM %s\n\n", model.ShortName)
 	fmt.Fprint(&sb, model.String())
@@ -724,7 +726,7 @@ func (s *Server) ListModelsHandler(c *gin.Context) {
 		return
 	}
 
-	var models []api.ModelResponse
+	models := []api.ModelResponse{}
 	if err := filepath.Walk(manifests, func(path string, info os.FileInfo, _ error) error {
 		if !info.IsDir() {
 			rel, err := filepath.Rel(manifests, path)
@@ -839,6 +841,25 @@ func (s *Server) HeadBlobHandler(c *gin.Context) {
 }
 
 func (s *Server) CreateBlobHandler(c *gin.Context) {
+	if ib, ok := intermediateBlobs[c.Param("digest")]; ok {
+		p, err := GetBlobsPath(ib)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+			return
+		}
+
+		if _, err := os.Stat(p); errors.Is(err, os.ErrNotExist) {
+			slog.Info("evicting intermediate blob which no longer exists", "digest", ib)
+			delete(intermediateBlobs, c.Param("digest"))
+		} else if err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+			return
+		} else {
+			c.Status(http.StatusOK)
+			return
+		}
+	}
+
 	path, err := GetBlobsPath(c.Param("digest"))
 	if err != nil {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
@@ -979,6 +1000,7 @@ func (s *Server) GenerateRoutes() http.Handler {
 	r.POST("/api/show", s.ShowModelHandler)
 	r.POST("/api/blobs/:digest", s.CreateBlobHandler)
 	r.HEAD("/api/blobs/:digest", s.HeadBlobHandler)
+	r.GET("/api/ps", s.ProcessHandler)
 
 	// Compatibility endpoints
 	r.POST("/v1/chat/completions", openai.Middleware(), s.ChatHandler)
@@ -1084,7 +1106,7 @@ func Serve(ln net.Listener) error {
 		return err
 	}
 	<-ctx.Done()
-	return err
+	return nil
 }
 
 func waitForStream(c *gin.Context, ch chan interface{}) {
@@ -1137,6 +1159,42 @@ func streamResponse(c *gin.Context, ch chan any) {
 	})
 }
 
+func (s *Server) ProcessHandler(c *gin.Context) {
+	models := []api.ModelResponse{}
+
+	for _, v := range s.sched.loaded {
+		model := v.model
+		modelDetails := api.ModelDetails{
+			Format:            model.Config.ModelFormat,
+			Family:            model.Config.ModelFamily,
+			Families:          model.Config.ModelFamilies,
+			ParameterSize:     model.Config.ModelType,
+			QuantizationLevel: model.Config.FileType,
+		}
+
+		mr := api.ModelResponse{
+			Model:     model.ShortName,
+			Name:      model.ShortName,
+			Size:      int64(v.estimatedTotal),
+			SizeVRAM:  int64(v.estimatedVRAM),
+			Digest:    model.Digest,
+			Details:   modelDetails,
+			ExpiresAt: v.expiresAt,
+		}
+		// The scheduler waits to set expiresAt, so if a model is loading it's
+		// possible that it will be set to the unix epoch. For those cases, just
+		// calculate the time w/ the sessionDuration instead.
+		var epoch time.Time
+		if v.expiresAt == epoch {
+			mr.ExpiresAt = time.Now().Add(v.sessionDuration)
+		}
+
+		models = append(models, mr)
+	}
+
+	c.JSON(http.StatusOK, api.ListResponse{Models: models})
+}
+
 // ChatPrompt builds up a prompt from a series of messages for the currently `loaded` model
 func chatPrompt(ctx context.Context, runner *runnerRef, template string, messages []api.Message, numCtx int) (string, error) {
 	encode := func(s string) ([]int, error) {

server/routes_test.go

@@ -17,7 +17,7 @@ import (
 	"github.com/stretchr/testify/assert"
 
 	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/types/model"
+	"github.com/ollama/ollama/parser"
 	"github.com/ollama/ollama/version"
 )
 
@@ -56,7 +56,7 @@ func Test_Routes(t *testing.T) {
 		fname := createTestFile(t, "ollama-model")
 
 		r := strings.NewReader(fmt.Sprintf("FROM %s\nPARAMETER seed 42\nPARAMETER top_p 0.9\nPARAMETER stop foo\nPARAMETER stop bar", fname))
-		modelfile, err := model.ParseFile(r)
+		modelfile, err := parser.ParseFile(r)
 		assert.Nil(t, err)
 		fn := func(resp api.ProgressResponse) {
 			t.Logf("Status: %s", resp.Status)
@@ -95,6 +95,7 @@ func Test_Routes(t *testing.T) {
 				err = json.Unmarshal(body, &modelList)
 				assert.Nil(t, err)
 
+				assert.NotNil(t, modelList.Models)
 				assert.Equal(t, 0, len(modelList.Models))
 			},
 		},

server/sched.go

@@ -6,6 +6,7 @@ import (
 	"fmt"
 	"log/slog"
 	"reflect"
+	"runtime"
 	"sort"
 	"strings"
 	"sync"
@@ -177,7 +178,7 @@ func (s *Scheduler) processPending(ctx context.Context) {
 				}
 				// Trigger an expiration to unload once it's done
 				runnerToExpire.refMu.Lock()
-				slog.Debug("resetting model to expire immediately to make room", "model", runnerToExpire.model, "refCount", runnerToExpire.refCount)
+				slog.Debug("resetting model to expire immediately to make room", "modelPath", runnerToExpire.modelPath, "refCount", runnerToExpire.refCount)
 				if runnerToExpire.expireTimer != nil {
 					runnerToExpire.expireTimer.Stop()
 					runnerToExpire.expireTimer = nil
@@ -190,13 +191,13 @@ func (s *Scheduler) processPending(ctx context.Context) {
 				// Wait for the unload to happen
 				// Note: at this point we're queueing up all incoming requests, even if they were for
 				// a different model that's loaded and not scheduled to be removed.
-				slog.Debug("waiting for pending requests to complete and unload to occur", "model", runnerToExpire.model)
+				slog.Debug("waiting for pending requests to complete and unload to occur", "modelPath", runnerToExpire.modelPath)
 				select {
 				case <-ctx.Done():
 					slog.Debug("shutting down scheduler pending loop")
 					return
 				case <-s.unloadedCh:
-					slog.Debug("unload completed", "model", runnerToExpire.model)
+					slog.Debug("unload completed", "modelPath", runnerToExpire.modelPath)
 					continue
 				}
 			}
@@ -219,23 +220,23 @@ func (s *Scheduler) processCompleted(ctx context.Context) {
 			runner := s.loaded[finished.model.ModelPath]
 			s.loadedMu.Unlock()
 			if runner == nil {
-				slog.Error("finished requeset signal received after model unloaded", "model", finished.model.ModelPath)
+				slog.Error("finished request signal received after model unloaded", "modelPath", finished.model.ModelPath)
 				continue
 			}
 			runner.refMu.Lock()
 			runner.refCount--
 			if runner.refCount <= 0 {
 				if runner.sessionDuration <= 0 {
-					slog.Debug("runner with zero duration has gone idle, expiring to unload", "model", runner.model)
+					slog.Debug("runner with zero duration has gone idle, expiring to unload", "modelPath", runner.modelPath)
 					if runner.expireTimer != nil {
 						runner.expireTimer.Stop()
 						runner.expireTimer = nil
 					}
 					s.expiredCh <- runner
 				} else if runner.expireTimer == nil {
-					slog.Debug("runner with non-zero duration has gone idle, adding timer", "model", runner.model, "duration", runner.sessionDuration)
+					slog.Debug("runner with non-zero duration has gone idle, adding timer", "modelPath", runner.modelPath, "duration", runner.sessionDuration)
 					runner.expireTimer = time.AfterFunc(runner.sessionDuration, func() {
-						slog.Debug("timer expired, expiring to unload", "model", runner.model)
+						slog.Debug("timer expired, expiring to unload", "modelPath", runner.modelPath)
 						runner.refMu.Lock()
 						defer runner.refMu.Unlock()
 						if runner.expireTimer != nil {
@@ -244,19 +245,21 @@ func (s *Scheduler) processCompleted(ctx context.Context) {
 						}
 						s.expiredCh <- runner
 					})
+					runner.expiresAt = time.Now().Add(runner.sessionDuration)
 				} else {
-					slog.Debug("runner with non-zero duration has gone idle, resetting timer", "model", runner.model, "duration", runner.sessionDuration)
+					slog.Debug("runner with non-zero duration has gone idle, resetting timer", "modelPath", runner.modelPath, "duration", runner.sessionDuration)
 					runner.expireTimer.Reset(runner.sessionDuration)
+					runner.expiresAt = time.Now().Add(runner.sessionDuration)
 				}
 			}
-			slog.Debug("after processing request finished event", "model", runner.model, "refCount", runner.refCount)
+			slog.Debug("after processing request finished event", "modelPath", runner.modelPath, "refCount", runner.refCount)
 			runner.refMu.Unlock()
 		case runner := <-s.expiredCh:
-			slog.Debug("runner expired event received", "model", runner.model)
+			slog.Debug("runner expired event received", "modelPath", runner.modelPath)
 			runner.refMu.Lock()
 			if runner.refCount > 0 {
 				// Shouldn't happen, but safeguard to ensure no leaked runners
-				slog.Debug("expired event with positive ref count, retrying", "model", runner.model, "refCount", runner.refCount)
+				slog.Debug("expired event with positive ref count, retrying", "modelPath", runner.modelPath, "refCount", runner.refCount)
 				go func(runner *runnerRef) {
 					// We can't unload yet, but want to as soon as the current request completes
 					// So queue up another expired event
@@ -268,16 +271,16 @@ func (s *Scheduler) processCompleted(ctx context.Context) {
 			}
 
 			s.loadedMu.Lock()
-			slog.Debug("got lock to unload", "model", runner.model)
+			slog.Debug("got lock to unload", "modelPath", runner.modelPath)
 			finished := runner.waitForVRAMRecovery()
 			runner.unload()
-			delete(s.loaded, runner.model)
+			delete(s.loaded, runner.modelPath)
 			s.loadedMu.Unlock()
-			slog.Debug("runner released", "model", runner.model)
+			slog.Debug("runner released", "modelPath", runner.modelPath)
 			runner.refMu.Unlock()
 
 			<-finished
-			slog.Debug("sending an unloaded event", "model", runner.model)
+			slog.Debug("sending an unloaded event", "modelPath", runner.modelPath)
 			s.unloadedCh <- struct{}{}
 		}
 	}
@@ -316,18 +319,20 @@ func (s *Scheduler) load(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList)
 		req.errCh <- err
 		return
 	}
-	runner := &runnerRef{}
+	runner := &runnerRef{
-	runner.model = req.model.ModelPath
+		model:           req.model,
-	runner.adapters = req.model.AdapterPaths
+		modelPath:       req.model.ModelPath,
-	runner.projectors = req.model.ProjectorPaths
+		llama:           llama,
-	runner.llama = llama
+		Options:         &req.opts,
-	runner.Options = &req.opts
+		sessionDuration: req.sessionDuration,
-	runner.sessionDuration = req.sessionDuration
+		gpus:            gpus,
-	runner.gpus = gpus
+		estimatedVRAM:   llama.EstimatedVRAM(),
-	runner.estimatedVRAM = llama.EstimatedVRAM()
+		estimatedTotal:  llama.EstimatedTotal(),
-	runner.loading = true
+		loading:         true,
-	runner.refCount = 1
+		refCount:        1,
+	}
 	runner.refMu.Lock()
+
 	s.loadedMu.Lock()
 	s.loaded[req.model.ModelPath] = runner
 	slog.Info("loaded runners", "count", len(s.loaded))
@@ -339,7 +344,7 @@ func (s *Scheduler) load(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList)
 			slog.Error("error loading llama server", "error", err)
 			runner.refCount--
 			req.errCh <- err
-			slog.Debug("triggering expiration for failed load", "model", runner.model)
+			slog.Debug("triggering expiration for failed load", "model", runner.modelPath)
 			s.expiredCh <- runner
 			return
 		}
@@ -408,17 +413,18 @@ type runnerRef struct {
 	refCount uint // prevent unloading if > 0
 	// unloading bool      // set to true when we are trying to unload the runner
 
-	llama         llm.LlamaServer
+	llama          llm.LlamaServer
-	loading       bool            // True only during initial load, then false forever
+	loading        bool            // True only during initial load, then false forever
-	gpus          gpu.GpuInfoList // Recorded at time of provisioning
+	gpus           gpu.GpuInfoList // Recorded at time of provisioning
-	estimatedVRAM uint64
+	estimatedVRAM  uint64
+	estimatedTotal uint64
 
 	sessionDuration time.Duration
 	expireTimer     *time.Timer
+	expiresAt       time.Time
 
-	model      string
+	model     *Model
-	adapters   []string
+	modelPath string
-	projectors []string
 	*api.Options
 }
 
@@ -431,9 +437,8 @@ func (runner *runnerRef) unload() {
 	if runner.llama != nil {
 		runner.llama.Close()
 	}
+	runner.model = nil
 	runner.llama = nil
-	runner.adapters = nil
-	runner.projectors = nil
 	runner.Options = nil
 	runner.gpus = nil
 }
@@ -462,8 +467,8 @@ func (runner *runnerRef) needsReload(ctx context.Context, req *LlmRequest) bool
 
 	ctx, cancel := context.WithTimeout(ctx, timeout)
 	defer cancel()
-	if !reflect.DeepEqual(runner.adapters, req.model.AdapterPaths) || // have the adapters changed?
+	if !reflect.DeepEqual(runner.model.AdapterPaths, req.model.AdapterPaths) || // have the adapters changed?
-		!reflect.DeepEqual(runner.projectors, req.model.ProjectorPaths) || // have the projectors changed?
+		!reflect.DeepEqual(runner.model.ProjectorPaths, req.model.ProjectorPaths) || // have the projectors changed?
 		!reflect.DeepEqual(optsExisting, optsNew) || // have the runner options changed?
 		runner.llama.Ping(ctx) != nil {
 		return true
@@ -483,8 +488,8 @@ func (runner *runnerRef) needsReload(ctx context.Context, req *LlmRequest) bool
 func (runner *runnerRef) waitForVRAMRecovery() chan interface{} {
 	finished := make(chan interface{}, 1)
 
-	// CPU or Metal don't need checking, so no waiting required
+	// CPU or Metal don't need checking, so no waiting required, windows can page VRAM, and the APIs we query tend to be optimistic on free space
-	if len(runner.gpus) == 1 && (runner.gpus[0].Library == "cpu" || runner.gpus[0].Library == "metal") {
+	if (len(runner.gpus) == 1 && (runner.gpus[0].Library == "cpu" || runner.gpus[0].Library == "metal")) || runtime.GOOS == "windows" {
 		finished <- struct{}{}
 		return finished
 	}

server/sched_test.go

@@ -151,7 +151,7 @@ func newScenario(t *testing.T, ctx context.Context, modelName string, estimatedV
 }
 
 func TestRequests(t *testing.T) {
-	ctx, done := context.WithTimeout(context.Background(), 500*time.Millisecond)
+	ctx, done := context.WithTimeout(context.Background(), time.Second)
 	defer done()
 
 	// Same model, same request
@@ -164,7 +164,8 @@ func TestRequests(t *testing.T) {
 
 	// simple reload of same model
 	scenario2a := newScenario(t, ctx, "ollama-model-1", 20)
-	scenario2a.req.model = scenario1a.req.model
+	tmpModel := *scenario1a.req.model
+	scenario2a.req.model = &tmpModel
 	scenario2a.ggml = scenario1a.ggml
 
 	// Multiple loaded models
@@ -496,10 +497,9 @@ func TestNeedsReload(t *testing.T) {
 	llm := &mockLlm{}
 	do := api.DefaultOptions()
 	runner := &runnerRef{
-		adapters:   []string{"adapter1"},
+		model:   &Model{AdapterPaths: []string{"adapter1"}, ProjectorPaths: []string{"projector1"}},
-		projectors: []string{"projector1"},
+		Options: &do,
-		Options:    &do,
+		llama:   llm,
-		llama:      llm,
 	}
 	req := &LlmRequest{
 		model: &Model{
@@ -510,10 +510,10 @@ func TestNeedsReload(t *testing.T) {
 	}
 	resp := runner.needsReload(ctx, req)
 	require.True(t, resp)
-	req.model.AdapterPaths = runner.adapters
+	req.model.AdapterPaths = runner.model.AdapterPaths
 	resp = runner.needsReload(ctx, req)
 	require.True(t, resp)
-	req.model.ProjectorPaths = runner.projectors
+	req.model.ProjectorPaths = runner.model.ProjectorPaths
 	runner.loading = true
 	req.opts.NumBatch = 1234
 	resp = runner.needsReload(ctx, req)
@@ -558,11 +558,11 @@ func TestUnloadAllRunners(t *testing.T) {
 func TestUnload(t *testing.T) {
 	llm1 := &mockLlm{}
 	r1 := &runnerRef{llama: llm1}
-	r2 := &runnerRef{adapters: []string{"A"}}
+	r2 := &runnerRef{model: &Model{AdapterPaths: []string{"A"}}}
 	r1.unload()
 	require.True(t, llm1.closeCalled)
 	r2.unload()
-	require.Nil(t, r2.adapters)
+	require.Nil(t, r2.model)
 }
 
 type mockLlm struct {
@@ -578,6 +578,7 @@ type mockLlm struct {
 	closeResp         error
 	closeCalled       bool
 	estimatedVRAM     uint64
+	estimatedTotal    uint64
 }
 
 func (s *mockLlm) Ping(ctx context.Context) error             { return s.pingResp }
@@ -598,4 +599,5 @@ func (s *mockLlm) Close() error {
 	s.closeCalled = true
 	return s.closeResp
 }
-func (s *mockLlm) EstimatedVRAM() uint64 { return s.estimatedVRAM }
+func (s *mockLlm) EstimatedVRAM() uint64  { return s.estimatedVRAM }
+func (s *mockLlm) EstimatedTotal() uint64 { return s.estimatedTotal }

types/errtypes/errtypes.go

@@ -7,6 +7,7 @@ import (
 )
 
 const UnknownOllamaKeyErrMsg = "unknown ollama key"
+const InvalidModelNameErrMsg = "invalid model name"
 
 // TODO: This should have a structured response from the API
 type UnknownOllamaKey struct {