handle intermediate blobs

- FROM /path/to/{safetensors,pytorch}
- FROM /path/to/fp{16,32}.bin
- FROM model:fp{16,32}

.gitignore

@@ -11,4 +11,5 @@ ggml-metal.metal
 .idea
 test_data
 *.crt
-llm/build
+llm/build
+__debug_bin*

README.md

@@ -1,5 +1,5 @@
 <div align="center">
-  <img alt="ollama" height="200px" src="https://github.com/ollama/ollama/assets/3325447/0d0b44e2-8f4a-4e99-9b52-a5c1c741c8f7">
+ <img alt="ollama" height="200px" src="https://github.com/ollama/ollama/assets/3325447/0d0b44e2-8f4a-4e99-9b52-a5c1c741c8f7">
 </div>
 
 # Ollama
@@ -173,7 +173,7 @@ I'm a basic program that prints the famous "Hello, world!" message to the consol
 The image features a yellow smiley face, which is likely the central focus of the picture.
 ```
 
-### Pass in prompt as arguments
+### Pass the prompt as an argument
 
 ```
 $ ollama run llama3 "Summarize this file: $(cat README.md)"
@@ -294,7 +294,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [RAGFlow: Open-source Retrieval-Augmented Generation engine based on deep document understanding](https://github.com/infiniflow/ragflow)
 - [chat: chat web app for teams](https://github.com/swuecho/chat)
 - [Lobe Chat](https://github.com/lobehub/lobe-chat) with [Integrating Doc](https://lobehub.com/docs/self-hosting/examples/ollama)
-- [Ollama RAG Chatbot: Local Chat with multiples PDFs using Ollama and RAG.](https://github.com/datvodinh/rag-chatbot.git)
+- [Ollama RAG Chatbot: Local Chat with multiple PDFs using Ollama and RAG.](https://github.com/datvodinh/rag-chatbot.git)
 
 ### Terminal
 
@@ -384,4 +384,4 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Discord-Ollama Chat Bot](https://github.com/kevinthedang/discord-ollama) (Generalized TypeScript Discord Bot w/ Tuning Documentation)
 
 ### Supported backends 
-- [llama.cpp](https://github.com/ggerganov/llama.cpp) project founded by Georgi Gerganov. 
+- [llama.cpp](https://github.com/ggerganov/llama.cpp) project founded by Georgi Gerganov. 

cmd/cmd.go

@@ -57,12 +57,13 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	p := progress.NewProgress(os.Stderr)
 	defer p.Stop()
 
-	modelfile, err := os.ReadFile(filename)
+	modelfile, err := os.Open(filename)
 	if err != nil {
 		return err
 	}
+	defer modelfile.Close()
 
-	commands, err := parser.Parse(bytes.NewReader(modelfile))
+	commands, err := parser.Parse(modelfile)
 	if err != nil {
 		return err
 	}
@@ -76,10 +77,10 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	spinner := progress.NewSpinner(status)
 	p.Add(status, spinner)
 
-	for _, c := range commands {
-		switch c.Name {
+	for i := range commands {
+		switch commands[i].Name {
 		case "model", "adapter":
-			path := c.Args
+			path := commands[i].Args
 			if path == "~" {
 				path = home
 			} else if strings.HasPrefix(path, "~/") {
@@ -91,7 +92,7 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 			}
 
 			fi, err := os.Stat(path)
-			if errors.Is(err, os.ErrNotExist) && c.Name == "model" {
+			if errors.Is(err, os.ErrNotExist) && commands[i].Name == "model" {
 				continue
 			} else if err != nil {
 				return err
@@ -114,13 +115,7 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 				return err
 			}
 
-			name := c.Name
-			if c.Name == "model" {
-				name = "from"
-			}
-
-			re := regexp.MustCompile(fmt.Sprintf(`(?im)^(%s)\s+%s\s*$`, name, c.Args))
-			modelfile = re.ReplaceAll(modelfile, []byte("$1 @"+digest))
+			commands[i].Args = "@"+digest
 		}
 	}
 
@@ -150,7 +145,7 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 
 	quantization, _ := cmd.Flags().GetString("quantization")
 
-	request := api.CreateRequest{Name: args[0], Modelfile: string(modelfile), Quantization: quantization}
+	request := api.CreateRequest{Name: args[0], Modelfile: parser.Format(commands), Quantization: quantization}
 	if err := client.Create(cmd.Context(), &request, fn); err != nil {
 		return err
 	}
@@ -1125,7 +1120,7 @@ Environment Variables:
 		RunE:    ListHandler,
 	}
 	copyCmd := &cobra.Command{
-		Use:     "cp SOURCE TARGET",
+		Use:     "cp SOURCE DESTINATION",
 		Short:   "Copy a model",
 		Args:    cobra.ExactArgs(2),
 		PreRunE: checkServerHeartbeat,

cmd/interactive.go

@@ -94,6 +94,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 		fmt.Fprintln(os.Stderr, "  /show           Show model information")
 		fmt.Fprintln(os.Stderr, "  /load <model>   Load a session or model")
 		fmt.Fprintln(os.Stderr, "  /save <model>   Save your current session")
+		fmt.Fprintln(os.Stderr, "  /clear          Clear session context")
 		fmt.Fprintln(os.Stderr, "  /bye            Exit")
 		fmt.Fprintln(os.Stderr, "  /?, /help       Help for a command")
 		fmt.Fprintln(os.Stderr, "  /? shortcuts    Help for keyboard shortcuts")
@@ -280,6 +281,10 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 			}
 			fmt.Printf("Created new model '%s'\n", args[1])
 			continue
+		case strings.HasPrefix(line, "/clear"):
+			opts.Messages = []api.Message{}
+			fmt.Println("Cleared session context")
+			continue
 		case strings.HasPrefix(line, "/set"):
 			args := strings.Fields(line)
 			if len(args) > 1 {

convert/convert.go

@@ -5,6 +5,7 @@ import (
 	"encoding/binary"
 	"encoding/json"
 	"fmt"
+	"io"
 	"log/slog"
 	"os"
 	"path/filepath"
@@ -47,7 +48,7 @@ type ByteOrder interface {
 type ModelArch interface {
 	GetTensors() error
 	LoadVocab() error
-	WriteGGUF() (string, error)
+	WriteGGUF(io.WriteSeeker) error
 }
 
 type ModelFormat interface {

convert/gemma.go

@@ -94,7 +94,7 @@ func (m *GemmaModel) LoadVocab() error {
 	return nil
 }
 
-func (m *GemmaModel) WriteGGUF() (string, error) {
+func (m *GemmaModel) WriteGGUF(ws io.WriteSeeker) error {
 	kv := llm.KV{
 		"general.architecture":                   "gemma",
 		"general.name":                           m.Name,
@@ -122,16 +122,5 @@ func (m *GemmaModel) WriteGGUF() (string, error) {
 		"tokenizer.ggml.add_eos_token":    false,
 	}
 
-	f, err := os.CreateTemp("", "ollama-gguf")
-	if err != nil {
-		return "", err
-	}
-	defer f.Close()
-
-	mod := llm.NewGGUFV3(m.Params.ByteOrder)
-	if err := mod.Encode(f, kv, m.Tensors); err != nil {
-		return "", err
-	}
-
-	return f.Name(), nil
+	return llm.NewGGUFV3(m.Params.ByteOrder).Encode(ws, kv, m.Tensors)
 }

convert/llama.go

@@ -132,7 +132,7 @@ func (m *LlamaModel) LoadVocab() error {
 	return nil
 }
 
-func (m *LlamaModel) WriteGGUF() (string, error) {
+func (m *LlamaModel) WriteGGUF(ws io.WriteSeeker) error {
 	kv := llm.KV{
 		"general.architecture":                   "llama",
 		"general.name":                           m.Name,
@@ -161,16 +161,9 @@ func (m *LlamaModel) WriteGGUF() (string, error) {
 
 	f, err := os.CreateTemp("", "ollama-gguf")
 	if err != nil {
-		return "", err
+		return err
 	}
 	defer f.Close()
 
-	mod := llm.NewGGUFV3(m.Params.ByteOrder)
-	if err := mod.Encode(f, kv, m.Tensors); err != nil {
-		return "", err
-	}
-
-	slog.Debug(fmt.Sprintf("gguf file = %s", f.Name()))
-
-	return f.Name(), nil
+	return llm.NewGGUFV3(m.Params.ByteOrder).Encode(f, kv, m.Tensors)
 }

convert/mistral.go

@@ -132,7 +132,7 @@ func (m *MistralModel) LoadVocab() error {
 	return nil
 }
 
-func (m *MistralModel) WriteGGUF() (string, error) {
+func (m *MistralModel) WriteGGUF(ws io.WriteSeeker) error {
 	kv := llm.KV{
 		"general.architecture":                   "llama",
 		"general.name":                           m.Name,
@@ -158,16 +158,5 @@ func (m *MistralModel) WriteGGUF() (string, error) {
 		"tokenizer.ggml.unknown_token_id": uint32(0),
 	}
 
-	f, err := os.CreateTemp("", "ollama-gguf")
-	if err != nil {
-		return "", err
-	}
-	defer f.Close()
-
-	mod := llm.NewGGUFV3(m.Params.ByteOrder)
-	if err := mod.Encode(f, kv, m.Tensors); err != nil {
-		return "", err
-	}
-
-	return f.Name(), nil
+	return llm.NewGGUFV3(m.Params.ByteOrder).Encode(ws, kv, m.Tensors)
 }

convert/mixtral.go

@@ -1,7 +1,7 @@
 package convert
 
 import (
-	"os"
+	"io"
 	"regexp"
 
 	"github.com/ollama/ollama/llm"
@@ -47,7 +47,7 @@ func (m *MixtralModel) LoadVocab() error {
 	return nil
 }
 
-func (m *MixtralModel) WriteGGUF() (string, error) {
+func (m *MixtralModel) WriteGGUF(ws io.WriteSeeker) error {
 	kv := llm.KV{
 		"general.architecture":          "llama",
 		"general.name":                  m.Name,
@@ -81,16 +81,5 @@ func (m *MixtralModel) WriteGGUF() (string, error) {
 		"tokenizer.ggml.add_eos_token":    false,
 	}
 
-	f, err := os.CreateTemp("", "ollama-gguf")
-	if err != nil {
-		return "", err
-	}
-	defer f.Close()
-
-	mod := llm.NewGGUFV3(m.Params.ByteOrder)
-	if err := mod.Encode(f, kv, m.Tensors); err != nil {
-		return "", err
-	}
-
-	return f.Name(), nil
+	return llm.NewGGUFV3(m.Params.ByteOrder).Encode(ws, kv, m.Tensors)
 }

docs/development.md

@@ -51,7 +51,7 @@ Typically the build scripts will auto-detect CUDA, however, if your Linux distro
 or installation approach uses unusual paths, you can specify the location by
 specifying an environment variable `CUDA_LIB_DIR` to the location of the shared
 libraries, and `CUDACXX` to the location of the nvcc compiler. You can customize
-set set of target CUDA architectues by setting `CMAKE_CUDA_ARCHITECTURES` (e.g. "50;60;70")
+a set of target CUDA architectures by setting `CMAKE_CUDA_ARCHITECTURES` (e.g. "50;60;70")
 
 Then generate dependencies:
 
@@ -142,4 +142,4 @@ In addition to the common Windows development tools described above, install AMD
 - [AMD HIP](https://www.amd.com/en/developer/resources/rocm-hub/hip-sdk.html)
 - [Strawberry Perl](https://strawberryperl.com/)
 
-Lastly, add `ninja.exe` included with MSVC to the system path (e.g. `C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\Common7\IDE\CommonExtensions\Microsoft\CMake\Ninja`).
+Lastly, add `ninja.exe` included with MSVC to the system path (e.g. `C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\Common7\IDE\CommonExtensions\Microsoft\CMake\Ninja`).

docs/windows.md

@@ -1,47 +1,47 @@
-# Ollama Windows Preview
-
-Welcome to the Ollama Windows preview.
-
-No more WSL required!
-
-Ollama now runs as a native Windows application, including NVIDIA and AMD Radeon GPU support.
-After installing Ollama Windows Preview, Ollama will run in the background and
-the `ollama` command line is available in `cmd`, `powershell` or your favorite
-terminal application. As usual the Ollama [api](./api.md) will be served on
-`http://localhost:11434`.
-
-As this is a preview release, you should expect a few bugs here and there.  If
-you run into a problem you can reach out on
-[Discord](https://discord.gg/ollama), or file an 
-[issue](https://github.com/ollama/ollama/issues).
-Logs will often be helpful in diagnosing the problem (see
-[Troubleshooting](#troubleshooting) below)
-
-## System Requirements
-
-* Windows 10 or newer, Home or Pro
-* NVIDIA 452.39 or newer Drivers if you have an NVIDIA card
-* AMD Radeon Driver https://www.amd.com/en/support if you have a Radeon card
-
-## API Access
-
-Here's a quick example showing API access from `powershell`
-```powershell
-(Invoke-WebRequest -method POST -Body '{"model":"llama2", "prompt":"Why is the sky blue?", "stream": false}' -uri http://localhost:11434/api/generate ).Content | ConvertFrom-json
-```
-
-## 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
-server.
-
-Ollama on Windows stores files in a few different locations.  You can view them in
-the explorer window by hitting `<cmd>+R` and type in:
-- `explorer %LOCALAPPDATA%\Ollama` contains logs, and downloaded updates
-    - *app.log* contains logs from the GUI application
-    - *server.log* contains the server logs
-    - *upgrade.log* contains log output for upgrades
-- `explorer %LOCALAPPDATA%\Programs\Ollama` contains the binaries (The installer adds this to your user PATH)
-- `explorer %HOMEPATH%\.ollama` contains models and configuration
-- `explorer %TEMP%` contains temporary executable files in one or more `ollama*` directories
+# Ollama Windows Preview
+
+Welcome to the Ollama Windows preview.
+
+No more WSL required!
+
+Ollama now runs as a native Windows application, including NVIDIA and AMD Radeon GPU support.
+After installing Ollama Windows Preview, Ollama will run in the background and
+the `ollama` command line is available in `cmd`, `powershell` or your favorite
+terminal application. As usual the Ollama [api](./api.md) will be served on
+`http://localhost:11434`.
+
+As this is a preview release, you should expect a few bugs here and there.  If
+you run into a problem you can reach out on
+[Discord](https://discord.gg/ollama), or file an
+[issue](https://github.com/ollama/ollama/issues).
+Logs will often be helpful in diagnosing the problem (see
+[Troubleshooting](#troubleshooting) below)
+
+## System Requirements
+
+* Windows 10 or newer, Home or Pro
+* NVIDIA 452.39 or newer Drivers if you have an NVIDIA card
+* AMD Radeon Driver https://www.amd.com/en/support if you have a Radeon card
+
+## API Access
+
+Here's a quick example showing API access from `powershell`
+```powershell
+(Invoke-WebRequest -method POST -Body '{"model":"llama2", "prompt":"Why is the sky blue?", "stream": false}' -uri http://localhost:11434/api/generate ).Content | ConvertFrom-json
+```
+
+## 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
+server.
+
+Ollama on Windows stores files in a few different locations.  You can view them in
+the explorer window by hitting `<cmd>+R` and type in:
+- `explorer %LOCALAPPDATA%\Ollama` contains logs, and downloaded updates
+    - *app.log* contains logs from the GUI application
+    - *server.log* contains the server logs
+    - *upgrade.log* contains log output for upgrades
+- `explorer %LOCALAPPDATA%\Programs\Ollama` contains the binaries (The installer adds this to your user PATH)
+- `explorer %HOMEPATH%\.ollama` contains models and configuration
+- `explorer %TEMP%` contains temporary executable files in one or more `ollama*` directories

integration/utils_test.go

@@ -107,7 +107,7 @@ func startServer(ctx context.Context, ollamaHost string) error {
 
 	if tmp := os.Getenv("OLLAMA_HOST"); tmp != ollamaHost {
 		slog.Info("setting env", "OLLAMA_HOST", ollamaHost)
-		os.Setenv("OLLAMA_HOST", ollamaHost)
+		t.Setenv("OLLAMA_HOST", ollamaHost)
 	}
 
 	slog.Info("starting server", "url", ollamaHost)

llm/filetype.go

@@ -0,0 +1,140 @@
+package llm
+
+import "fmt"
+
+type fileType uint32
+
+const (
+	fileTypeF32 fileType = iota
+	fileTypeF16
+	fileTypeQ4_0
+	fileTypeQ4_1
+	fileTypeQ4_1_F16
+	fileTypeQ4_2 // unused
+	fileTypeQ4_3 // unused
+	fileTypeQ8_0
+	fileTypeQ5_0
+	fileTypeQ5_1
+	fileTypeQ2_K
+	fileTypeQ3_K_S
+	fileTypeQ3_K_M
+	fileTypeQ3_K_L
+	fileTypeQ4_K_S
+	fileTypeQ4_K_M
+	fileTypeQ5_K_S
+	fileTypeQ5_K_M
+	fileTypeQ6_K
+	fileTypeIQ2_XXS
+	fileTypeIQ2_XS
+	fileTypeQ2_K_S
+	fileTypeQ3_K_XS
+	fileTypeIQ3_XXS
+
+	fileTypeUnknown
+)
+
+func ParseFileType(s string) (fileType, error) {
+	switch s {
+	case "F32":
+		return fileTypeF32, nil
+	case "F16":
+		return fileTypeF16, nil
+	case "Q4_0":
+		return fileTypeQ4_0, nil
+	case "Q4_1":
+		return fileTypeQ4_1, nil
+	case "Q4_1_F16":
+		return fileTypeQ4_1_F16, nil
+	case "Q8_0":
+		return fileTypeQ8_0, nil
+	case "Q5_0":
+		return fileTypeQ5_0, nil
+	case "Q5_1":
+		return fileTypeQ5_1, nil
+	case "Q2_K":
+		return fileTypeQ2_K, nil
+	case "Q3_K_S":
+		return fileTypeQ3_K_S, nil
+	case "Q3_K_M":
+		return fileTypeQ3_K_M, nil
+	case "Q3_K_L":
+		return fileTypeQ3_K_L, nil
+	case "Q4_K_S":
+		return fileTypeQ4_K_S, nil
+	case "Q4_K_M":
+		return fileTypeQ4_K_M, nil
+	case "Q5_K_S":
+		return fileTypeQ5_K_S, nil
+	case "Q5_K_M":
+		return fileTypeQ5_K_M, nil
+	case "Q6_K":
+		return fileTypeQ6_K, nil
+	case "IQ2_XXS":
+		return fileTypeIQ2_XXS, nil
+	case "IQ2_XS":
+		return fileTypeIQ2_XS, nil
+	case "Q2_K_S":
+		return fileTypeQ2_K_S, nil
+	case "Q3_K_XS":
+		return fileTypeQ3_K_XS, nil
+	case "IQ3_XXS":
+		return fileTypeIQ3_XXS, nil
+	default:
+		return fileTypeUnknown, fmt.Errorf("unknown fileType: %s", s)
+	}
+}
+
+func (t fileType) String() string {
+	switch t {
+	case fileTypeF32:
+		return "F32"
+	case fileTypeF16:
+		return "F16"
+	case fileTypeQ4_0:
+		return "Q4_0"
+	case fileTypeQ4_1:
+		return "Q4_1"
+	case fileTypeQ4_1_F16:
+		return "Q4_1_F16"
+	case fileTypeQ8_0:
+		return "Q8_0"
+	case fileTypeQ5_0:
+		return "Q5_0"
+	case fileTypeQ5_1:
+		return "Q5_1"
+	case fileTypeQ2_K:
+		return "Q2_K"
+	case fileTypeQ3_K_S:
+		return "Q3_K_S"
+	case fileTypeQ3_K_M:
+		return "Q3_K_M"
+	case fileTypeQ3_K_L:
+		return "Q3_K_L"
+	case fileTypeQ4_K_S:
+		return "Q4_K_S"
+	case fileTypeQ4_K_M:
+		return "Q4_K_M"
+	case fileTypeQ5_K_S:
+		return "Q5_K_S"
+	case fileTypeQ5_K_M:
+		return "Q5_K_M"
+	case fileTypeQ6_K:
+		return "Q6_K"
+	case fileTypeIQ2_XXS:
+		return "IQ2_XXS"
+	case fileTypeIQ2_XS:
+		return "IQ2_XS"
+	case fileTypeQ2_K_S:
+		return "Q2_K_S"
+	case fileTypeQ3_K_XS:
+		return "Q3_K_XS"
+	case fileTypeIQ3_XXS:
+		return "IQ3_XXS"
+	default:
+		return "unknown"
+	}
+}
+
+func (t fileType) Value() uint32 {
+	return uint32(t)
+}

llm/ggla.go

@@ -33,6 +33,7 @@ func (c *containerGGLA) Decode(rs io.ReadSeeker) (model, error) {
 
 type ggla struct {
 	*containerGGLA
+	offset int64
 
 	kv      KV
 	tensors []*Tensor
@@ -53,6 +54,10 @@ func (llm *ggla) Tensors() Tensors {
 	return llm.tensors
 }
 
+func (llm *ggla) Offset() int64 {
+	return llm.offset
+}
+
 func (llm *ggla) decode(rs io.ReadSeeker) error {
 	var r uint32
 	if err := binary.Read(rs, binary.LittleEndian, &r); err != nil {
@@ -66,6 +71,13 @@ func (llm *ggla) decode(rs io.ReadSeeker) error {
 	}
 	llm.kv["alpha"] = alpha
 
+	offset, err := rs.Seek(0, io.SeekCurrent)
+	if err != nil {
+		return err
+	}
+
+	llm.offset = offset
+
 	for {
 		var dims uint32
 		if err := binary.Read(rs, binary.LittleEndian, &dims); err != nil {

llm/ggml.go

@@ -13,85 +13,10 @@ type GGML struct {
 	model
 }
 
-const (
-	fileTypeF32 uint32 = iota
-	fileTypeF16
-	fileTypeQ4_0
-	fileTypeQ4_1
-	fileTypeQ4_1_F16
-	fileTypeQ8_0 uint32 = iota + 2
-	fileTypeQ5_0
-	fileTypeQ5_1
-	fileTypeQ2_K
-	fileTypeQ3_K_S
-	fileTypeQ3_K_M
-	fileTypeQ3_K_L
-	fileTypeQ4_K_S
-	fileTypeQ4_K_M
-	fileTypeQ5_K_S
-	fileTypeQ5_K_M
-	fileTypeQ6_K
-	fileTypeIQ2_XXS
-	fileTypeIQ2_XS
-	fileTypeQ2_K_S
-	fileTypeQ3_K_XS
-	fileTypeIQ3_XXS
-)
-
-func fileType(fileType uint32) string {
-	switch fileType {
-	case fileTypeF32:
-		return "F32"
-	case fileTypeF16:
-		return "F16"
-	case fileTypeQ4_0:
-		return "Q4_0"
-	case fileTypeQ4_1:
-		return "Q4_1"
-	case fileTypeQ4_1_F16:
-		return "Q4_1_F16"
-	case fileTypeQ8_0:
-		return "Q8_0"
-	case fileTypeQ5_0:
-		return "Q5_0"
-	case fileTypeQ5_1:
-		return "Q5_1"
-	case fileTypeQ2_K:
-		return "Q2_K"
-	case fileTypeQ3_K_S:
-		return "Q3_K_S"
-	case fileTypeQ3_K_M:
-		return "Q3_K_M"
-	case fileTypeQ3_K_L:
-		return "Q3_K_L"
-	case fileTypeQ4_K_S:
-		return "Q4_K_S"
-	case fileTypeQ4_K_M:
-		return "Q4_K_M"
-	case fileTypeQ5_K_S:
-		return "Q5_K_S"
-	case fileTypeQ5_K_M:
-		return "Q5_K_M"
-	case fileTypeQ6_K:
-		return "Q6_K"
-	case fileTypeIQ2_XXS:
-		return "IQ2_XXS"
-	case fileTypeIQ2_XS:
-		return "IQ2_XS"
-	case fileTypeQ2_K_S:
-		return "Q2_K_S"
-	case fileTypeQ3_K_XS:
-		return "Q3_K_XS"
-	case fileTypeIQ3_XXS:
-		return "IQ3_XXS"
-	default:
-		return "unknown"
-	}
-}
-
 type model interface {
 	KV() KV
 	Tensors() Tensors
+	Offset() int64
 }
 
 type KV map[string]any
@@ -123,7 +48,7 @@ func (kv KV) ParameterCount() uint64 {
 
 func (kv KV) FileType() string {
 	if u64 := kv.u64("general.file_type"); u64 > 0 {
-		return fileType(uint32(u64))
+		return fileType(uint32(u64)).String()
 	}
 
 	return "unknown"
@@ -286,6 +211,23 @@ const (
 
 var ErrUnsupportedFormat = errors.New("unsupported model format")
 
+func DetectGGMLType(b []byte) string {
+	switch binary.LittleEndian.Uint32(b[:4]) {
+	case FILE_MAGIC_GGML:
+		return "ggml"
+	case FILE_MAGIC_GGMF:
+		return "ggmf"
+	case FILE_MAGIC_GGJT:
+		return "ggjt"
+	case FILE_MAGIC_GGLA:
+		return "ggla"
+	case FILE_MAGIC_GGUF_LE, FILE_MAGIC_GGUF_BE:
+		return "gguf"
+	default:
+		return ""
+	}
+}
+
 func DecodeGGML(rs io.ReadSeeker) (*GGML, int64, error) {
 	var magic uint32
 	if err := binary.Read(rs, binary.LittleEndian, &magic); err != nil {

llm/gguf.go

@@ -55,7 +55,7 @@ func (c *containerGGUF) Decode(rs io.ReadSeeker) (model, error) {
 
 	model := newGGUF(c)
 	slog.Debug(fmt.Sprintf("model = %#v", model))
-	if err := model.Decode(rs); err != nil {
+	if err := model.decode(rs); err != nil {
 		return nil, err
 	}
 
@@ -90,6 +90,7 @@ const (
 
 type gguf struct {
 	*containerGGUF
+	offset int64
 
 	kv      KV
 	tensors []*Tensor
@@ -116,6 +117,10 @@ func (llm *gguf) Tensors() Tensors {
 	return llm.tensors
 }
 
+func (llm *gguf) Offset() int64 {
+	return llm.offset
+}
+
 func (llm *gguf) numTensor() uint64 {
 	switch llm.Version {
 	case 1:
@@ -138,7 +143,7 @@ func (llm *gguf) numKV() uint64 {
 	}
 }
 
-func (llm *gguf) Decode(rs io.ReadSeeker) error {
+func (llm *gguf) decode(rs io.ReadSeeker) error {
 	// decode key-values
 	for i := 0; uint64(i) < llm.numKV(); i++ {
 		k, err := readGGUFString(llm, rs)
@@ -250,6 +255,8 @@ func (llm *gguf) Decode(rs io.ReadSeeker) error {
 		return err
 	}
 
+	llm.offset = offset + padding
+
 	for _, tensor := range llm.tensors {
 		if _, err := rs.Seek(int64(tensor.size()), io.SeekCurrent); err != nil {
 			return err

llm/llm.go

@@ -20,7 +20,7 @@ func SystemInfo() string {
 	return C.GoString(C.llama_print_system_info())
 }
 
-func Quantize(infile, outfile, filetype string) error {
+func Quantize(infile, outfile string, ftype fileType) error {
 	cinfile := C.CString(infile)
 	defer C.free(unsafe.Pointer(cinfile))
 
@@ -29,58 +29,10 @@ func Quantize(infile, outfile, filetype string) error {
 
 	params := C.llama_model_quantize_default_params()
 	params.nthread = -1
+	params.ftype = ftype.Value()
 
-	switch filetype {
-	case "F32":
-		params.ftype = fileTypeF32
-	case "F16":
-		params.ftype = fileTypeF16
-	case "Q4_0":
-		params.ftype = fileTypeQ4_0
-	case "Q4_1":
-		params.ftype = fileTypeQ4_1
-	case "Q4_1_F16":
-		params.ftype = fileTypeQ4_1_F16
-	case "Q8_0":
-		params.ftype = fileTypeQ8_0
-	case "Q5_0":
-		params.ftype = fileTypeQ5_0
-	case "Q5_1":
-		params.ftype = fileTypeQ5_1
-	case "Q2_K":
-		params.ftype = fileTypeQ2_K
-	case "Q3_K_S":
-		params.ftype = fileTypeQ3_K_S
-	case "Q3_K_M":
-		params.ftype = fileTypeQ3_K_M
-	case "Q3_K_L":
-		params.ftype = fileTypeQ3_K_L
-	case "Q4_K_S":
-		params.ftype = fileTypeQ4_K_S
-	case "Q4_K_M":
-		params.ftype = fileTypeQ4_K_M
-	case "Q5_K_S":
-		params.ftype = fileTypeQ5_K_S
-	case "Q5_K_M":
-		params.ftype = fileTypeQ5_K_M
-	case "Q6_K":
-		params.ftype = fileTypeQ6_K
-	case "IQ2_XXS":
-		params.ftype = fileTypeIQ2_XXS
-	case "IQ2_XS":
-		params.ftype = fileTypeIQ2_XS
-	case "Q2_K_S":
-		params.ftype = fileTypeQ2_K_S
-	case "Q3_K_XS":
-		params.ftype = fileTypeQ3_K_XS
-	case "IQ3_XXS":
-		params.ftype = fileTypeIQ3_XXS
-	default:
-		return fmt.Errorf("unknown filetype: %s", filetype)
-	}
-
-	if retval := C.llama_model_quantize(cinfile, coutfile, &params); retval != 0 {
-		return fmt.Errorf("llama_model_quantize: %d", retval)
+	if rc := C.llama_model_quantize(cinfile, coutfile, &params); rc != 0 {
+		return fmt.Errorf("llama_model_quantize: %d", rc)
 	}
 
 	return nil

llm/server.go

@@ -300,12 +300,6 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 			continue
 		}
 
-		// reap subprocess when it exits
-		go func() {
-			// Exit status managed via getServerStatus
-			_ = s.cmd.Wait()
-		}()
-
 		// TODO - make sure this is all wired up correctly
 		// if err = s.WaitUntilRunning(); err != nil {
 		// 	slog.Error("error starting llama server", "server", servers[i], "error", err)
@@ -899,7 +893,13 @@ func (s *llmServer) Detokenize(ctx context.Context, tokens []int) (string, error
 func (s *llmServer) Close() error {
 	if s.cmd != nil {
 		slog.Debug("stopping llama server")
-		return s.cmd.Process.Kill()
+		if err := s.cmd.Process.Kill(); err != nil {
+			return err
+		}
+
+		_ = s.cmd.Wait()
+
+		slog.Debug("llama server stopped")
 	}
 
 	return nil

parser/parser.go

@@ -6,8 +6,8 @@ import (
 	"errors"
 	"fmt"
 	"io"
-	"log/slog"
-	"slices"
+	"strconv"
+	"strings"
 )
 
 type Command struct {
@@ -15,118 +15,283 @@ type Command struct {
 	Args string
 }
 
-func (c *Command) Reset() {
-	c.Name = ""
-	c.Args = ""
-}
+type state int
 
-func Parse(reader io.Reader) ([]Command, error) {
-	var commands []Command
-	var command, modelCommand Command
+const (
+	stateNil state = iota
+	stateName
+	stateValue
+	stateParameter
+	stateMessage
+	stateComment
+)
 
-	scanner := bufio.NewScanner(reader)
-	scanner.Buffer(make([]byte, 0, bufio.MaxScanTokenSize), bufio.MaxScanTokenSize)
-	scanner.Split(scanModelfile)
-	for scanner.Scan() {
-		line := scanner.Bytes()
+var (
+	errMissingFrom        = errors.New("no FROM line")
+	errInvalidMessageRole = errors.New("message role must be one of \"system\", \"user\", or \"assistant\"")
+	errInvalidCommand     = errors.New("command must be one of \"from\", \"license\", \"template\", \"system\", \"adapter\", \"parameter\", or \"message\"")
+)
 
-		fields := bytes.SplitN(line, []byte(" "), 2)
-		if len(fields) == 0 || len(fields[0]) == 0 {
-			continue
-		}
+func Format(cmds []Command) string {
+	var sb strings.Builder
+	for _, cmd := range cmds {
+		name := cmd.Name
+		args := cmd.Args
 
-		switch string(bytes.ToUpper(fields[0])) {
-		case "FROM":
-			command.Name = "model"
-			command.Args = string(bytes.TrimSpace(fields[1]))
-			// copy command for validation
-			modelCommand = command
-		case "ADAPTER":
-			command.Name = string(bytes.ToLower(fields[0]))
-			command.Args = string(bytes.TrimSpace(fields[1]))
-		case "LICENSE", "TEMPLATE", "SYSTEM", "PROMPT":
-			command.Name = string(bytes.ToLower(fields[0]))
-			command.Args = string(fields[1])
-		case "PARAMETER":
-			fields = bytes.SplitN(fields[1], []byte(" "), 2)
-			if len(fields) < 2 {
-				return nil, fmt.Errorf("missing value for %s", fields)
-			}
-
-			command.Name = string(fields[0])
-			command.Args = string(bytes.TrimSpace(fields[1]))
-		case "EMBED":
-			return nil, fmt.Errorf("deprecated command: EMBED is no longer supported, use the /embed API endpoint instead")
-		case "MESSAGE":
-			command.Name = string(bytes.ToLower(fields[0]))
-			fields = bytes.SplitN(fields[1], []byte(" "), 2)
-			if len(fields) < 2 {
-				return nil, fmt.Errorf("should be in the format <role> <message>")
-			}
-			if !slices.Contains([]string{"system", "user", "assistant"}, string(bytes.ToLower(fields[0]))) {
-				return nil, fmt.Errorf("role must be one of \"system\", \"user\", or \"assistant\"")
-			}
-			command.Args = fmt.Sprintf("%s: %s", string(bytes.ToLower(fields[0])), string(fields[1]))
+		switch cmd.Name {
+		case "model":
+			name = "from"
+			args = cmd.Args
+		case "license", "template", "system", "adapter":
+			args = quote(args)
+		case "message":
+			role, message, _ := strings.Cut(cmd.Args, ": ")
+			args = role + " " + quote(message)
 		default:
-			if !bytes.HasPrefix(fields[0], []byte("#")) {
-				// log a warning for unknown commands
-				slog.Warn(fmt.Sprintf("Unknown command: %s", fields[0]))
-			}
-			continue
+			name = "parameter"
+			args = cmd.Name + " " + quote(cmd.Args)
 		}
 
-		commands = append(commands, command)
-		command.Reset()
+		fmt.Fprintln(&sb, strings.ToUpper(name), args)
 	}
 
-	if modelCommand.Args == "" {
-		return nil, errors.New("no FROM line for the model was specified")
-	}
-
-	return commands, scanner.Err()
+	return sb.String()
 }
 
-func scanModelfile(data []byte, atEOF bool) (advance int, token []byte, err error) {
-	advance, token, err = scan([]byte(`"""`), []byte(`"""`), data, atEOF)
-	if err != nil {
-		return 0, nil, err
-	}
+func Parse(r io.Reader) (cmds []Command, err error) {
+	var cmd Command
+	var curr state
+	var b bytes.Buffer
+	var role string
 
-	if advance > 0 && token != nil {
-		return advance, token, nil
-	}
-
-	advance, token, err = scan([]byte(`"`), []byte(`"`), data, atEOF)
-	if err != nil {
-		return 0, nil, err
-	}
-
-	if advance > 0 && token != nil {
-		return advance, token, nil
-	}
-
-	return bufio.ScanLines(data, atEOF)
-}
-
-func scan(openBytes, closeBytes, data []byte, atEOF bool) (advance int, token []byte, err error) {
-	newline := bytes.IndexByte(data, '\n')
-
-	if start := bytes.Index(data, openBytes); start >= 0 && start < newline {
-		end := bytes.Index(data[start+len(openBytes):], closeBytes)
-		if end < 0 {
-			if atEOF {
-				return 0, nil, fmt.Errorf("unterminated %s: expecting %s", openBytes, closeBytes)
-			} else {
-				return 0, nil, nil
-			}
+	br := bufio.NewReader(r)
+	for {
+		r, _, err := br.ReadRune()
+		if errors.Is(err, io.EOF) {
+			break
+		} else if err != nil {
+			return nil, err
 		}
 
-		n := start + len(openBytes) + end + len(closeBytes)
+		next, r, err := parseRuneForState(r, curr)
+		if errors.Is(err, io.ErrUnexpectedEOF) {
+			return nil, fmt.Errorf("%w: %s", err, b.String())
+		} else if err != nil {
+			return nil, err
+		}
 
-		newData := data[:start]
-		newData = append(newData, data[start+len(openBytes):n-len(closeBytes)]...)
-		return n, newData, nil
+		// process the state transition, some transitions need to be intercepted and redirected
+		if next != curr {
+			switch curr {
+			case stateName:
+				if !isValidCommand(b.String()) {
+					return nil, errInvalidCommand
+				}
+
+				// next state sometimes depends on the current buffer value
+				switch s := strings.ToLower(b.String()); s {
+				case "from":
+					cmd.Name = "model"
+				case "parameter":
+					// transition to stateParameter which sets command name
+					next = stateParameter
+				case "message":
+					// transition to stateMessage which validates the message role
+					next = stateMessage
+					fallthrough
+				default:
+					cmd.Name = s
+				}
+			case stateParameter:
+				cmd.Name = b.String()
+			case stateMessage:
+				if !isValidMessageRole(b.String()) {
+					return nil, errInvalidMessageRole
+				}
+
+				role = b.String()
+			case stateComment, stateNil:
+				// pass
+			case stateValue:
+				s, ok := unquote(b.String())
+				if !ok || isSpace(r) {
+					if _, err := b.WriteRune(r); err != nil {
+						return nil, err
+					}
+
+					continue
+				}
+
+				if role != "" {
+					s = role + ": " + s
+					role = ""
+				}
+
+				cmd.Args = s
+				cmds = append(cmds, cmd)
+			}
+
+			b.Reset()
+			curr = next
+		}
+
+		if strconv.IsPrint(r) {
+			if _, err := b.WriteRune(r); err != nil {
+				return nil, err
+			}
+		}
 	}
 
-	return 0, nil, nil
+	// flush the buffer
+	switch curr {
+	case stateComment, stateNil:
+		// pass; nothing to flush
+	case stateValue:
+		s, ok := unquote(b.String())
+		if !ok {
+			return nil, io.ErrUnexpectedEOF
+		}
+
+		if role != "" {
+			s = role + ": " + s
+		}
+
+		cmd.Args = s
+		cmds = append(cmds, cmd)
+	default:
+		return nil, io.ErrUnexpectedEOF
+	}
+
+	for _, cmd := range cmds {
+		if cmd.Name == "model" {
+			return cmds, nil
+		}
+	}
+
+	return nil, errMissingFrom
+}
+
+func parseRuneForState(r rune, cs state) (state, rune, error) {
+	switch cs {
+	case stateNil:
+		switch {
+		case r == '#':
+			return stateComment, 0, nil
+		case isSpace(r), isNewline(r):
+			return stateNil, 0, nil
+		default:
+			return stateName, r, nil
+		}
+	case stateName:
+		switch {
+		case isAlpha(r):
+			return stateName, r, nil
+		case isSpace(r):
+			return stateValue, 0, nil
+		default:
+			return stateNil, 0, errInvalidCommand
+		}
+	case stateValue:
+		switch {
+		case isNewline(r):
+			return stateNil, r, nil
+		case isSpace(r):
+			return stateNil, r, nil
+		default:
+			return stateValue, r, nil
+		}
+	case stateParameter:
+		switch {
+		case isAlpha(r), isNumber(r), r == '_':
+			return stateParameter, r, nil
+		case isSpace(r):
+			return stateValue, 0, nil
+		default:
+			return stateNil, 0, io.ErrUnexpectedEOF
+		}
+	case stateMessage:
+		switch {
+		case isAlpha(r):
+			return stateMessage, r, nil
+		case isSpace(r):
+			return stateValue, 0, nil
+		default:
+			return stateNil, 0, io.ErrUnexpectedEOF
+		}
+	case stateComment:
+		switch {
+		case isNewline(r):
+			return stateNil, 0, nil
+		default:
+			return stateComment, 0, nil
+		}
+	default:
+		return stateNil, 0, errors.New("")
+	}
+}
+
+func quote(s string) string {
+	if strings.Contains(s, "\n") || strings.HasPrefix(s, " ") || strings.HasSuffix(s, " ") {
+		if strings.Contains(s, "\"") {
+			return `"""` + s + `"""`
+		}
+
+		return `"` + s + `"`
+	}
+
+	return s
+}
+
+func unquote(s string) (string, bool) {
+	if len(s) == 0 {
+		return "", false
+	}
+
+	// TODO: single quotes
+	if len(s) >= 3 && s[:3] == `"""` {
+		if len(s) >= 6 && s[len(s)-3:] == `"""` {
+			return s[3 : len(s)-3], true
+		}
+
+		return "", false
+	}
+
+	if len(s) >= 1 && s[0] == '"' {
+		if len(s) >= 2 && s[len(s)-1] == '"' {
+			return s[1 : len(s)-1], true
+		}
+
+		return "", false
+	}
+
+	return s, true
+}
+
+func isAlpha(r rune) bool {
+	return r >= 'a' && r <= 'z' || r >= 'A' && r <= 'Z'
+}
+
+func isNumber(r rune) bool {
+	return r >= '0' && r <= '9'
+}
+
+func isSpace(r rune) bool {
+	return r == ' ' || r == '\t'
+}
+
+func isNewline(r rune) bool {
+	return r == '\r' || r == '\n'
+}
+
+func isValidMessageRole(role string) bool {
+	return role == "system" || role == "user" || role == "assistant"
+}
+
+func isValidCommand(cmd string) bool {
+	switch strings.ToLower(cmd) {
+	case "from", "license", "template", "system", "adapter", "parameter", "message":
+		return true
+	default:
+		return false
+	}
 }

parser/parser_test.go

@@ -1,14 +1,16 @@
 package parser
 
 import (
+	"bytes"
+	"fmt"
+	"io"
 	"strings"
 	"testing"
 
 	"github.com/stretchr/testify/assert"
 )
 
-func Test_Parser(t *testing.T) {
-
+func TestParser(t *testing.T) {
 	input := `
 FROM model1
 ADAPTER adapter1
@@ -35,21 +37,62 @@ TEMPLATE template1
 	assert.Equal(t, expectedCommands, commands)
 }
 
-func Test_Parser_NoFromLine(t *testing.T) {
+func TestParserFrom(t *testing.T) {
+	var cases = []struct {
+		input    string
+		expected []Command
+		err      error
+	}{
+		{
+			"FROM foo",
+			[]Command{{Name: "model", Args: "foo"}},
+			nil,
+		},
+		{
+			"FROM /path/to/model",
+			[]Command{{Name: "model", Args: "/path/to/model"}},
+			nil,
+		},
+		{
+			"FROM /path/to/model/fp16.bin",
+			[]Command{{Name: "model", Args: "/path/to/model/fp16.bin"}},
+			nil,
+		},
+		{
+			"FROM llama3:latest",
+			[]Command{{Name: "model", Args: "llama3:latest"}},
+			nil,
+		},
+		{
+			"FROM llama3:7b-instruct-q4_K_M",
+			[]Command{{Name: "model", Args: "llama3:7b-instruct-q4_K_M"}},
+			nil,
+		},
+		{
+			"", nil, errMissingFrom,
+		},
+		{
+			"PARAMETER param1 value1",
+			nil,
+			errMissingFrom,
+		},
+		{
+			"PARAMETER param1 value1\nFROM foo",
+			[]Command{{Name: "param1", Args: "value1"}, {Name: "model", Args: "foo"}},
+			nil,
+		},
+	}
 
-	input := `
-PARAMETER param1 value1
-PARAMETER param2 value2
-`
-
-	reader := strings.NewReader(input)
-
-	_, err := Parse(reader)
-	assert.ErrorContains(t, err, "no FROM line")
+	for _, c := range cases {
+		t.Run("", func(t *testing.T) {
+			commands, err := Parse(strings.NewReader(c.input))
+			assert.ErrorIs(t, err, c.err)
+			assert.Equal(t, c.expected, commands)
+		})
+	}
 }
 
-func Test_Parser_MissingValue(t *testing.T) {
-
+func TestParserParametersMissingValue(t *testing.T) {
 	input := `
 FROM foo
 PARAMETER param1
@@ -58,41 +101,401 @@ PARAMETER param1
 	reader := strings.NewReader(input)
 
 	_, err := Parse(reader)
-	assert.ErrorContains(t, err, "missing value for [param1]")
+	assert.ErrorIs(t, err, io.ErrUnexpectedEOF)
+}
+
+func TestParserBadCommand(t *testing.T) {
+	input := `
+FROM foo
+BADCOMMAND param1 value1
+`
+	_, err := Parse(strings.NewReader(input))
+	assert.ErrorIs(t, err, errInvalidCommand)
 
 }
 
-func Test_Parser_Messages(t *testing.T) {
-
-	input := `
+func TestParserMessages(t *testing.T) {
+	var cases = []struct {
+		input    string
+		expected []Command
+		err      error
+	}{
+		{
+			`
+FROM foo
+MESSAGE system You are a Parser. Always Parse things.
+`,
+			[]Command{
+				{Name: "model", Args: "foo"},
+				{Name: "message", Args: "system: You are a Parser. Always Parse things."},
+			},
+			nil,
+		},
+		{
+			`
+FROM foo
+MESSAGE system You are a Parser. Always Parse things.`,
+			[]Command{
+				{Name: "model", Args: "foo"},
+				{Name: "message", Args: "system: You are a Parser. Always Parse things."},
+			},
+			nil,
+		},
+		{
+			`
 FROM foo
 MESSAGE system You are a Parser. Always Parse things.
 MESSAGE user Hey there!
 MESSAGE assistant Hello, I want to parse all the things!
-`
-
-	reader := strings.NewReader(input)
-	commands, err := Parse(reader)
-	assert.Nil(t, err)
-
-	expectedCommands := []Command{
-		{Name: "model", Args: "foo"},
-		{Name: "message", Args: "system: You are a Parser. Always Parse things."},
-		{Name: "message", Args: "user: Hey there!"},
-		{Name: "message", Args: "assistant: Hello, I want to parse all the things!"},
-	}
-
-	assert.Equal(t, expectedCommands, commands)
-}
-
-func Test_Parser_Messages_BadRole(t *testing.T) {
-
-	input := `
+`,
+			[]Command{
+				{Name: "model", Args: "foo"},
+				{Name: "message", Args: "system: You are a Parser. Always Parse things."},
+				{Name: "message", Args: "user: Hey there!"},
+				{Name: "message", Args: "assistant: Hello, I want to parse all the things!"},
+			},
+			nil,
+		},
+		{
+			`
+FROM foo
+MESSAGE system """
+You are a multiline Parser. Always Parse things.
+"""
+			`,
+			[]Command{
+				{Name: "model", Args: "foo"},
+				{Name: "message", Args: "system: \nYou are a multiline Parser. Always Parse things.\n"},
+			},
+			nil,
+		},
+		{
+			`
 FROM foo
 MESSAGE badguy I'm a bad guy!
-`
+`,
+			nil,
+			errInvalidMessageRole,
+		},
+		{
+			`
+FROM foo
+MESSAGE system
+`,
+			nil,
+			io.ErrUnexpectedEOF,
+		},
+		{
+			`
+FROM foo
+MESSAGE system`,
+			nil,
+			io.ErrUnexpectedEOF,
+		},
+	}
+
+	for _, c := range cases {
+		t.Run("", func(t *testing.T) {
+			commands, err := Parse(strings.NewReader(c.input))
+			assert.ErrorIs(t, err, c.err)
+			assert.Equal(t, c.expected, commands)
+		})
+	}
+}
+
+func TestParserQuoted(t *testing.T) {
+	var cases = []struct {
+		multiline string
+		expected  []Command
+		err       error
+	}{
+		{
+			`
+FROM foo
+SYSTEM """
+This is a
+multiline system.
+"""
+			`,
+			[]Command{
+				{Name: "model", Args: "foo"},
+				{Name: "system", Args: "\nThis is a\nmultiline system.\n"},
+			},
+			nil,
+		},
+		{
+			`
+FROM foo
+SYSTEM """
+This is a
+multiline system."""
+			`,
+			[]Command{
+				{Name: "model", Args: "foo"},
+				{Name: "system", Args: "\nThis is a\nmultiline system."},
+			},
+			nil,
+		},
+		{
+			`
+FROM foo
+SYSTEM """This is a
+multiline system."""
+			`,
+			[]Command{
+				{Name: "model", Args: "foo"},
+				{Name: "system", Args: "This is a\nmultiline system."},
+			},
+			nil,
+		},
+		{
+			`
+FROM foo
+SYSTEM """This is a multiline system."""
+			`,
+			[]Command{
+				{Name: "model", Args: "foo"},
+				{Name: "system", Args: "This is a multiline system."},
+			},
+			nil,
+		},
+		{
+			`
+FROM foo
+SYSTEM """This is a multiline system.""
+			`,
+			nil,
+			io.ErrUnexpectedEOF,
+		},
+		{
+			`
+FROM foo
+SYSTEM "
+			`,
+			nil,
+			io.ErrUnexpectedEOF,
+		},
+		{
+			`
+FROM foo
+SYSTEM """
+This is a multiline system with "quotes".
+"""
+`,
+			[]Command{
+				{Name: "model", Args: "foo"},
+				{Name: "system", Args: "\nThis is a multiline system with \"quotes\".\n"},
+			},
+			nil,
+		},
+		{
+			`
+FROM foo
+SYSTEM """"""
+`,
+			[]Command{
+				{Name: "model", Args: "foo"},
+				{Name: "system", Args: ""},
+			},
+			nil,
+		},
+		{
+			`
+FROM foo
+SYSTEM ""
+`,
+			[]Command{
+				{Name: "model", Args: "foo"},
+				{Name: "system", Args: ""},
+			},
+			nil,
+		},
+		{
+			`
+FROM foo
+SYSTEM "'"
+`,
+			[]Command{
+				{Name: "model", Args: "foo"},
+				{Name: "system", Args: "'"},
+			},
+			nil,
+		},
+		{
+			`
+FROM foo
+SYSTEM """''"'""'""'"'''''""'""'"""
+`,
+			[]Command{
+				{Name: "model", Args: "foo"},
+				{Name: "system", Args: `''"'""'""'"'''''""'""'`},
+			},
+			nil,
+		},
+		{
+			`
+FROM foo
+TEMPLATE """
+{{ .Prompt }}
+"""`,
+			[]Command{
+				{Name: "model", Args: "foo"},
+				{Name: "template", Args: "\n{{ .Prompt }}\n"},
+			},
+			nil,
+		},
+	}
+
+	for _, c := range cases {
+		t.Run("", func(t *testing.T) {
+			commands, err := Parse(strings.NewReader(c.multiline))
+			assert.ErrorIs(t, err, c.err)
+			assert.Equal(t, c.expected, commands)
+		})
+	}
+}
+
+func TestParserParameters(t *testing.T) {
+	var cases = map[string]struct {
+		name, value string
+	}{
+		"numa true":                    {"numa", "true"},
+		"num_ctx 1":                    {"num_ctx", "1"},
+		"num_batch 1":                  {"num_batch", "1"},
+		"num_gqa 1":                    {"num_gqa", "1"},
+		"num_gpu 1":                    {"num_gpu", "1"},
+		"main_gpu 1":                   {"main_gpu", "1"},
+		"low_vram true":                {"low_vram", "true"},
+		"f16_kv true":                  {"f16_kv", "true"},
+		"logits_all true":              {"logits_all", "true"},
+		"vocab_only true":              {"vocab_only", "true"},
+		"use_mmap true":                {"use_mmap", "true"},
+		"use_mlock true":               {"use_mlock", "true"},
+		"num_thread 1":                 {"num_thread", "1"},
+		"num_keep 1":                   {"num_keep", "1"},
+		"seed 1":                       {"seed", "1"},
+		"num_predict 1":                {"num_predict", "1"},
+		"top_k 1":                      {"top_k", "1"},
+		"top_p 1.0":                    {"top_p", "1.0"},
+		"tfs_z 1.0":                    {"tfs_z", "1.0"},
+		"typical_p 1.0":                {"typical_p", "1.0"},
+		"repeat_last_n 1":              {"repeat_last_n", "1"},
+		"temperature 1.0":              {"temperature", "1.0"},
+		"repeat_penalty 1.0":           {"repeat_penalty", "1.0"},
+		"presence_penalty 1.0":         {"presence_penalty", "1.0"},
+		"frequency_penalty 1.0":        {"frequency_penalty", "1.0"},
+		"mirostat 1":                   {"mirostat", "1"},
+		"mirostat_tau 1.0":             {"mirostat_tau", "1.0"},
+		"mirostat_eta 1.0":             {"mirostat_eta", "1.0"},
+		"penalize_newline true":        {"penalize_newline", "true"},
+		"stop ### User:":               {"stop", "### User:"},
+		"stop ### User: ":              {"stop", "### User: "},
+		"stop \"### User:\"":           {"stop", "### User:"},
+		"stop \"### User: \"":          {"stop", "### User: "},
+		"stop \"\"\"### User:\"\"\"":   {"stop", "### User:"},
+		"stop \"\"\"### User:\n\"\"\"": {"stop", "### User:\n"},
+		"stop <|endoftext|>":           {"stop", "<|endoftext|>"},
+		"stop <|eot_id|>":              {"stop", "<|eot_id|>"},
+		"stop </s>":                    {"stop", "</s>"},
+	}
+
+	for k, v := range cases {
+		t.Run(k, func(t *testing.T) {
+			var b bytes.Buffer
+			fmt.Fprintln(&b, "FROM foo")
+			fmt.Fprintln(&b, "PARAMETER", k)
+			commands, err := Parse(&b)
+			assert.Nil(t, err)
+
+			assert.Equal(t, []Command{
+				{Name: "model", Args: "foo"},
+				{Name: v.name, Args: v.value},
+			}, commands)
+		})
+	}
+}
+
+func TestParserComments(t *testing.T) {
+	var cases = []struct {
+		input    string
+		expected []Command
+	}{
+		{
+			`
+# comment
+FROM foo
+	`,
+			[]Command{
+				{Name: "model", Args: "foo"},
+			},
+		},
+	}
+
+	for _, c := range cases {
+		t.Run("", func(t *testing.T) {
+			commands, err := Parse(strings.NewReader(c.input))
+			assert.Nil(t, err)
+			assert.Equal(t, c.expected, commands)
+		})
+	}
+}
+
+func TestParseFormatParse(t *testing.T) {
+	var cases = []string{
+		`
+FROM foo
+ADAPTER adapter1
+LICENSE MIT
+PARAMETER param1 value1
+PARAMETER param2 value2
+TEMPLATE template1
+MESSAGE system You are a Parser. Always Parse things.
+MESSAGE user Hey there!
+MESSAGE assistant Hello, I want to parse all the things!
+`,
+		`
+FROM foo
+ADAPTER adapter1
+LICENSE MIT
+PARAMETER param1 value1
+PARAMETER param2 value2
+TEMPLATE template1
+MESSAGE system """
+You are a store greeter. Always responsed with "Hello!".
+"""
+MESSAGE user Hey there!
+MESSAGE assistant Hello, I want to parse all the things!
+`,
+		`
+FROM foo
+ADAPTER adapter1
+LICENSE """
+Very long and boring legal text.
+Blah blah blah.
+"Oh look, a quote!"
+"""
+
+PARAMETER param1 value1
+PARAMETER param2 value2
+TEMPLATE template1
+MESSAGE system """
+You are a store greeter. Always responsed with "Hello!".
+"""
+MESSAGE user Hey there!
+MESSAGE assistant Hello, I want to parse all the things!
+`,
+	}
+
+	for _, c := range cases {
+		t.Run("", func(t *testing.T) {
+			commands, err := Parse(strings.NewReader(c))
+			assert.NoError(t, err)
+
+			commands2, err := Parse(strings.NewReader(Format(commands)))
+			assert.NoError(t, err)
+
+			assert.Equal(t, commands, commands2)
+		})
+	}
 
-	reader := strings.NewReader(input)
-	_, err := Parse(reader)
-	assert.ErrorContains(t, err, "role must be one of \"system\", \"user\", or \"assistant\"")
 }

server/images.go

@@ -1,8 +1,8 @@
 package server
 
 import (
-	"archive/zip"
 	"bytes"
+	"cmp"
 	"context"
 	"crypto/sha256"
 	"encoding/base64"
@@ -11,7 +11,6 @@ import (
 	"errors"
 	"fmt"
 	"io"
-	"io/fs"
 	"log"
 	"log/slog"
 	"net/http"
@@ -21,13 +20,11 @@ import (
 	"runtime"
 	"strconv"
 	"strings"
-	"text/template"
 
 	"golang.org/x/exp/slices"
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/auth"
-	"github.com/ollama/ollama/convert"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/llm"
 	"github.com/ollama/ollama/parser"
@@ -64,6 +61,48 @@ func (m *Model) IsEmbedding() bool {
 	return slices.Contains(m.Config.ModelFamilies, "bert") || slices.Contains(m.Config.ModelFamilies, "nomic-bert")
 }
 
+func (m *Model) Commands() (cmds []parser.Command) {
+	cmds = append(cmds, parser.Command{Name: "model", Args: m.ModelPath})
+
+	if m.Template != "" {
+		cmds = append(cmds, parser.Command{Name: "template", Args: m.Template})
+	}
+
+	if m.System != "" {
+		cmds = append(cmds, parser.Command{Name: "system", Args: m.System})
+	}
+
+	for _, adapter := range m.AdapterPaths {
+		cmds = append(cmds, parser.Command{Name: "adapter", Args: adapter})
+	}
+
+	for _, projector := range m.ProjectorPaths {
+		cmds = append(cmds, parser.Command{Name: "projector", Args: projector})
+	}
+
+	for k, v := range m.Options {
+		switch v := v.(type) {
+		case []any:
+			for _, s := range v {
+				cmds = append(cmds, parser.Command{Name: k, Args: fmt.Sprintf("%v", s)})
+			}
+		default:
+			cmds = append(cmds, parser.Command{Name: k, Args: fmt.Sprintf("%v", v)})
+		}
+	}
+
+	for _, license := range m.License {
+		cmds = append(cmds, parser.Command{Name: "license", Args: license})
+	}
+
+	for _, msg := range m.Messages {
+		cmds = append(cmds, parser.Command{Name: "message", Args: fmt.Sprintf("%s %s", msg.Role, msg.Content)})
+	}
+
+	return cmds
+
+}
+
 type Message struct {
 	Role    string `json:"role"`
 	Content string `json:"content"`
@@ -89,36 +128,6 @@ type ConfigV2 struct {
 	RootFS       RootFS `json:"rootfs"`
 }
 
-func (c *ConfigV2) SetModelFormat(format string) {
-	if c.ModelFormat == "" {
-		c.ModelFormat = format
-	}
-}
-
-func (c *ConfigV2) SetModelFamily(families ...string) {
-	for _, family := range families {
-		if c.ModelFamily == "" {
-			c.ModelFamily = family
-		}
-
-		if !slices.Contains(c.ModelFamilies, family) {
-			c.ModelFamilies = append(c.ModelFamilies, family)
-		}
-	}
-}
-
-func (c *ConfigV2) SetModelType(modelType string) {
-	if c.ModelType == "" {
-		c.ModelType = modelType
-	}
-}
-
-func (c *ConfigV2) SetFileType(fileType string) {
-	if c.FileType == "" {
-		c.FileType = fileType
-	}
-}
-
 type RootFS struct {
 	Type    string   `json:"type"`
 	DiffIDs []string `json:"diff_ids"`
@@ -199,6 +208,14 @@ func GetModel(name string) (*Model, error) {
 
 		switch layer.MediaType {
 		case "application/vnd.ollama.image.model":
+			model.ModelPath = filename
+			model.ParentModel = layer.From
+		case "application/vnd.ollama.image.model+metadata", "application/vnd.ollama.image.model+data":
+			filename, err = GetBlobsPath(layer.MergeBase)
+			if err != nil {
+				return nil, err
+			}
+
 			model.ModelPath = filename
 			model.ParentModel = layer.From
 		case "application/vnd.ollama.image.embed":
@@ -263,7 +280,7 @@ func GetModel(name string) (*Model, error) {
 	return model, nil
 }
 
-func realpath(mfDir, from string) string {
+func realpath(rel, from string) string {
 	abspath, err := filepath.Abs(from)
 	if err != nil {
 		return from
@@ -280,22 +297,15 @@ func realpath(mfDir, from string) string {
 		return filepath.Join(home, from[2:])
 	}
 
-	if _, err := os.Stat(filepath.Join(mfDir, from)); err == nil {
+	if _, err := os.Stat(filepath.Join(rel, from)); err == nil {
 		// this is a file relative to the Modelfile
-		return filepath.Join(mfDir, from)
+		return filepath.Join(rel, from)
 	}
 
 	return abspath
 }
 
-func CreateModel(ctx context.Context, name, modelFileDir, quantization string, commands []parser.Command, fn func(resp api.ProgressResponse)) error {
-	deleteMap := make(map[string]struct{})
-	if manifest, _, err := GetManifest(ParseModelPath(name)); err == nil {
-		for _, layer := range append(manifest.Layers, manifest.Config) {
-			deleteMap[layer.Digest] = struct{}{}
-		}
-	}
-
+func CreateModel(ctx context.Context, name, modelFileDir, quantization string, commands []parser.Command, fn func(resp api.ProgressResponse)) (err error) {
 	config := ConfigV2{
 		OS:           "linux",
 		Architecture: "amd64",
@@ -304,250 +314,222 @@ func CreateModel(ctx context.Context, name, modelFileDir, quantization string, c
 		},
 	}
 
-	var layers Layers
-	messages := []string{}
-
-	params := make(map[string][]string)
-	fromParams := make(map[string]any)
+	var messages []*api.Message
+	parameters := make(map[string]any)
 
+	var layers []*Layer
 	for _, c := range commands {
 		mediatype := fmt.Sprintf("application/vnd.ollama.image.%s", c.Name)
 
 		switch c.Name {
-		case "model":
-			if strings.HasPrefix(c.Args, "@") {
-				blobPath, err := GetBlobsPath(strings.TrimPrefix(c.Args, "@"))
+		case "model", "adapter":
+			var baseLayers []*layerWithGGML
+			if name := model.ParseName(c.Args); name.IsValid() {
+				baseLayers, err = parseFromModel(ctx, name, fn)
+				if err != nil {
+					return err
+				}
+			} else if strings.HasPrefix(c.Args, "@") {
+				blobpath, err := GetBlobsPath(strings.TrimPrefix(c.Args, "@"))
 				if err != nil {
 					return err
 				}
 
-				c.Args = blobPath
-			}
-
-			pathName := realpath(modelFileDir, c.Args)
-
-			ggufName, err := convertModel(name, pathName, fn)
-			if err != nil {
-				var pathErr *fs.PathError
-				switch {
-				case errors.Is(err, zip.ErrFormat):
-					// it's not a safetensor archive
-				case errors.As(err, &pathErr):
-					// it's not a file on disk, could be a model reference
-				default:
-					return err
-				}
-			}
-
-			if ggufName != "" {
-				pathName = ggufName
-				defer os.RemoveAll(ggufName)
-
-				if quantization != "" {
-					quantization = strings.ToUpper(quantization)
-					fn(api.ProgressResponse{Status: fmt.Sprintf("quantizing %s model to %s", "F16", quantization)})
-					tempfile, err := os.CreateTemp(filepath.Dir(ggufName), quantization)
-					if err != nil {
-						return err
-					}
-					defer os.RemoveAll(tempfile.Name())
-
-					if err := llm.Quantize(ggufName, tempfile.Name(), quantization); err != nil {
-						return err
-					}
-
-					if err := tempfile.Close(); err != nil {
-						return err
-					}
-
-					pathName = tempfile.Name()
-				}
-			}
-
-			bin, err := os.Open(pathName)
-			if err != nil {
-				// not a file on disk so must be a model reference
-				modelpath := ParseModelPath(c.Args)
-				manifest, _, err := GetManifest(modelpath)
-				switch {
-				case errors.Is(err, os.ErrNotExist):
-					fn(api.ProgressResponse{Status: "pulling model"})
-					if err := PullModel(ctx, c.Args, &registryOptions{}, fn); err != nil {
-						return err
-					}
-
-					manifest, _, err = GetManifest(modelpath)
-					if err != nil {
-						return err
-					}
-				case err != nil:
-					return err
-				}
-
-				fn(api.ProgressResponse{Status: "reading model metadata"})
-				fromConfigPath, err := GetBlobsPath(manifest.Config.Digest)
+				blob, err := os.Open(blobpath)
 				if err != nil {
 					return err
 				}
+				defer blob.Close()
 
-				fromConfigFile, err := os.Open(fromConfigPath)
+				baseLayers, err = parseFromFile(ctx, blob, fn)
 				if err != nil {
 					return err
 				}
-				defer fromConfigFile.Close()
+			} else if file, err := os.Open(realpath(modelFileDir, c.Args)); err == nil {
+				defer file.Close()
 
-				var fromConfig ConfigV2
-				if err := json.NewDecoder(fromConfigFile).Decode(&fromConfig); err != nil {
+				baseLayers, err = parseFromFile(ctx, file, fn)
+				if err != nil {
 					return err
 				}
+			} else {
+				return fmt.Errorf("invalid model reference: %s", c.Args)
+			}
 
-				// if the model is still not in gguf format, error out
-				if fromConfig.ModelFormat != "gguf" {
-					return fmt.Errorf("%s is not in gguf format, this base model is not compatible with this version of ollama", c.Args)
-				}
-
-				config.SetModelFormat(fromConfig.ModelFormat)
-				config.SetModelFamily(append(fromConfig.ModelFamilies, fromConfig.ModelFamily)...)
-				config.SetModelType(fromConfig.ModelType)
-				config.SetFileType(fromConfig.FileType)
-
-				for _, layer := range manifest.Layers {
-					deleteMap[layer.Digest] = struct{}{}
-					if layer.MediaType == "application/vnd.ollama.image.params" {
-						fromParamsPath, err := GetBlobsPath(layer.Digest)
-						if err != nil {
-							return err
-						}
-
-						fromParamsFile, err := os.Open(fromParamsPath)
-						if err != nil {
-							return err
-						}
-						defer fromParamsFile.Close()
-
-						if err := json.NewDecoder(fromParamsFile).Decode(&fromParams); err != nil {
-							return err
-						}
-					}
-
-					layer, err := NewLayerFromLayer(layer.Digest, layer.MediaType, modelpath.GetShortTagname())
+			for _, baseLayer := range baseLayers {
+				if quantization != "" && baseLayer.MediaType == "application/vnd.ollama.image.model" {
+					ftype, err := llm.ParseFileType(quantization)
 					if err != nil {
 						return err
 					}
 
-					layers.Add(layer)
+					filetype := baseLayer.GGML.KV().FileType()
+					if !slices.Contains([]string{"F16", "F32"}, filetype) {
+						return errors.New("quantization is only supported for F16 and F32 models")
+					}
+
+					fn(api.ProgressResponse{Status: fmt.Sprintf("quantizing %s model to %s", filetype, quantization)})
+
+					blob, err := GetBlobsPath(baseLayer.Digest)
+					if err != nil {
+						return err
+					}
+
+					temp, err := os.CreateTemp(filepath.Dir(blob), quantization)
+					if err != nil {
+						return err
+					}
+					defer temp.Close()
+					defer os.Remove(temp.Name())
+
+					if err := llm.Quantize(blob, temp.Name(), ftype); err != nil {
+						return err
+					}
+
+					baseLayer.Layer, err = NewLayer(temp, baseLayer.Layer.MediaType)
+					if err != nil {
+						return err
+					}
 				}
 
-				deleteMap[manifest.Config.Digest] = struct{}{}
-				continue
+				if baseLayer.GGML != nil {
+					config.ModelFormat = cmp.Or(config.ModelFormat, baseLayer.GGML.Name())
+					config.ModelFamily = cmp.Or(config.ModelFamily, baseLayer.GGML.KV().Architecture())
+					config.ModelType = cmp.Or(config.ModelType, format.HumanNumber(baseLayer.GGML.KV().ParameterCount()))
+					config.FileType = cmp.Or(config.FileType, baseLayer.GGML.KV().FileType())
+					config.ModelFamilies = append(config.ModelFamilies, baseLayer.GGML.KV().Architecture())
+
+					f, err := baseLayer.Layer.Open()
+					if err != nil {
+						return err
+					}
+					defer f.Close()
+
+					metadata := io.NewSectionReader(f, 0, baseLayer.GGML.Offset())
+					metadataLayer, err := NewLayer(metadata, "application/vnd.ollama.image.model+metadata")
+					if err != nil {
+						return err
+					}
+					metadataLayer.Intermediate = true
+					metadataLayer.MergeBase = baseLayer.Digest
+
+					layers = append(layers, metadataLayer)
+
+					metadataPath, err := GetBlobsPath(metadataLayer.Digest)
+					if err != nil {
+						return err
+					}
+					defer os.Remove(metadataPath)
+
+					stat, err := f.Stat()
+					if err != nil {
+						return err
+					}
+
+					data := io.NewSectionReader(f, baseLayer.GGML.Offset(), stat.Size())
+					dataLayer, err := NewLayer(data, "application/vnd.ollama.image.model+data")
+					if err != nil {
+						return err
+					}
+					dataLayer.Intermediate = true
+					dataLayer.MergeBase = baseLayer.Digest
+
+					layers = append(layers, dataLayer)
+
+					dataPath, err := GetBlobsPath(dataLayer.Digest)
+					if err != nil {
+						return err
+					}
+					defer os.Remove(dataPath)
+					continue
+				}
+
+				layers = append(layers, baseLayer.Layer)
 			}
-			defer bin.Close()
-
-			var offset int64
-			for {
-				fn(api.ProgressResponse{Status: "creating model layer"})
-				if _, err := bin.Seek(offset, io.SeekStart); err != nil {
-					return err
-				}
-
-				ggml, size, err := llm.DecodeGGML(bin)
-				if errors.Is(err, io.EOF) {
-					break
-				} else if errors.Is(err, llm.ErrUnsupportedFormat) {
-					return fmt.Errorf("model binary specified in FROM field is not a valid gguf format model, %w", err)
-				} else if err != nil {
-					return err
-				}
-
-				config.SetModelFormat(ggml.Name())
-				config.SetModelFamily(ggml.KV().Architecture())
-				config.SetModelType(format.HumanNumber(ggml.KV().ParameterCount()))
-				config.SetFileType(ggml.KV().FileType())
-
-				mediatype := mediatype
-				if ggml.KV().Architecture() == "clip" {
-					mediatype = "application/vnd.ollama.image.projector"
-				}
-
-				sr := io.NewSectionReader(bin, offset, size)
-				layer, err := NewLayer(sr, mediatype)
-				if err != nil {
-					return err
-				}
-
-				layers.Add(layer)
-
-				offset += size
-			}
-		case "adapter":
-			if strings.HasPrefix(c.Args, "@") {
-				blobPath, err := GetBlobsPath(strings.TrimPrefix(c.Args, "@"))
-				if err != nil {
-					return err
-				}
-
-				c.Args = blobPath
-			}
-
-			fn(api.ProgressResponse{Status: "creating adapter layer"})
-			bin, err := os.Open(realpath(modelFileDir, c.Args))
-			if err != nil {
-				return err
-			}
-			defer bin.Close()
-
-			_, size, err := llm.DecodeGGML(bin)
+		case "license", "template", "system":
+			blob := strings.NewReader(c.Args)
+			layer, err := NewLayer(blob, mediatype)
 			if err != nil {
 				return err
 			}
 
-			sr := io.NewSectionReader(bin, 0, size)
-			layer, err := NewLayer(sr, mediatype)
-			if err != nil {
-				return err
+			if c.Name != "license" {
+				// replace
+				layers = slices.DeleteFunc(layers, func(layer *Layer) bool {
+					return layer.MediaType == mediatype
+				})
 			}
 
-			layers.Add(layer)
-		case "license":
-			fn(api.ProgressResponse{Status: "creating license layer"})
-
-			bin := strings.NewReader(c.Args)
-			layer, err := NewLayer(bin, mediatype)
-			if err != nil {
-				return err
-			}
-
-			layers.Add(layer)
-		case "template", "system":
-			fn(api.ProgressResponse{Status: fmt.Sprintf("creating %s layer", c.Name)})
-
-			bin := strings.NewReader(c.Args)
-			layer, err := NewLayer(bin, mediatype)
-			if err != nil {
-				return err
-			}
-
-			layers.Replace(layer)
+			layers = append(layers, layer)
 		case "message":
-			messages = append(messages, c.Args)
+			role, content, ok := strings.Cut(c.Args, ": ")
+			if !ok {
+				return fmt.Errorf("invalid message: %s", c.Args)
+			}
+
+			messages = append(messages, &api.Message{Role: role, Content: content})
 		default:
-			params[c.Name] = append(params[c.Name], c.Args)
+			ps, err := api.FormatParams(map[string][]string{c.Name: {c.Args}})
+			if err != nil {
+				return err
+			}
+
+			for k, v := range ps {
+				if ks, ok := parameters[k].([]string); ok {
+					parameters[k] = append(ks, v.([]string)...)
+				} else if vs, ok := v.([]string); ok {
+					parameters[k] = vs
+				} else {
+					parameters[k] = v
+				}
+			}
 		}
 	}
 
-	if len(messages) > 0 {
-		fn(api.ProgressResponse{Status: "creating parameters layer"})
+	var err2 error
+	layers = slices.DeleteFunc(layers, func(layer *Layer) bool {
+		switch layer.MediaType {
+		case "application/vnd.ollama.image.message":
+			// if there are new messages, remove the inherited ones
+			if len(messages) > 0 {
+				return true
+			}
 
-		msgs := make([]api.Message, 0)
+			return false
+		case "application/vnd.ollama.image.params":
+			// merge inherited parameters with new ones
+			r, err := layer.Open()
+			if err != nil {
+				err2 = err
+				return false
+			}
+			defer r.Close()
 
-		for _, m := range messages {
-			// todo: handle images
-			msg := strings.SplitN(m, ": ", 2)
-			msgs = append(msgs, api.Message{Role: msg[0], Content: msg[1]})
+			var ps map[string]any
+			if err := json.NewDecoder(r).Decode(&ps); err != nil {
+				err2 = err
+				return false
+			}
+
+			for k, v := range ps {
+				if _, ok := parameters[k]; !ok {
+					parameters[k] = v
+				}
+			}
+
+			return true
+		default:
+			return false
 		}
+	})
 
+	if err2 != nil {
+		return err2
+	}
+
+	if len(messages) > 0 {
 		var b bytes.Buffer
-		if err := json.NewEncoder(&b).Encode(msgs); err != nil {
+		if err := json.NewEncoder(&b).Encode(messages); err != nil {
 			return err
 		}
 
@@ -556,39 +538,25 @@ func CreateModel(ctx context.Context, name, modelFileDir, quantization string, c
 			return err
 		}
 
-		layers.Replace(layer)
+		layers = append(layers, layer)
 	}
 
-	if len(params) > 0 {
-		fn(api.ProgressResponse{Status: "creating parameters layer"})
-
-		formattedParams, err := api.FormatParams(params)
-		if err != nil {
-			return err
-		}
-
-		for k, v := range fromParams {
-			if _, ok := formattedParams[k]; !ok {
-				formattedParams[k] = v
-			}
-		}
-
+	if len(parameters) > 0 {
 		var b bytes.Buffer
-		if err := json.NewEncoder(&b).Encode(formattedParams); err != nil {
+		if err := json.NewEncoder(&b).Encode(parameters); err != nil {
 			return err
 		}
 
-		fn(api.ProgressResponse{Status: "creating config layer"})
 		layer, err := NewLayer(&b, "application/vnd.ollama.image.params")
 		if err != nil {
 			return err
 		}
 
-		layers.Replace(layer)
+		layers = append(layers, layer)
 	}
 
-	digests := make([]string, len(layers.items))
-	for i, layer := range layers.items {
+	digests := make([]string, len(layers))
+	for i, layer := range layers {
 		digests[i] = layer.Digest
 	}
 
@@ -599,36 +567,38 @@ func CreateModel(ctx context.Context, name, modelFileDir, quantization string, c
 		return err
 	}
 
-	configLayer, err := NewLayer(&b, "application/vnd.docker.container.image.v1+json")
+	layer, err := NewLayer(&b, "application/vnd.docker.container.image.v1+json")
 	if err != nil {
 		return err
 	}
 
-	delete(deleteMap, configLayer.Digest)
+	for _, layer := range append(layers, layer) {
+		if layer.message != "" {
+			fn(api.ProgressResponse{Status: layer.message})
+		}
+	}
 
-	for _, layer := range append(layers.items, configLayer) {
-		committed, err := layer.Commit()
-		if err != nil {
-			return err
+	unref := make(map[string]struct{})
+	if manifest, _, err := GetManifest(ParseModelPath(name)); err == nil {
+		for _, layer := range manifest.Layers {
+			if !slices.Contains(digests, layer.Digest) {
+				unref[layer.Digest] = struct{}{}
+			}
 		}
 
-		status := "writing layer"
-		if !committed {
-			status = "using already created layer"
+		if manifest.Config.Digest != layer.Digest {
+			unref[manifest.Config.Digest] = struct{}{}
 		}
-
-		fn(api.ProgressResponse{Status: fmt.Sprintf("%s %s", status, layer.Digest)})
-
-		delete(deleteMap, layer.Digest)
 	}
 
 	fn(api.ProgressResponse{Status: "writing manifest"})
-	if err := WriteManifest(name, configLayer, layers.items); err != nil {
+	if err := WriteManifest(name, layer, layers); err != nil {
 		return err
 	}
 
-	if noprune := os.Getenv("OLLAMA_NOPRUNE"); noprune == "" {
-		if err := deleteUnusedLayers(nil, deleteMap, false); err != nil {
+	if os.Getenv("OLLAMA_NOPRUNE") == "" && len(unref) > 0 {
+		fn(api.ProgressResponse{Status: "removing unused layers"})
+		if err := deleteUnusedLayers(nil, unref, false); err != nil {
 			return err
 		}
 	}
@@ -637,74 +607,6 @@ func CreateModel(ctx context.Context, name, modelFileDir, quantization string, c
 	return nil
 }
 
-func convertModel(name, path string, fn func(resp api.ProgressResponse)) (string, error) {
-	r, err := zip.OpenReader(path)
-	if err != nil {
-		return "", err
-	}
-	defer r.Close()
-
-	tempDir, err := os.MkdirTemp("", "ollama-convert")
-	if err != nil {
-		return "", err
-	}
-	defer os.RemoveAll(tempDir)
-
-	fn(api.ProgressResponse{Status: "unpacking model metadata"})
-	for _, f := range r.File {
-		fpath := filepath.Join(tempDir, f.Name)
-		outFile, err := os.OpenFile(fpath, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, f.Mode())
-		if err != nil {
-			return "", err
-		}
-
-		rc, err := f.Open()
-		if err != nil {
-			return "", err
-		}
-
-		_, err = io.Copy(outFile, rc)
-		if err != nil {
-			return "", err
-		}
-
-		outFile.Close()
-		rc.Close()
-	}
-
-	mf, err := convert.GetModelFormat(tempDir)
-	if err != nil {
-		return "", err
-	}
-
-	params, err := mf.GetParams(tempDir)
-	if err != nil {
-		return "", err
-	}
-
-	mArch, err := mf.GetModelArch(name, tempDir, params)
-	if err != nil {
-		return "", err
-	}
-
-	fn(api.ProgressResponse{Status: "processing tensors"})
-	if err := mArch.GetTensors(); err != nil {
-		return "", err
-	}
-
-	if err := mArch.LoadVocab(); err != nil {
-		return "", err
-	}
-
-	fn(api.ProgressResponse{Status: "converting model"})
-	path, err = mArch.WriteGGUF()
-	if err != nil {
-		return "", err
-	}
-
-	return path, nil
-}
-
 func CopyModel(src, dst model.Name) error {
 	if !dst.IsFullyQualified() {
 		return model.Unqualified(dst)
@@ -774,6 +676,9 @@ func deleteUnusedLayers(skipModelPath *ModelPath, deleteMap map[string]struct{},
 
 		for _, layer := range manifest.Layers {
 			delete(deleteMap, layer.Digest)
+			if layer.MergeBase != "" {
+				delete(deleteMap, layer.MergeBase)
+			}
 		}
 
 		delete(deleteMap, manifest.Config.Digest)
@@ -880,6 +785,9 @@ func DeleteModel(name string) error {
 	deleteMap := make(map[string]struct{})
 	for _, layer := range manifest.Layers {
 		deleteMap[layer.Digest] = struct{}{}
+		if layer.MergeBase != "" {
+			deleteMap[layer.MergeBase] = struct{}{}
+		}
 	}
 	deleteMap[manifest.Config.Digest] = struct{}{}
 
@@ -901,67 +809,6 @@ func DeleteModel(name string) error {
 	return nil
 }
 
-func ShowModelfile(model *Model) (string, error) {
-	var mt struct {
-		*Model
-		From       string
-		Parameters map[string][]any
-	}
-
-	mt.Parameters = make(map[string][]any)
-	for k, v := range model.Options {
-		if s, ok := v.([]any); ok {
-			mt.Parameters[k] = s
-			continue
-		}
-
-		mt.Parameters[k] = []any{v}
-	}
-
-	mt.Model = model
-	mt.From = model.ModelPath
-
-	if model.ParentModel != "" {
-		mt.From = model.ParentModel
-	}
-
-	modelFile := `# Modelfile generated by "ollama show"
-# To build a new Modelfile based on this one, replace the FROM line with:
-# FROM {{ .ShortName }}
-
-FROM {{ .From }}
-TEMPLATE """{{ .Template }}"""
-
-{{- if .System }}
-SYSTEM """{{ .System }}"""
-{{- end }}
-
-{{- range $adapter := .AdapterPaths }}
-ADAPTER {{ $adapter }}
-{{- end }}
-
-{{- range $k, $v := .Parameters }}
-{{- range $parameter := $v }}
-PARAMETER {{ $k }} {{ printf "%#v" $parameter }}
-{{- end }}
-{{- end }}`
-
-	tmpl, err := template.New("").Parse(modelFile)
-	if err != nil {
-		slog.Info(fmt.Sprintf("error parsing template: %q", err))
-		return "", err
-	}
-
-	var buf bytes.Buffer
-
-	if err = tmpl.Execute(&buf, mt); err != nil {
-		slog.Info(fmt.Sprintf("error executing template: %q", err))
-		return "", err
-	}
-
-	return buf.String(), nil
-}
-
 func PushModel(ctx context.Context, name string, regOpts *registryOptions, fn func(api.ProgressResponse)) error {
 	mp := ParseModelPath(name)
 	fn(api.ProgressResponse{Status: "retrieving manifest"})
@@ -980,6 +827,49 @@ func PushModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 	layers = append(layers, manifest.Layers...)
 	layers = append(layers, manifest.Config)
 
+	for _, layer := range layers {
+		if !layer.Intermediate {
+			continue
+		}
+
+		switch layer.MediaType {
+		case "application/vnd.ollama.image.model+metadata", "application/vnd.ollama.image.model+data":
+			if _, err := GetBlobsPath(layer.MergeBase); errors.Is(err, os.ErrNotExist) {
+				filename, err := GetBlobsPath(layer.MergeBase)
+				if err != nil {
+					return err
+				}
+
+				f, err := os.Open(filename)
+				if err != nil {
+					return err
+				}
+				defer f.Close()
+
+				ggml, size, err := llm.DecodeGGML(f)
+				if err != nil {
+					return err
+				}
+
+				if _, err := f.Seek(0, io.SeekStart); err != nil {
+					return err
+				}
+
+				metadata := io.NewSectionReader(f, 0, ggml.Offset())
+				if _, err := NewLayer(metadata, "application/vnd.ollama.image.model+metadata"); err != nil {
+					return err
+				}
+
+				data := io.NewSectionReader(f, ggml.Offset(), size)
+				if _, err := NewLayer(data, "application/vnd.ollama.image.model+metadata"); err != nil {
+					return err
+				}
+			} else if err != nil {
+				return err
+			}
+		}
+	}
+
 	for _, layer := range layers {
 		if err := uploadBlob(ctx, mp, layer, regOpts, fn); err != nil {
 			slog.Info(fmt.Sprintf("error uploading blob: %v", err))
@@ -1049,6 +939,27 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 	layers = append(layers, manifest.Config)
 
 	for _, layer := range layers {
+		if layer.Intermediate {
+			filename, err := GetBlobsPath(layer.MergeBase)
+			if err != nil {
+				return err
+			}
+
+			if _, err := os.Stat(filename); errors.Is(err, os.ErrNotExist) {
+				// pass
+			} else if err != nil {
+				return err
+			} else {
+				fn(api.ProgressResponse{
+					Status:    fmt.Sprintf("pulling %s", layer.Digest[7:19]),
+					Digest:    layer.Digest,
+					Total:     layer.Size,
+					Completed: layer.Size,
+				})
+				continue
+			}
+		}
+
 		if err := downloadBlob(
 			ctx,
 			downloadOpts{
@@ -1063,9 +974,59 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 	}
 	delete(deleteMap, manifest.Config.Digest)
 
+	type mergedLayer struct {
+		Metadata, Data *Layer
+	}
+
+	mergedLayers := make(map[string]mergedLayer)
+	for _, layer := range manifest.Layers {
+		filename, err := GetBlobsPath(layer.MergeBase)
+		if err != nil {
+			return err
+		}
+
+		if _, err := os.Stat(filename); errors.Is(err, os.ErrNotExist) {
+			merged := mergedLayers[layer.MergeBase]
+			if layer.MediaType == "application/vnd.ollama.image.model+metadata" {
+				merged.Metadata = layer
+			} else if layer.MediaType == "application/vnd.ollama.image.model+data" {
+				merged.Data = layer
+			} else {
+				continue
+			}
+
+			mergedLayers[layer.MergeBase] = merged
+		} else if err != nil {
+			return err
+		} else {
+			continue
+		}
+	}
+
+	for _, mergedLayer := range mergedLayers {
+		fn(api.ProgressResponse{Status: "merging layers"})
+		metadata, err := mergedLayer.Metadata.Open()
+		if err != nil {
+			return err
+		}
+		defer metadata.Close()
+
+		data, err := mergedLayer.Data.Open()
+		if err != nil {
+			return err
+		}
+		defer data.Close()
+
+		if _, err := NewLayer(io.MultiReader(metadata, data), "application/vnd.ollama.image.model"); err != nil {
+			return err
+		}
+	}
+
 	fn(api.ProgressResponse{Status: "verifying sha256 digest"})
 	for _, layer := range layers {
-		if err := verifyBlob(layer.Digest); err != nil {
+		if err := verifyBlob(layer.Digest); errors.Is(err, os.ErrNotExist) && layer.Intermediate {
+			// pass
+		} else if err != nil {
 			if errors.Is(err, errDigestMismatch) {
 				// something went wrong, delete the blob
 				fp, err := GetBlobsPath(layer.Digest)

server/layer.go

@@ -5,39 +5,18 @@ import (
 	"fmt"
 	"io"
 	"os"
-	"strings"
-
-	"golang.org/x/exp/slices"
 )
 
-type Layers struct {
-	items []*Layer
-}
-
-func (ls *Layers) Add(layer *Layer) {
-	if layer.Size > 0 {
-		ls.items = append(ls.items, layer)
-	}
-}
-
-func (ls *Layers) Replace(layer *Layer) {
-	if layer.Size > 0 {
-		mediatype := layer.MediaType
-		layers := slices.DeleteFunc(ls.items, func(l *Layer) bool {
-			return l.MediaType == mediatype
-		})
-
-		ls.items = append(layers, layer)
-	}
-}
-
 type Layer struct {
 	MediaType string `json:"mediaType"`
 	Digest    string `json:"digest"`
 	Size      int64  `json:"size"`
 	From      string `json:"from,omitempty"`
 
-	tempFileName string
+	Intermediate bool   `json:"intermediate,omitempty"`
+	MergeBase    string `json:"merge_base,omitempty"`
+
+	message string
 }
 
 func NewLayer(r io.Reader, mediatype string) (*Layer, error) {
@@ -46,14 +25,12 @@ func NewLayer(r io.Reader, mediatype string) (*Layer, error) {
 		return nil, err
 	}
 
-	const delimiter = "-"
-
-	pattern := strings.Join([]string{"sha256", "*-partial"}, delimiter)
-	temp, err := os.CreateTemp(blobs, pattern)
+	temp, err := os.CreateTemp(blobs, "sha256-")
 	if err != nil {
 		return nil, err
 	}
 	defer temp.Close()
+	defer os.Remove(temp.Name())
 
 	sha256sum := sha256.New()
 	n, err := io.Copy(io.MultiWriter(temp, sha256sum), r)
@@ -61,11 +38,29 @@ func NewLayer(r io.Reader, mediatype string) (*Layer, error) {
 		return nil, err
 	}
 
+	if err := temp.Close(); err != nil {
+		return nil, err
+	}
+
+	digest := fmt.Sprintf("sha256:%x", sha256sum.Sum(nil))
+	blob, err := GetBlobsPath(digest)
+	if err != nil {
+		return nil, err
+	}
+
+	status := "using existing layer"
+	if _, err := os.Stat(blob); err != nil {
+		status = "creating new layer"
+		if err := os.Rename(temp.Name(), blob); err != nil {
+			return nil, err
+		}
+	}
+
 	return &Layer{
-		MediaType:    mediatype,
-		Digest:       fmt.Sprintf("sha256:%x", sha256sum.Sum(nil)),
-		Size:         n,
-		tempFileName: temp.Name(),
+		MediaType: mediatype,
+		Digest:    digest,
+		Size:      n,
+		message:   fmt.Sprintf("%s %s", status, digest),
 	}, nil
 }
 
@@ -85,21 +80,15 @@ func NewLayerFromLayer(digest, mediatype, from string) (*Layer, error) {
 		Digest:    digest,
 		Size:      fi.Size(),
 		From:      from,
+		message:   fmt.Sprintf("using existing layer %s", digest),
 	}, nil
 }
 
-func (l *Layer) Commit() (bool, error) {
-	// always remove temp
-	defer os.Remove(l.tempFileName)
-
+func (l *Layer) Open() (*os.File, error) {
 	blob, err := GetBlobsPath(l.Digest)
 	if err != nil {
-		return false, err
+		return nil, err
 	}
 
-	if _, err := os.Stat(blob); err != nil {
-		return true, os.Rename(l.tempFileName, blob)
-	}
-
-	return false, nil
+	return os.Open(blob)
 }

server/model.go

@@ -0,0 +1,259 @@
+package server
+
+import (
+	"archive/zip"
+	"bytes"
+	"context"
+	"errors"
+	"fmt"
+	"io"
+	"net/http"
+	"os"
+	"path/filepath"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/convert"
+	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/types/model"
+)
+
+type layerWithGGML struct {
+	*Layer
+	*llm.GGML
+}
+
+func parseFromModel(ctx context.Context, name model.Name, fn func(api.ProgressResponse)) (layers []*layerWithGGML, err error) {
+	modelpath := ParseModelPath(name.String())
+	manifest, _, err := GetManifest(modelpath)
+	switch {
+	case errors.Is(err, os.ErrNotExist):
+		if err := PullModel(ctx, name.String(), &registryOptions{}, fn); err != nil {
+			return nil, err
+		}
+
+		modelpath = ParseModelPath(name.String())
+		manifest, _, err = GetManifest(modelpath)
+		if err != nil {
+			return nil, err
+		}
+	case err != nil:
+		return nil, err
+	}
+
+	for _, layer := range manifest.Layers {
+		layer, err := NewLayerFromLayer(layer.Digest, layer.MediaType, modelpath.GetShortTagname())
+		if err != nil {
+			return nil, err
+		}
+
+		switch layer.MediaType {
+		case "application/vnd.ollama.image.model",
+			"application/vnd.ollama.image.projector",
+			"application/vnd.ollama.image.adapter":
+			blobpath, err := GetBlobsPath(layer.Digest)
+			if err != nil {
+				return nil, err
+			}
+
+			blob, err := os.Open(blobpath)
+			if err != nil {
+				return nil, err
+			}
+			defer blob.Close()
+
+			ggml, _, err := llm.DecodeGGML(blob)
+			if err != nil {
+				return nil, err
+			}
+
+			layers = append(layers, &layerWithGGML{layer, ggml})
+		default:
+			layers = append(layers, &layerWithGGML{layer, nil})
+		}
+
+	}
+
+	return layers, nil
+}
+
+func parseFromZipFile(_ context.Context, file *os.File, fn func(api.ProgressResponse)) (layers []*layerWithGGML, err error) {
+	stat, err := file.Stat()
+	if err != nil {
+		return nil, err
+	}
+
+	r, err := zip.NewReader(file, stat.Size())
+	if err != nil {
+		return nil, err
+	}
+
+	tempdir, err := os.MkdirTemp(filepath.Dir(file.Name()), "")
+	if err != nil {
+		return nil, err
+	}
+	defer os.RemoveAll(tempdir)
+
+	fn(api.ProgressResponse{Status: "unpacking model metadata"})
+	for _, f := range r.File {
+		// TODO(mxyng): this should not write out all files to disk
+		outfile, err := os.Create(filepath.Join(tempdir, f.Name))
+		if err != nil {
+			return nil, err
+		}
+
+		infile, err := f.Open()
+		if err != nil {
+			return nil, err
+		}
+
+		if _, err = io.Copy(outfile, infile); err != nil {
+			return nil, err
+		}
+
+		if err := outfile.Close(); err != nil {
+			return nil, err
+		}
+
+		if err := infile.Close(); err != nil {
+			return nil, err
+		}
+	}
+
+	mf, err := convert.GetModelFormat(tempdir)
+	if err != nil {
+		return nil, err
+	}
+
+	params, err := mf.GetParams(tempdir)
+	if err != nil {
+		return nil, err
+	}
+
+	mArch, err := mf.GetModelArch("", tempdir, params)
+	if err != nil {
+		return nil, err
+	}
+
+	fn(api.ProgressResponse{Status: "processing tensors"})
+	if err := mArch.GetTensors(); err != nil {
+		return nil, err
+	}
+
+	if err := mArch.LoadVocab(); err != nil {
+		return nil, err
+	}
+
+	fn(api.ProgressResponse{Status: "converting model"})
+
+	// TODO(mxyng): this should write directly into a layer
+	// e.g. NewLayer(arch.Reader(), "application/vnd.ollama.image.model")
+	temp, err := os.CreateTemp(tempdir, "fp16")
+	if err != nil {
+		return nil, err
+	}
+	defer temp.Close()
+	defer os.Remove(temp.Name())
+
+	if err = mArch.WriteGGUF(temp); err != nil {
+		return nil, err
+	}
+
+	if _, err := temp.Seek(0, io.SeekStart); err != nil {
+		return nil, err
+	}
+
+	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)
+	if err != nil {
+		return nil, err
+	}
+	defer bin.Close()
+
+	ggml, _, err := llm.DecodeGGML(bin)
+	if err != nil {
+		return nil, err
+	}
+
+	layer, err = NewLayerFromLayer(layer.Digest, layer.MediaType, "")
+	if err != nil {
+		return nil, err
+	}
+
+	layers = append(layers, &layerWithGGML{layer, ggml})
+	return layers, nil
+}
+
+func parseFromFile(ctx context.Context, file *os.File, fn func(api.ProgressResponse)) (layers []*layerWithGGML, err error) {
+	sr := io.NewSectionReader(file, 0, 512)
+	contentType, err := detectContentType(sr)
+	if err != nil {
+		return nil, err
+	}
+
+	switch contentType {
+	case "gguf", "ggla":
+		// noop
+	case "application/zip":
+		return parseFromZipFile(ctx, file, fn)
+	default:
+		return nil, fmt.Errorf("unsupported content type: %s", contentType)
+	}
+
+	stat, err := file.Stat()
+	if err != nil {
+		return nil, err
+	}
+
+	var offset int64
+	for offset < stat.Size() {
+		ggml, n, err := llm.DecodeGGML(file)
+		if errors.Is(err, io.EOF) {
+			break
+		} else if err != nil {
+			return nil, err
+		}
+
+		mediatype := "application/vnd.ollama.image.model"
+		if ggml.Name() == "ggla" {
+			mediatype = "application/vnd.ollama.image.adapter"
+		} else if ggml.KV().Architecture() == "clip" {
+			mediatype = "application/vnd.ollama.image.projector"
+		}
+
+		layer, err := NewLayer(io.NewSectionReader(file, offset, n), mediatype)
+		if err != nil {
+			return nil, err
+		}
+
+		layers = append(layers, &layerWithGGML{layer, ggml})
+		offset = n
+	}
+
+	return layers, nil
+}
+
+func detectContentType(r io.Reader) (string, error) {
+	var b bytes.Buffer
+	if _, err := io.Copy(&b, r); err != nil {
+		return "", err
+	}
+
+	if contentType := llm.DetectGGMLType(b.Bytes()); contentType != "" {
+		return contentType, nil
+	}
+
+	if contentType := http.DetectContentType(b.Bytes()); contentType != "application/octet-stream" {
+		return contentType, nil
+	}
+
+	return "unknown", nil
+}

server/routes.go

@@ -580,7 +580,7 @@ func (s *Server) CreateModelHandler(c *gin.Context) {
 		ctx, cancel := context.WithCancel(c.Request.Context())
 		defer cancel()
 
-		if err := CreateModel(ctx, model, filepath.Dir(req.Path), req.Quantization, commands, fn); err != nil {
+		if err := CreateModel(ctx, model, filepath.Dir(req.Path), strings.ToUpper(req.Quantization), commands, fn); err != nil {
 			ch <- gin.H{"error": err.Error()}
 		}
 	}()
@@ -728,12 +728,12 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 		}
 	}
 
-	mf, err := ShowModelfile(model)
-	if err != nil {
-		return nil, err
-	}
-
-	resp.Modelfile = mf
+	var sb strings.Builder
+	fmt.Fprintln(&sb, "# Modelfile generate 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, parser.Format(model.Commands()))
+	resp.Modelfile = sb.String()
 
 	return resp, nil
 }
@@ -810,16 +810,13 @@ func (s *Server) CopyModelHandler(c *gin.Context) {
 
 	src := model.ParseName(r.Source)
 	if !src.IsValid() {
-		_ = c.Error(fmt.Errorf("source %q is invalid", r.Source))
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("source %q is invalid", r.Source)})
+		return
 	}
 
 	dst := model.ParseName(r.Destination)
 	if !dst.IsValid() {
-		_ = c.Error(fmt.Errorf("destination %q is invalid", r.Destination))
-	}
-
-	if len(c.Errors) > 0 {
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": c.Errors.Errors()})
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("destination %q is invalid", r.Source)})
 		return
 	}
 
@@ -875,11 +872,6 @@ func (s *Server) CreateBlobHandler(c *gin.Context) {
 		return
 	}
 
-	if _, err := layer.Commit(); err != nil {
-		c.AbortWithStatusJSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-		return
-	}
-
 	c.Status(http.StatusCreated)
 }
 

server/routes_test.go

@@ -124,14 +124,12 @@ func Test_Routes(t *testing.T) {
 			Method: http.MethodPost,
 			Path:   "/api/create",
 			Setup: func(t *testing.T, req *http.Request) {
-				f, err := os.CreateTemp(t.TempDir(), "ollama-model")
-				assert.Nil(t, err)
-				defer f.Close()
+				fname := createTestFile(t, "ollama-model")
 
 				stream := false
 				createReq := api.CreateRequest{
 					Name:      "t-bone",
-					Modelfile: fmt.Sprintf("FROM %s", f.Name()),
+					Modelfile: fmt.Sprintf("FROM %s", fname),
 					Stream:    &stream,
 				}
 				jsonData, err := json.Marshal(createReq)
@@ -216,28 +214,25 @@ func Test_Routes(t *testing.T) {
 	httpSrv := httptest.NewServer(router)
 	t.Cleanup(httpSrv.Close)
 
-	workDir, err := os.MkdirTemp("", "ollama-test")
-	assert.Nil(t, err)
-	defer os.RemoveAll(workDir)
-	os.Setenv("OLLAMA_MODELS", workDir)
+	t.Setenv("OLLAMA_MODELS", t.TempDir())
 
 	for _, tc := range testCases {
-		t.Logf("Running Test: [%s]", tc.Name)
-		u := httpSrv.URL + tc.Path
-		req, err := http.NewRequestWithContext(context.TODO(), tc.Method, u, nil)
-		assert.Nil(t, err)
+		t.Run(tc.Name, func(t *testing.T) {
+			u := httpSrv.URL + tc.Path
+			req, err := http.NewRequestWithContext(context.TODO(), tc.Method, u, nil)
+			assert.Nil(t, err)
 
-		if tc.Setup != nil {
-			tc.Setup(t, req)
-		}
+			if tc.Setup != nil {
+				tc.Setup(t, req)
+			}
 
-		resp, err := httpSrv.Client().Do(req)
-		assert.Nil(t, err)
-		defer resp.Body.Close()
-
-		if tc.Expected != nil {
-			tc.Expected(t, resp)
-		}
+			resp, err := httpSrv.Client().Do(req)
+			assert.Nil(t, err)
+			defer resp.Body.Close()
 
+			if tc.Expected != nil {
+				tc.Expected(t, resp)
+			}
+		})
 	}
 }

server/sched.go

@@ -250,6 +250,7 @@ func (s *Scheduler) processCompleted(ctx context.Context) {
 						defer runner.refMu.Unlock()
 						if runner.expireTimer != nil {
 							runner.expireTimer.Stop()
+							runner.expireTimer = nil
 						}
 						s.expiredCh <- runner
 					})
@@ -296,6 +297,10 @@ func (pending *LlmRequest) useLoadedRunner(runner *runnerRef, finished chan *Llm
 	runner.refMu.Lock()
 	defer runner.refMu.Unlock()
 	runner.refCount++
+	if runner.expireTimer != nil {
+		runner.expireTimer.Stop()
+		runner.expireTimer = nil
+	}
 	runner.sessionDuration = pending.sessionDuration
 	pending.successCh <- runner
 	go func() {
@@ -426,6 +431,10 @@ type runnerRef struct {
 
 // The refMu must already be held when calling unload
 func (runner *runnerRef) unload() {
+	if runner.expireTimer != nil {
+		runner.expireTimer.Stop()
+		runner.expireTimer = nil
+	}
 	if runner.llama != nil {
 		runner.llama.Close()
 	}

types/model/name.go

@@ -143,18 +143,28 @@ func ParseNameBare(s string) Name {
 		n.RawDigest = MissingPart
 	}
 
-	s, n.Tag, _ = cutPromised(s, ":")
+	// "/" is an illegal tag character, so we can use it to split the host
+	if strings.LastIndex(s, ":") > strings.LastIndex(s, "/") {
+		s, n.Tag, _ = cutPromised(s, ":")
+	}
+
 	s, n.Model, promised = cutPromised(s, "/")
 	if !promised {
 		n.Model = s
 		return n
 	}
+
 	s, n.Namespace, promised = cutPromised(s, "/")
 	if !promised {
 		n.Namespace = s
 		return n
 	}
-	n.Host = s
+
+	scheme, host, ok := strings.Cut(s, "://")
+	if ! ok {
+		host = scheme
+	}
+	n.Host = host
 
 	return n
 }

types/model/name_test.go

@@ -1,6 +1,7 @@
 package model
 
 import (
+	"path/filepath"
 	"reflect"
 	"runtime"
 	"testing"
@@ -15,8 +16,19 @@ func TestParseNameParts(t *testing.T) {
 	cases := []struct {
 		in              string
 		want            Name
+		wantFilepath    string
 		wantValidDigest bool
 	}{
+		{
+			in: "scheme://host:port/namespace/model:tag",
+			want: Name{
+				Host:      "host:port",
+				Namespace: "namespace",
+				Model:     "model",
+				Tag:       "tag",
+			},
+			wantFilepath: filepath.Join("host:port", "namespace", "model", "tag"),
+		},
 		{
 			in: "host/namespace/model:tag",
 			want: Name{
@@ -25,6 +37,17 @@ func TestParseNameParts(t *testing.T) {
 				Model:     "model",
 				Tag:       "tag",
 			},
+			wantFilepath: filepath.Join("host", "namespace", "model", "tag"),
+		},
+		{
+			in: "host:port/namespace/model:tag",
+			want: Name{
+				Host:      "host:port",
+				Namespace: "namespace",
+				Model:     "model",
+				Tag:       "tag",
+			},
+			wantFilepath: filepath.Join("host:port", "namespace", "model", "tag"),
 		},
 		{
 			in: "host/namespace/model",
@@ -33,6 +56,16 @@ func TestParseNameParts(t *testing.T) {
 				Namespace: "namespace",
 				Model:     "model",
 			},
+			wantFilepath: filepath.Join("host", "namespace", "model", "latest"),
+		},
+		{
+			in: "host:port/namespace/model",
+			want: Name{
+				Host:      "host:port",
+				Namespace: "namespace",
+				Model:     "model",
+			},
+			wantFilepath: filepath.Join("host:port", "namespace", "model", "latest"),
 		},
 		{
 			in: "namespace/model",
@@ -40,12 +73,14 @@ func TestParseNameParts(t *testing.T) {
 				Namespace: "namespace",
 				Model:     "model",
 			},
+			wantFilepath: filepath.Join("registry.ollama.ai", "namespace", "model", "latest"),
 		},
 		{
 			in: "model",
 			want: Name{
 				Model: "model",
 			},
+			wantFilepath: filepath.Join("registry.ollama.ai", "library", "model", "latest"),
 		},
 		{
 			in: "h/nn/mm:t",
@@ -55,6 +90,7 @@ func TestParseNameParts(t *testing.T) {
 				Model:     "mm",
 				Tag:       "t",
 			},
+			wantFilepath: filepath.Join("h", "nn", "mm", "t"),
 		},
 		{
 			in: part80 + "/" + part80 + "/" + part80 + ":" + part80,
@@ -64,6 +100,7 @@ func TestParseNameParts(t *testing.T) {
 				Model:     part80,
 				Tag:       part80,
 			},
+			wantFilepath: filepath.Join(part80, part80, part80, part80),
 		},
 		{
 			in: part350 + "/" + part80 + "/" + part80 + ":" + part80,
@@ -73,6 +110,7 @@ func TestParseNameParts(t *testing.T) {
 				Model:     part80,
 				Tag:       part80,
 			},
+			wantFilepath: filepath.Join(part350, part80, part80, part80),
 		},
 		{
 			in: "@digest",
@@ -97,6 +135,11 @@ func TestParseNameParts(t *testing.T) {
 			if !reflect.DeepEqual(got, tt.want) {
 				t.Errorf("parseName(%q) = %v; want %v", tt.in, got, tt.want)
 			}
+
+			got = ParseName(tt.in)
+			if tt.wantFilepath != "" && got.Filepath() != tt.wantFilepath {
+				t.Errorf("parseName(%q).Filepath() = %q; want %q", tt.in, got.Filepath(), tt.wantFilepath)
+			}
 		})
 	}
 }