fix quantization

---------

Co-authored-by: Richard Lyons <frob@cloudstaff.com>

.github/workflows/release.yaml

@@ -23,7 +23,7 @@ jobs:
           echo GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=${GITHUB_REF_NAME#v}\" \"-X=github.com/ollama/ollama/server.mode=release\"'" >>$GITHUB_OUTPUT
 
   darwin-build:
-    runs-on: macos-13
+    runs-on: macos-13-xlarge
     environment: release
     needs: setup-environment
     strategy:

README.md

@@ -360,7 +360,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Tkinter-based client](https://github.com/chyok/ollama-gui) (Python tkinter-based Client for Ollama)
 - [LLMChat](https://github.com/trendy-design/llmchat) (Privacy focused, 100% local, intuitive all-in-one chat interface)
 - [Local Multimodal AI Chat](https://github.com/Leon-Sander/Local-Multimodal-AI-Chat) (Ollama-based LLM Chat with support for multiple features, including PDF RAG, voice chat, image-based interactions, and integration with OpenAI.)
-- [ARGO](https://github.com/xark-argo/argo) (Locally download and run Ollama and Huggingface models with RAG on Mac/Windows/Linux)
+- [ARGO](https://github.com/xark-argo/argo) (Locally download and run Ollama and Huggingface models with RAG and deep research on Mac/Windows/Linux)
 - [OrionChat](https://github.com/EliasPereirah/OrionChat) - OrionChat is a web interface for chatting with different AI providers
 - [G1](https://github.com/bklieger-groq/g1) (Prototype of using prompting strategies to improve the LLM's reasoning through o1-like reasoning chains.)
 - [Web management](https://github.com/lemonit-eric-mao/ollama-web-management) (Web management page)
@@ -595,10 +595,11 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [SimpleOllamaUnity](https://github.com/HardCodeDev777/SimpleOllamaUnity) (Unity Engine extension for communicating with Ollama in a few lines of code. Also works at runtime)
 - [UnityCodeLama](https://github.com/HardCodeDev777/UnityCodeLama) (Unity Edtior tool to analyze scripts via Ollama)
 - [NativeMind](https://github.com/NativeMindBrowser/NativeMindExtension) (Private, on-device AI Assistant, no cloud dependencies)
+- [GMAI - Gradle Managed AI](https://gmai.premex.se/) (Gradle plugin for automated Ollama lifecycle management during build phases)
 
 ### Supported backends
 
-- [llama.cpp](https://github.com/ggerganov/llama.cpp) project founded by Georgi Gerganov.
+- [llama.cpp](https://github.com/ggml-org/llama.cpp) project founded by Georgi Gerganov.
 
 ### Observability
 - [Opik](https://www.comet.com/docs/opik/cookbook/ollama) is an open-source platform to debug, evaluate, and monitor your LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards. Opik supports native intergration to Ollama.

api/client.go

@@ -222,10 +222,6 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 			return fmt.Errorf("unmarshal: %w", err)
 		}
 
-		if errorResponse.Error != "" {
-			return errors.New(errorResponse.Error)
-		}
-
 		if response.StatusCode >= http.StatusBadRequest {
 			return StatusError{
 				StatusCode:   response.StatusCode,
@@ -234,6 +230,10 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 			}
 		}
 
+		if errorResponse.Error != "" {
+			return errors.New(errorResponse.Error)
+		}
+
 		if err := fn(bts); err != nil {
 			return err
 		}

api/client_test.go

@@ -89,6 +89,16 @@ func TestClientStream(t *testing.T) {
 			},
 			wantErr: "mid-stream error",
 		},
+		{
+			name: "http status error takes precedence over general error",
+			responses: []any{
+				testError{
+					message:    "custom error message",
+					statusCode: http.StatusInternalServerError,
+				},
+			},
+			wantErr: "500",
+		},
 		{
 			name: "successful stream completion",
 			responses: []any{

api/types.go

@@ -468,13 +468,14 @@ type ListModelResponse struct {
 
 // ProcessModelResponse is a single model description in [ProcessResponse].
 type ProcessModelResponse struct {
-	Name      string       `json:"name"`
-	Model     string       `json:"model"`
-	Size      int64        `json:"size"`
-	Digest    string       `json:"digest"`
-	Details   ModelDetails `json:"details,omitempty"`
-	ExpiresAt time.Time    `json:"expires_at"`
-	SizeVRAM  int64        `json:"size_vram"`
+	Name          string       `json:"name"`
+	Model         string       `json:"model"`
+	Size          int64        `json:"size"`
+	Digest        string       `json:"digest"`
+	Details       ModelDetails `json:"details,omitempty"`
+	ExpiresAt     time.Time    `json:"expires_at"`
+	SizeVRAM      int64        `json:"size_vram"`
+	ContextLength int          `json:"context_length"`
 }
 
 type TokenResponse struct {

cmd/cmd.go

@@ -583,12 +583,13 @@ func ListRunningHandler(cmd *cobra.Command, args []string) error {
 			} else {
 				until = format.HumanTime(m.ExpiresAt, "Never")
 			}
-			data = append(data, []string{m.Name, m.Digest[:12], format.HumanBytes(m.Size), procStr, until})
+			ctxStr := strconv.Itoa(m.ContextLength)
+			data = append(data, []string{m.Name, m.Digest[:12], format.HumanBytes(m.Size), procStr, ctxStr, until})
 		}
 	}
 
 	table := tablewriter.NewWriter(os.Stdout)
-	table.SetHeader([]string{"NAME", "ID", "SIZE", "PROCESSOR", "UNTIL"})
+	table.SetHeader([]string{"NAME", "ID", "SIZE", "PROCESSOR", "CONTEXT", "UNTIL"})
 	table.SetHeaderAlignment(tablewriter.ALIGN_LEFT)
 	table.SetAlignment(tablewriter.ALIGN_LEFT)
 	table.SetHeaderLine(false)
@@ -1079,10 +1080,11 @@ func chat(cmd *cobra.Command, opts runOptions) (*api.Message, error) {
 	var state *displayResponseState = &displayResponseState{}
 	var latest api.ChatResponse
 	var fullResponse strings.Builder
-	var role string
 	var thinkTagOpened bool = false
 	var thinkTagClosed bool = false
 
+	role := "assistant"
+
 	fn := func(response api.ChatResponse) error {
 		if response.Message.Content != "" || !opts.HideThinking {
 			p.StopAndClear()
@@ -1416,13 +1418,13 @@ func NewCLI() *cobra.Command {
 
 	createCmd := &cobra.Command{
 		Use:     "create MODEL",
-		Short:   "Create a model from a Modelfile",
+		Short:   "Create a model",
 		Args:    cobra.ExactArgs(1),
 		PreRunE: checkServerHeartbeat,
 		RunE:    CreateHandler,
 	}
 
-	createCmd.Flags().StringP("file", "f", "", "Name of the Modelfile (default \"Modelfile\"")
+	createCmd.Flags().StringP("file", "f", "", "Name of the Modelfile (default \"Modelfile\")")
 	createCmd.Flags().StringP("quantize", "q", "", "Quantize model to this level (e.g. q4_K_M)")
 
 	showCmd := &cobra.Command{

cmd/interactive.go

@@ -385,18 +385,21 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 				case "modelfile":
 					fmt.Println(resp.Modelfile)
 				case "parameters":
+					fmt.Println("Model defined parameters:")
 					if resp.Parameters == "" {
-						fmt.Println("No parameters were specified for this model.")
+						fmt.Println("  No additional parameters were specified for this model.")
 					} else {
-						if len(opts.Options) > 0 {
-							fmt.Println("User defined parameters:")
-							for k, v := range opts.Options {
-								fmt.Printf("%-*s %v\n", 30, k, v)
-							}
-							fmt.Println()
+						for _, l := range strings.Split(resp.Parameters, "\n") {
+							fmt.Printf("  %s\n", l)
 						}
-						fmt.Println("Model defined parameters:")
-						fmt.Println(resp.Parameters)
+					}
+					fmt.Println()
+					if len(opts.Options) > 0 {
+						fmt.Println("User defined parameters:")
+						for k, v := range opts.Options {
+							fmt.Printf("  %-*s %v\n", 30, k, v)
+						}
+						fmt.Println()
 					}
 				case "system":
 					switch {

docs/README.md

@@ -4,6 +4,7 @@
 * [Quickstart](../README.md#quickstart)
 * [Examples](./examples.md)
 * [Importing models](./import.md)
+* [MacOS Documentation](./macos.md)
 * [Linux Documentation](./linux.md)
 * [Windows Documentation](./windows.md)
 * [Docker Documentation](./docker.md)

docs/faq.md

@@ -292,7 +292,7 @@ If too many requests are sent to the server, it will respond with a 503 error in
 
 ## How does Ollama handle concurrent requests?
 
-Ollama supports two levels of concurrent processing.  If your system has sufficient available memory (system memory when using CPU inference, or VRAM for GPU inference) then multiple models can be loaded at the same time.  For a given model, if there is sufficient available memory when the model is loaded, it is configured to allow parallel request processing.
+Ollama supports two levels of concurrent processing.  If your system has sufficient available memory (system memory when using CPU inference, or VRAM for GPU inference) then multiple models can be loaded at the same time.  For a given model, if there is sufficient available memory when the model is loaded, it can be configured to allow parallel request processing.
 
 If there is insufficient available memory to load a new model request while one or more models are already loaded, all new requests will be queued until the new model can be loaded.  As prior models become idle, one or more will be unloaded to make room for the new model.  Queued requests will be processed in order.  When using GPU inference new models must be able to completely fit in VRAM to allow concurrent model loads.
 
@@ -301,7 +301,7 @@ Parallel request processing for a given model results in increasing the context
 The following server settings may be used to adjust how Ollama handles concurrent requests on most platforms:
 
 - `OLLAMA_MAX_LOADED_MODELS` - The maximum number of models that can be loaded concurrently provided they fit in available memory.  The default is 3 * the number of GPUs or 3 for CPU inference.
-- `OLLAMA_NUM_PARALLEL` - The maximum number of parallel requests each model will process at the same time.  The default will auto-select either 4 or 1 based on available memory.
+- `OLLAMA_NUM_PARALLEL` - The maximum number of parallel requests each model will process at the same time.  The default is 1, and will handle 1 request per model at a time.
 - `OLLAMA_MAX_QUEUE` - The maximum number of requests Ollama will queue when busy before rejecting additional requests. The default is 512
 
 Note: Windows with Radeon GPUs currently default to 1 model maximum due to limitations in ROCm v5.7 for available VRAM reporting.  Once ROCm v6.2 is available, Windows Radeon will follow the defaults above.  You may enable concurrent model loads on Radeon on Windows, but ensure you don't load more models than will fit into your GPUs VRAM.
@@ -333,3 +333,16 @@ The currently available K/V cache quantization types are:
 How much the cache quantization impacts the model's response quality will depend on the model and the task.  Models that have a high GQA count (e.g. Qwen2) may see a larger impact on precision from quantization than models with a low GQA count.
 
 You may need to experiment with different quantization types to find the best balance between memory usage and quality.
+
+## How can I stop Ollama from starting when I login to my computer
+
+Ollama for Windows and macOS register as a login item during installation.  You can disable this if you prefer not to have Ollama automatically start.  Ollama will respect this setting across upgrades, unless you uninstall the application.
+
+**Windows**
+- Remove `%APPDATA%\Microsoft\Windows\Start Menu\Programs\Startup\Ollama.lnk`
+
+**MacOS Monterey (v12)**
+- Open `Settings` -> `Users & Groups` -> `Login Items` and find the `Ollama` entry, then click the `-` (minus) to remove
+
+**MacOS Ventura (v13) and later**
+- Open `Settings` and search for "Login Items", find the `Ollama` entry under "Allow in the Background`, then click the slider to disable.

docs/import.md

@@ -53,6 +53,8 @@ FROM /path/to/safetensors/directory
 
 If you create the Modelfile in the same directory as the weights, you can use the command `FROM .`.
 
+If you do not create the Modelfile, ollama will act as if there was a Modelfile with the command `FROM .`.
+
 Now run the `ollama create` command from the directory where you created the `Modelfile`:
 
 ```shell

docs/linux.md

@@ -16,7 +16,7 @@ curl -fsSL https://ollama.com/install.sh | sh
 Download and extract the package:
 
 ```shell
-curl -L https://ollama.com/download/ollama-linux-amd64.tgz -o ollama-linux-amd64.tgz
+curl -LO https://ollama.com/download/ollama-linux-amd64.tgz
 sudo tar -C /usr -xzf ollama-linux-amd64.tgz
 ```
 

docs/macos.md

@@ -0,0 +1,42 @@
+# Ollama for macOS
+
+## System Requirements
+
+* MacOS Monterey (v12) or newer
+* Apple M series (CPU and GPU support) or x86 (CPU only)
+
+
+## Filesystem Requirements
+
+The preferred method of installation is to mount the `ollama.dmg` and drag-and-drop the Ollama application to the system-wide `Applications` folder.  Upon startup, the Ollama app will verify the `ollama` CLI is present in your PATH, and if not detected, will prompt for permission to create a link in `/usr/local/bin`
+
+Once you've installed Ollama, you'll need additional space for storing the Large Language models, which can be tens to hundreds of GB in size.  If your home directory doesn't have enough space, you can change where the binaries are installed, and where the models are stored.
+
+### Changing Install Location
+
+To install the Ollama application somewhere other than `Applications`, place the Ollama application in the desired location, and ensure the CLI `Ollama.app/Contents/Resources/ollama` or a sym-link to the CLI can be found in your path.  Upon first start decline the "Move to Applications?" request.
+
+
+## Troubleshooting
+
+Ollama on MacOS stores files in a few different locations.
+- `~/.ollama` contains models and configuration
+- `~/.ollama/logs` contains logs
+    - *app.log* contains most recent logs from the GUI application
+    - *server.log* contains the most recent server logs
+- `<install location>/Ollama.app/Contents/Resources/ollama` the CLI binary
+
+## Uninstall
+
+To fully remove Ollama from your system, remove the following files and folders:
+
+```
+sudo rm -rf /Applications/Ollama.app
+sudo rm /usr/local/bin/ollama
+rm -rf "~/Library/Application Support/Ollama"
+rm -rf "~/Library/Saved Application State/com.electron.ollama.savedState"
+rm -rf ~/Library/Caches/com.electron.ollama/
+rm -rf ~/Library/Caches/ollama
+rm -rf ~/Library/WebKit/com.electron.ollama
+rm -rf ~/.ollama
+```

docs/modelfile.md

@@ -150,7 +150,7 @@ PARAMETER <parameter> <parametervalue>
 
 | Parameter      | Description                                                                                                                                                                                                                                             | Value Type | Example Usage        |
 | -------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------- | -------------------- |
-| num_ctx        | Sets the size of the context window used to generate the next token. (Default: 2048)                                                                                                                                                                    | int        | num_ctx 4096         |
+| num_ctx        | Sets the size of the context window used to generate the next token. (Default: 4096)                                                                                                                                                                    | int        | num_ctx 4096         |
 | repeat_last_n  | Sets how far back for the model to look back to prevent repetition. (Default: 64, 0 = disabled, -1 = num_ctx)                                                                                                                                           | int        | repeat_last_n 64     |
 | repeat_penalty | Sets how strongly to penalize repetitions. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. (Default: 1.1)                                                                     | float      | repeat_penalty 1.1   |
 | temperature    | The temperature of the model. Increasing the temperature will make the model answer more creatively. (Default: 0.8)                                                                                                                                     | float      | temperature 0.7      |

docs/windows.md

@@ -30,20 +30,6 @@ To install the Ollama application in a location different than your home directo
 OllamaSetup.exe /DIR="d:\some\location"
 ```
 
-### Changing Model Location
-
-To change where Ollama stores the downloaded models instead of using your home directory, set the environment variable `OLLAMA_MODELS` in your user account.
-
-1. Start the Settings (Windows 11) or Control Panel (Windows 10) application and search for _environment variables_.
-
-2. Click on _Edit environment variables for your account_.
-
-3. Edit or create a new variable for your user account for `OLLAMA_MODELS` where you want the models stored
-
-4. Click OK/Apply to save.
-
-If Ollama is already running, Quit the tray application and relaunch it from the Start menu, or a new terminal started after you saved the environment variables.
-
 ## API Access
 
 Here's a quick example showing API access from `powershell`

envconfig/config.go

@@ -219,7 +219,7 @@ func Uint(key string, defaultValue uint) func() uint {
 
 var (
 	// NumParallel sets the number of parallel model requests. NumParallel can be configured via the OLLAMA_NUM_PARALLEL environment variable.
-	NumParallel = Uint("OLLAMA_NUM_PARALLEL", 0)
+	NumParallel = Uint("OLLAMA_NUM_PARALLEL", 1)
 	// MaxRunners sets the maximum number of loaded models. MaxRunners can be configured via the OLLAMA_MAX_LOADED_MODELS environment variable.
 	MaxRunners = Uint("OLLAMA_MAX_LOADED_MODELS", 0)
 	// MaxQueue sets the maximum number of queued requests. MaxQueue can be configured via the OLLAMA_MAX_QUEUE environment variable.

integration/library_models_test.go

@@ -0,0 +1,57 @@
+//go:build integration && library
+
+package integration
+
+import (
+	"context"
+	"log/slog"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+// First run of this scenario on a target system will take a long time to download
+// ~1.5TB of models.  Set a sufficiently large -timeout for your network speed
+func TestLibraryModelsGenerate(t *testing.T) {
+	softTimeout, hardTimeout := getTimeouts(t)
+	slog.Info("Setting timeouts", "soft", softTimeout, "hard", hardTimeout)
+	ctx, cancel := context.WithTimeout(context.Background(), hardTimeout)
+	defer cancel()
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+
+	chatModels := libraryChatModels
+	for _, model := range chatModels {
+		t.Run(model, func(t *testing.T) {
+			if time.Now().Sub(started) > softTimeout {
+				t.Skip("skipping remaining tests to avoid excessive runtime")
+			}
+			if err := PullIfMissing(ctx, client, model); err != nil {
+				t.Fatalf("pull failed %s", err)
+			}
+			req := api.GenerateRequest{
+				Model:     model,
+				Prompt:    "why is the sky blue?",
+				KeepAlive: &api.Duration{Duration: 10 * time.Second},
+				Options: map[string]interface{}{
+					"temperature": 0.1,
+					"seed":        123,
+				},
+			}
+			anyResp := []string{"rayleigh", "scatter", "atmosphere", "nitrogen", "oxygen", "wavelength"}
+			// Special cases
+			if model == "duckdb-nsql" {
+				anyResp = []string{"select", "from"}
+			} else if model == "granite3-guardian" || model == "shieldgemma" || model == "llama-guard3" || model == "bespoke-minicheck" {
+				anyResp = []string{"yes", "no", "safe", "unsafe"}
+			} else if model == "openthinker" || model == "nexusraven" {
+				anyResp = []string{"plugin", "im_sep", "components", "function call"}
+			} else if model == "starcoder" || model == "starcoder2" || model == "magicoder" || model == "deepseek-coder" {
+				req.Prompt = "def fibonacci():"
+				anyResp = []string{"f(n)", "sequence", "n-1", "main()", "__main__", "while"}
+			}
+			DoGenerate(ctx, t, client, req, anyResp, 120*time.Second, 30*time.Second)
+		})
+	}
+}

integration/utils_test.go

@@ -72,6 +72,187 @@ var (
 		"stablelm2:latest", // Predictions are off, crashes on small VRAM GPUs
 		"falcon:latest",
 	}
+
+	// Some library models are quite large - ensure large VRAM and sufficient disk space
+	// before running scenarios based on this set
+	libraryChatModels = []string{
+		"alfred",
+		"athene-v2",
+		"aya-expanse",
+		"aya",
+		"bakllava",
+		"bespoke-minicheck",
+		"codebooga",
+		"codegeex4",
+		"codegemma",
+		"codellama",
+		"codeqwen",
+		"codestral",
+		"codeup",
+		"cogito",
+		"command-a",
+		"command-r-plus",
+		"command-r",
+		"command-r7b-arabic",
+		"command-r7b",
+		"dbrx",
+		"deepcoder",
+		"deepscaler",
+		"deepseek-coder-v2",
+		"deepseek-coder",
+		"deepseek-llm",
+		"deepseek-r1",
+		// "deepseek-v2.5", // requires 155 GB VRAM
+		"deepseek-v2",
+		// "deepseek-v3", // requires 482 GB VRAM
+		"devstral",
+		"dolphin-llama3",
+		"dolphin-mistral",
+		"dolphin-mixtral",
+		"dolphin-phi",
+		"dolphin3",
+		"dolphincoder",
+		"duckdb-nsql",
+		"everythinglm",
+		"exaone-deep",
+		"exaone3.5",
+		"falcon",
+		"falcon2",
+		"falcon3",
+		"firefunction-v2",
+		"gemma",
+		"gemma2",
+		"gemma3",
+		"gemma3n",
+		"glm4",
+		"goliath",
+		"granite-code",
+		"granite3-dense",
+		"granite3-guardian",
+		"granite3-moe",
+		"granite3.1-dense",
+		"granite3.1-moe",
+		"granite3.2-vision",
+		"granite3.2",
+		"granite3.3",
+		"hermes3",
+		"internlm2",
+		"llama-guard3",
+		"llama-pro",
+		"llama2-chinese",
+		"llama2-uncensored",
+		"llama2",
+		"llama3-chatqa",
+		"llama3-gradient",
+		"llama3-groq-tool-use",
+		"llama3.1",
+		"llama3.2-vision",
+		"llama3.2",
+		"llama3.3",
+		"llama3",
+		"llama4",
+		"llava-llama3",
+		"llava-phi3",
+		"llava",
+		"magicoder",
+		"magistral",
+		"marco-o1",
+		"mathstral",
+		"meditron",
+		"medllama2",
+		"megadolphin",
+		"minicpm-v",
+		"mistral-large",
+		"mistral-nemo",
+		"mistral-openorca",
+		"mistral-small",
+		"mistral-small3.1",
+		"mistral-small3.2",
+		"mistral",
+		"mistrallite",
+		"mixtral",
+		"moondream",
+		"nemotron-mini",
+		"nemotron",
+		"neural-chat",
+		"nexusraven",
+		"notus",
+		"nous-hermes",
+		"nous-hermes2-mixtral",
+		"nous-hermes2",
+		"nuextract",
+		"olmo2",
+		"open-orca-platypus2",
+		"openchat",
+		"opencoder",
+		"openhermes",
+		"openthinker",
+		"orca-mini",
+		"orca2",
+		// "phi", // unreliable
+		"phi3.5",
+		"phi3",
+		"phi4-mini-reasoning",
+		"phi4-mini",
+		"phi4-reasoning",
+		"phi4",
+		"phind-codellama",
+		"qwen",
+		"qwen2-math",
+		"qwen2.5-coder",
+		"qwen2.5",
+		"qwen2.5vl",
+		"qwen2",
+		"qwen3:0.6b", // dense
+		"qwen3:30b",  // MOE
+		"qwq",
+		"r1-1776",
+		"reader-lm",
+		"reflection",
+		"sailor2",
+		"samantha-mistral",
+		"shieldgemma",
+		"smallthinker",
+		"smollm",
+		"smollm2",
+		"solar-pro",
+		"solar",
+		"sqlcoder",
+		"stable-beluga",
+		"stable-code",
+		"stablelm-zephyr",
+		"stablelm2",
+		"starcoder",
+		"starcoder2",
+		"starling-lm",
+		"tinydolphin",
+		"tinyllama",
+		"tulu3",
+		"vicuna",
+		"wizard-math",
+		"wizard-vicuna-uncensored",
+		"wizard-vicuna",
+		"wizardcoder",
+		"wizardlm-uncensored",
+		"wizardlm2",
+		"xwinlm",
+		"yarn-llama2",
+		"yarn-mistral",
+		"yi-coder",
+		"yi",
+		"zephyr",
+	}
+	libraryEmbedModels = []string{
+		"all-minilm",
+		"bge-large",
+		"bge-m3",
+		"granite-embedding",
+		"mxbai-embed-large",
+		"nomic-embed-text",
+		"paraphrase-multilingual",
+		"snowflake-arctic-embed",
+		"snowflake-arctic-embed2",
+	}
 )
 
 func Init() {
@@ -313,6 +494,10 @@ func DoGenerate(ctx context.Context, t *testing.T, client *api.Client, genReq ap
 			t.Errorf("generate stalled.  Response so far:%s", buf.String())
 		}
 	case <-done:
+		if genErr != nil && strings.Contains(genErr.Error(), "model requires more system memory") {
+			slog.Warn("model is too large for the target test system", "model", genReq.Model, "error", genErr)
+			return
+		}
 		require.NoError(t, genErr, "failed with %s request prompt %s ", genReq.Model, genReq.Prompt)
 		// Verify the response contains the expected data
 		response := buf.String()

llama/patches/0017-ggml-Export-GPU-UUIDs.patch

@@ -7,31 +7,31 @@ This enables matching up devices and information reported by the backend
 with tools (e.g. nvidia-smi) and system management libraries (e.g. nvml).
 ---
  ggml/include/ggml-backend.h      |  1 +
- ggml/src/ggml-cuda/ggml-cuda.cu  | 33 ++++++++++++++++++++++++++++++++
+ ggml/src/ggml-cuda/ggml-cuda.cu  | 39 ++++++++++++++++++++++++++++++++
  ggml/src/ggml-metal/ggml-metal.m |  1 +
- 3 files changed, 35 insertions(+)
+ 3 files changed, 41 insertions(+)
 
 diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h
-index 74e46716..a880df33 100644
+index 74e46716..48839339 100644
 --- a/ggml/include/ggml-backend.h
 +++ b/ggml/include/ggml-backend.h
 @@ -152,6 +152,7 @@ extern "C" {
      struct ggml_backend_dev_props {
          const char * name;
          const char * description;
-+        const char * uuid;
++        const char * id;
          size_t memory_free;
          size_t memory_total;
          enum ggml_backend_dev_type type;
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index cb0d8528..4c829153 100644
+index cb0d8528..d6960174 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
 @@ -2884,6 +2884,7 @@ struct ggml_backend_cuda_device_context {
      int device;
      std::string name;
      std::string description;
-+    std::string uuid;
++    std::string id;
  };
  
  static const char * ggml_backend_cuda_device_get_name(ggml_backend_dev_t dev) {
@@ -39,9 +39,9 @@ index cb0d8528..4c829153 100644
      return ctx->description.c_str();
  }
  
-+static const char * ggml_backend_cuda_device_get_uuid(ggml_backend_dev_t dev) {
++static const char * ggml_backend_cuda_device_get_id(ggml_backend_dev_t dev) {
 +    ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
-+    return ctx->uuid.c_str();
++    return ctx->id.c_str();
 +}
 +
  static void ggml_backend_cuda_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
@@ -51,17 +51,17 @@ index cb0d8528..4c829153 100644
  static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
      props->name        = ggml_backend_cuda_device_get_name(dev);
      props->description = ggml_backend_cuda_device_get_description(dev);
-+    props->uuid        = ggml_backend_cuda_device_get_uuid(dev);
++    props->id          = ggml_backend_cuda_device_get_id(dev);
      props->type        = ggml_backend_cuda_device_get_type(dev);
      ggml_backend_cuda_device_get_memory(dev, &props->memory_free, &props->memory_total);
  
-@@ -3458,6 +3465,32 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
+@@ -3458,6 +3465,38 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
                  CUDA_CHECK(cudaGetDeviceProperties(&prop, i));
                  dev_ctx->description = prop.name;
  
 +                #if !defined(GGML_USE_HIP)
-+                char uuid[64];
-+                snprintf(uuid, sizeof(uuid),
++                char id[64];
++                snprintf(id, sizeof(id),
 +                    "GPU-%02x%02x%02x%02x-%02x%02x-%02x%02x-%02x%02x-%02x%02x%02x%02x%02x%02x",
 +                    (unsigned char)prop.uuid.bytes[0],
 +                    (unsigned char)prop.uuid.bytes[1],
@@ -80,23 +80,29 @@ index cb0d8528..4c829153 100644
 +                    (unsigned char)prop.uuid.bytes[14],
 +                    (unsigned char)prop.uuid.bytes[15]
 +                  );
-+                dev_ctx->uuid = uuid;
++                dev_ctx->id = id;
 +                #else
-+                dev_ctx->uuid = "GPU-" + std::string(prop.uuid.bytes, 16);
++                #ifdef _WIN32
++                char id[16];
++                snprintf(id, sizeof(id), "%d", i);
++                dev_ctx->id = id;
++                #else
++                dev_ctx->id = "GPU-" + std::string(prop.uuid.bytes, 16);
++                #endif
 +                #endif
 +
                  ggml_backend_dev_t dev = new ggml_backend_device {
                      /* .iface   = */ ggml_backend_cuda_device_interface,
                      /* .reg     = */ &reg,
 diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
-index 1b56f858..ee4f2dcb 100644
+index 1b56f858..a9eeebc6 100644
 --- a/ggml/src/ggml-metal/ggml-metal.m
 +++ b/ggml/src/ggml-metal/ggml-metal.m
 @@ -5703,6 +5703,7 @@ static enum ggml_backend_dev_type ggml_backend_metal_device_get_type(ggml_backen
  static void ggml_backend_metal_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
      props->name        = ggml_backend_metal_device_get_name(dev);
      props->description = ggml_backend_metal_device_get_description(dev);
-+    props->uuid        = "0";
++    props->id          = "0";
      props->type        = ggml_backend_metal_device_get_type(dev);
      ggml_backend_metal_device_get_memory(dev, &props->memory_free, &props->memory_total);
      props->caps = (struct ggml_backend_dev_caps) {

ml/backend.go

@@ -124,9 +124,9 @@ type DeviceMemory struct {
 	// may not be persistent across instances of the runner.
 	Name string
 
-	// UUID is a unique persistent identifier for the device for matching
-	// with system management libraries
-	UUID string
+	// ID is an identifier for the device for matching with system
+	// management libraries.
+	ID string
 
 	// Weights is the per-layer memory needed for the model weights.
 	Weights []Memory
@@ -156,8 +156,8 @@ func (m DeviceMemory) LogValue() slog.Value {
 		attrs = append(attrs, slog.Any("Graph", m.Graph))
 	}
 
-	if len(attrs) > 0 && m.UUID != "" {
-		attrs = append([]slog.Attr{slog.String("UUID", m.UUID)}, attrs...)
+	if len(attrs) > 0 && m.ID != "" {
+		attrs = append([]slog.Attr{slog.String("ID", m.ID)}, attrs...)
 	}
 
 	return slog.GroupValue(attrs...)

ml/backend/ggml/ggml.go

@@ -138,7 +138,7 @@ func New(modelPath string, params ml.BackendParams) (ml.Backend, error) {
 	requiredMemory.CPU.Name = C.GoString(C.ggml_backend_dev_name(cpuDeviceBufferType.d))
 	var props C.struct_ggml_backend_dev_props
 	C.ggml_backend_dev_get_props(cpuDeviceBufferType.d, &props)
-	requiredMemory.CPU.UUID = C.GoString(props.uuid)
+	requiredMemory.CPU.ID = C.GoString(props.id)
 	requiredMemory.CPU.Weights = make([]ml.Memory, blocks+1)
 	requiredMemory.CPU.Cache = make([]ml.Memory, blocks+1)
 
@@ -155,7 +155,7 @@ func New(modelPath string, params ml.BackendParams) (ml.Backend, error) {
 		requiredMemory.GPUs[i].Name = C.GoString(C.ggml_backend_dev_name(d))
 		var props C.struct_ggml_backend_dev_props
 		C.ggml_backend_dev_get_props(d, &props)
-		requiredMemory.GPUs[i].UUID = C.GoString(props.uuid)
+		requiredMemory.GPUs[i].ID = C.GoString(props.id)
 		requiredMemory.GPUs[i].Weights = make([]ml.Memory, blocks+1)
 		requiredMemory.GPUs[i].Cache = make([]ml.Memory, blocks+1)
 	}
@@ -356,23 +356,25 @@ func New(modelPath string, params ml.BackendParams) (ml.Backend, error) {
 	}
 
 	// Mimic llama runner logs summarizing layers and memory
-	slog.Info(fmt.Sprintf("offloading %d repeating layers to GPU", max(0, params.NumGPULayers-1)))
 	gpuLayers := 0
-	switch C.ggml_backend_dev_type(output.d) {
-	case 0: // CPU
-		slog.Info("offloading output layer to CPU")
-	case 1: // GPU
-		slog.Info("offloading output layer to GPU")
-		gpuLayers++
-	case 2: // ACCEL
-		slog.Info("offloading output layer to ACCEL")
-	}
 	for _, layer := range layers {
-		if C.ggml_backend_dev_type(layer.d) == 1 {
+		if C.ggml_backend_dev_type(layer.d) == C.GGML_BACKEND_DEVICE_TYPE_GPU {
 			gpuLayers++
 		}
 	}
+	slog.Info(fmt.Sprintf("offloading %d repeating layers to GPU", gpuLayers))
+
+	switch C.ggml_backend_dev_type(output.d) {
+	case C.GGML_BACKEND_DEVICE_TYPE_CPU:
+		slog.Info("offloading output layer to CPU")
+	case C.GGML_BACKEND_DEVICE_TYPE_GPU:
+		slog.Info("offloading output layer to GPU")
+		gpuLayers++
+	case C.GGML_BACKEND_DEVICE_TYPE_ACCEL:
+		slog.Info("offloading output layer to ACCEL")
+	}
 	slog.Info(fmt.Sprintf("offloaded %d/%d layers to GPU", gpuLayers, len(layers)+1))
+
 	for bs := range maps.Values(bbs) {
 		slog.Info("model weights", "buffer", C.GoString(C.ggml_backend_buffer_name(bs)), "size", format.HumanBytes2(uint64(C.ggml_backend_buffer_get_size(bs))))
 	}
@@ -418,7 +420,7 @@ func New(modelPath string, params ml.BackendParams) (ml.Backend, error) {
 			(*C.ggml_backend_buffer_type_t)(unsafe.Pointer(&schedBufts[0])),
 			C.int(len(schedBackends)),
 			C.size_t(maxGraphNodes),
-			C._Bool(len(gpus) > 1 && slices.Contains(gpus, output.d)),
+			C._Bool(false),
 			C._Bool(false),
 		),
 		schedBackends: schedBackends,

ml/backend/ggml/ggml/include/ggml-backend.h

@@ -152,7 +152,7 @@ extern "C" {
     struct ggml_backend_dev_props {
         const char * name;
         const char * description;
-        const char * uuid;
+        const char * id;
         size_t memory_free;
         size_t memory_total;
         enum ggml_backend_dev_type type;

ml/backend/ggml/ggml/src/ggml-cuda/ggml-cuda.cu

@@ -2888,7 +2888,7 @@ struct ggml_backend_cuda_device_context {
     int device;
     std::string name;
     std::string description;
-    std::string uuid;
+    std::string id;
 };
 
 static const char * ggml_backend_cuda_device_get_name(ggml_backend_dev_t dev) {
@@ -2901,9 +2901,9 @@ static const char * ggml_backend_cuda_device_get_description(ggml_backend_dev_t
     return ctx->description.c_str();
 }
 
-static const char * ggml_backend_cuda_device_get_uuid(ggml_backend_dev_t dev) {
+static const char * ggml_backend_cuda_device_get_id(ggml_backend_dev_t dev) {
     ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
-    return ctx->uuid.c_str();
+    return ctx->id.c_str();
 }
 
 static void ggml_backend_cuda_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
@@ -2920,7 +2920,7 @@ static enum ggml_backend_dev_type ggml_backend_cuda_device_get_type(ggml_backend
 static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
     props->name        = ggml_backend_cuda_device_get_name(dev);
     props->description = ggml_backend_cuda_device_get_description(dev);
-    props->uuid        = ggml_backend_cuda_device_get_uuid(dev);
+    props->id          = ggml_backend_cuda_device_get_id(dev);
     props->type        = ggml_backend_cuda_device_get_type(dev);
     ggml_backend_cuda_device_get_memory(dev, &props->memory_free, &props->memory_total);
 
@@ -3471,8 +3471,8 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
                 dev_ctx->description = prop.name;
 
                 #if !defined(GGML_USE_HIP)
-                char uuid[64];
-                snprintf(uuid, sizeof(uuid),
+                char id[64];
+                snprintf(id, sizeof(id),
                     "GPU-%02x%02x%02x%02x-%02x%02x-%02x%02x-%02x%02x-%02x%02x%02x%02x%02x%02x",
                     (unsigned char)prop.uuid.bytes[0],
                     (unsigned char)prop.uuid.bytes[1],
@@ -3491,9 +3491,15 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
                     (unsigned char)prop.uuid.bytes[14],
                     (unsigned char)prop.uuid.bytes[15]
                   );
-                dev_ctx->uuid = uuid;
+                dev_ctx->id = id;
                 #else
-                dev_ctx->uuid = "GPU-" + std::string(prop.uuid.bytes, 16);
+                #ifdef _WIN32
+                char id[16];
+                snprintf(id, sizeof(id), "%d", i);
+                dev_ctx->id = id;
+                #else
+                dev_ctx->id = "GPU-" + std::string(prop.uuid.bytes, 16);
+                #endif
                 #endif
 
                 ggml_backend_dev_t dev = new ggml_backend_device {

ml/backend/ggml/ggml/src/ggml-metal/ggml-metal.m

@@ -5726,7 +5726,7 @@ static enum ggml_backend_dev_type ggml_backend_metal_device_get_type(ggml_backen
 static void ggml_backend_metal_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
     props->name        = ggml_backend_metal_device_get_name(dev);
     props->description = ggml_backend_metal_device_get_description(dev);
-    props->uuid        = "0";
+    props->id          = "0";
     props->type        = ggml_backend_metal_device_get_type(dev);
     ggml_backend_metal_device_get_memory(dev, &props->memory_free, &props->memory_total);
     props->caps = (struct ggml_backend_dev_caps) {

ml/backend/ggml/ggml/src/ggml-metal/metal.go

@@ -4,6 +4,6 @@ package metal
 
 //go:generate sh -c "{ echo // Code generated by 'go generate'. DO NOT EDIT.; sed -e '/__embed_ggml-common.h__/r ../ggml-common.h' -e '/__embed_ggml-common.h__/d' -e '/#include \"ggml-metal-impl.h\"/r ggml-metal-impl.h' -e '/#include \"ggml-metal-impl.h\"/d' ggml-metal.metal; } >ggml-metal-embed.metal"
 
-// #cgo CPPFLAGS: -DGGML_METAL_NDEBUG -DGGML_METAL_EMBED_LIBRARY -I.. -I../../include
+// #cgo CPPFLAGS: -DGGML_METAL_NDEBUG -DGGML_METAL_EMBED_LIBRARY -DGGML_METAL_USE_BF16 -I.. -I../../include
 // #cgo LDFLAGS: -framework Metal -framework MetalKit
 import "C"

ml/backend/ggml/ggml/src/ggml.go

@@ -5,6 +5,7 @@ package ggml
 // #cgo CPPFLAGS: -I${SRCDIR}/../include -I${SRCDIR}/ggml-cpu
 // #cgo windows CFLAGS: -Wno-dll-attribute-on-redeclaration
 // #cgo windows LDFLAGS: -lmsvcrt -static -static-libgcc -static-libstdc++
+// #cgo windows linux CPPFLAGS: -DGGML_FP16_TO_FP32=ggml_compute_fp16_to_fp32
 // #include <stdlib.h>
 // #include "ggml-backend.h"
 // extern void sink(int level, char *text, void *user_data);

ml/backend/ggml/quantization.go

@@ -10,6 +10,8 @@ package ggml
 import "C"
 
 import (
+	"iter"
+	"slices"
 	"unsafe"
 
 	fsggml "github.com/ollama/ollama/fs/ggml"
@@ -50,34 +52,30 @@ func ConvertToF32(data []byte, dtype uint32, nelements uint64) []float32 {
 	return f32s
 }
 
-func Quantize(newType fsggml.TensorType, f32s []float32, shape []uint64) []byte {
-	buf := make([]byte, len(f32s)*4) // upper bound on size
-	nPerRow := C.int64_t(shape[0])
-	nrows := C.int64_t(1)
-	if len(shape) > 1 {
-		nrows = C.int64_t(shape[1])
-	}
-	shape2 := C.int64_t(1)
-	if len(shape) > 2 {
-		shape2 = C.int64_t(shape[2])
-	}
-	nelements_matrix := nPerRow * nrows
-	newSize := C.size_t(0)
-	for i03 := C.int64_t(0); i03 < shape2; i03++ {
-		f32s_03 := i03 * nelements_matrix
-		buf_03 := C.int64_t(C.ggml_row_size(uint32(newType), nPerRow)) * i03 * nrows
-		newSize += C.ggml_quantize_chunk(
-			uint32(newType),
-			(*C.float)(&f32s[f32s_03]),
-			unsafe.Pointer((uintptr)(unsafe.Pointer(&buf[0]))+uintptr(buf_03)),
-			0,
-			nrows,
-			nPerRow,
-			nil)
-	}
-	return buf[:newSize]
-}
+func Quantize(newType fsggml.TensorType, f32s []float32, shape []uint64) iter.Seq[[]byte] {
+	return func(yield func([]byte) bool) {
+		C.ggml_quantize_init(uint32(newType))
+		defer C.ggml_quantize_free()
 
-func QuantizationVersion() uint32 {
-	return uint32(C.GGML_QNT_VERSION)
+		dims := slices.Repeat([]C.int64_t{1}, 4)
+		for i, s := range shape {
+			dims[i] = C.int64_t(s)
+		}
+
+		bts := make([]byte, C.ggml_row_size(uint32(newType), dims[0])*C.size_t(dims[1]))
+		for chunk := range dims[2] {
+			offset := chunk * dims[0] * dims[1]
+
+			n := C.ggml_quantize_chunk(
+				uint32(newType),
+				(*C.float)(&f32s[0]),
+				unsafe.Pointer(&bts[0]),
+				offset, dims[1], dims[0], nil,
+			)
+
+			if !yield(bts[:n]) {
+				return
+			}
+		}
+	}
 }

model/models/llama/model.go

@@ -2,6 +2,7 @@ package llama
 
 import (
 	"cmp"
+	"fmt"
 	"math"
 
 	"github.com/ollama/ollama/fs"
@@ -33,6 +34,14 @@ type Model struct {
 }
 
 func New(c fs.Config) (model.Model, error) {
+	// This model currently only supports the gpt2 tokenizer
+	if c.String("tokenizer.ggml.model") == "llama" {
+		return nil, fmt.Errorf("unsupported tokenizer: llama")
+	}
+	// Best effort detection of library/deepseek-coder model(s) which are incompatible
+	if c.String("general.name") == "deepseek-ai" {
+		return nil, fmt.Errorf("unsupported model: %s", c.String("general.name"))
+	}
 	m := Model{
 		BytePairEncoding: model.NewBytePairEncoding(
 			c.String("tokenizer.ggml.pretokenizer", `(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+`),

model/models/qwen2/model.go

@@ -2,7 +2,9 @@ package qwen2
 
 import (
 	"cmp"
+	"fmt"
 	"math"
+	"strings"
 
 	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/kvcache"
@@ -126,6 +128,14 @@ func (m Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor
 }
 
 func New(c fs.Config) (model.Model, error) {
+	// This model currently only supports the gpt2 tokenizer
+	if c.String("tokenizer.ggml.model") == "llama" {
+		return nil, fmt.Errorf("unsupported tokenizer: llama")
+	}
+	// detect library/qwen model(s) which are incompatible
+	if strings.HasPrefix(c.String("general.name"), "Qwen2-beta") {
+		return nil, fmt.Errorf("unsupported model: %s", c.String("general.name"))
+	}
 	m := Model{
 		Layers: make([]DecoderLayer, c.Uint("block_count")),
 		BytePairEncoding: model.NewBytePairEncoding(

openai/openai.go

@@ -423,7 +423,7 @@ func fromChatRequest(r ChatCompletionRequest) (*api.ChatRequest, error) {
 						}
 					}
 
-					types := []string{"jpeg", "jpg", "png"}
+					types := []string{"jpeg", "jpg", "png", "webp"}
 					valid := false
 					for _, t := range types {
 						prefix := "data:image/" + t + ";base64,"

server/quantization.go

@@ -40,10 +40,19 @@ func (q quantizer) WriteTo(w io.Writer) (int64, error) {
 	} else {
 		f32s = ggml.ConvertToF32(data, q.from.Kind, q.from.Elements())
 	}
-	data = ggml.Quantize(newType, f32s, q.from.Shape)
-	n, err := w.Write(data)
-	q.progressFn(q.from.Size())
-	return int64(n), err
+
+	var n int64
+	for bts := range ggml.Quantize(newType, f32s, q.from.Shape) {
+		nn, err := w.Write(bts)
+		if err != nil {
+			return 0, err
+		}
+
+		q.progressFn(uint64(nn))
+		n += int64(nn)
+	}
+
+	return n, err
 }
 
 type quantizeState struct {

server/quantization_test.go

@@ -2,12 +2,14 @@ package server
 
 import (
 	"bytes"
+	"encoding/binary"
 	"fmt"
 	"math"
 	"os"
 	"strings"
 	"testing"
 
+	"github.com/google/go-cmp/cmp"
 	fsggml "github.com/ollama/ollama/fs/ggml"
 	"github.com/ollama/ollama/ml/backend/ggml"
 )
@@ -649,3 +651,55 @@ var (
 		},
 	}
 )
+
+func TestQuantizer(t *testing.T) {
+	from := fsggml.Tensor{
+		Name:  "fp32",
+		Shape: []uint64{256},
+		Kind:  uint32(fsggml.TensorTypeF32),
+	}
+
+	temp, err := os.CreateTemp(t.TempDir(), "*.bin")
+	if err != nil {
+		t.Fatalf("failed to create temp file: %v", err)
+	}
+
+	f32s := make([]float32, 256)
+	for i := range f32s {
+		f32s[i] = float32(i)
+	}
+
+	if err := binary.Write(temp, binary.LittleEndian, f32s); err != nil {
+		t.Fatalf("failed to write to temp file: %v", err)
+	}
+
+	for type_, want := range quantBytes {
+		t.Run(type_.String(), func(t *testing.T) {
+			f, err := os.Open(temp.Name())
+			if err != nil {
+				t.Fatalf("failed to open temp file: %v", err)
+			}
+			defer f.Close()
+
+			q := quantizer{
+				File: f,
+				from: &from,
+				to: &fsggml.Tensor{
+					Name:  type_.String(),
+					Shape: from.Shape,
+					Kind:  uint32(type_),
+				},
+				progressFn: func(uint64) {},
+			}
+
+			var b bytes.Buffer
+			if _, err := q.WriteTo(&b); err != nil {
+				t.Fatalf("WriteTo failed: %v", err)
+			}
+
+			if diff := cmp.Diff(b.Bytes(), want); diff != "" {
+				t.Errorf("quantized data mismatch for %s (-got +want):\n%s", type_, diff)
+			}
+		})
+	}
+}

server/routes.go

@@ -842,8 +842,11 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 	}
 	resp.Parameters = strings.Join(params, "\n")
 
-	for k, v := range req.Options {
-		if _, ok := req.Options[k]; ok {
+	if len(req.Options) > 0 {
+		if m.Options == nil {
+			m.Options = make(map[string]any)
+		}
+		for k, v := range req.Options {
 			m.Options[k] = v
 		}
 	}
@@ -1404,6 +1407,9 @@ func (s *Server) PsHandler(c *gin.Context) {
 			Details:   modelDetails,
 			ExpiresAt: v.expiresAt,
 		}
+		if v.Options != nil {
+			mr.ContextLength = v.Options.NumCtx / v.numParallel
+		}
 		// 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.

server/sched.go

@@ -57,9 +57,7 @@ type Scheduler struct {
 var defaultModelsPerGPU = 3
 
 // Default automatic value for parallel setting
-// Model will still need to fit in VRAM.  If this setting won't fit
-// we'll back off down to 1 to try to get it to fit
-var defaultParallel = 2
+var defaultParallel = 1
 
 var ErrMaxQueue = errors.New("server busy, please try again.  maximum pending requests exceeded")
 

tools/tools.go

@@ -115,21 +115,7 @@ func (p *Parser) findTag() (int, bool) {
 // parseToolCall finds the next complete tool call in the buffer
 // incrementing n and advancing the buffer.
 func (p *Parser) parseToolCall() *api.ToolCall {
-	var tool *api.Tool
-	var end int = len(p.buffer)
-	var i int
-
-	// find tool name
-	for _, t := range p.tools {
-		n := t.Function.Name
-		if i = bytes.Index(p.buffer, []byte(n)); i != -1 {
-			if i+len(n) < end {
-				tool = &t
-				end = i + len(n)
-			}
-		}
-	}
-
+	tool, end := findTool(p.tools, p.buffer)
 	if tool == nil {
 		return nil
 	}
@@ -139,10 +125,10 @@ func (p *Parser) parseToolCall() *api.ToolCall {
 	// parsing arguments before the tool name, which may be needed in the future
 	args := map[string]any{}
 	if len(tool.Function.Parameters.Properties) > 0 {
+		var i int
 		if args, i = findArguments(*tool, p.buffer[end:]); args == nil {
 			return nil
 		}
-
 		end += i
 	}
 
@@ -159,9 +145,74 @@ func (p *Parser) parseToolCall() *api.ToolCall {
 	return tc
 }
 
+// findTool finds the first tool name in the list that matches the
+// beginning of the buffer, returning nil if no tool is found
+// or if the buffer ends with a partial tool name since we need
+// to wait for more data to disambiguate.
+// The second return value is the end position of the tool name
+// if one is found, otherwise 0.
+func findTool(tools []api.Tool, buf []byte) (*api.Tool, int) {
+	if len(buf) == 0 {
+		return nil, 0
+	}
+
+	// check if buffer ends with a partial tool name
+	// this prevents matching "get" when seeing "get_weather"
+	var longest string
+	for _, t := range tools {
+		if len(t.Function.Name) > len(longest) {
+			longest = t.Function.Name
+		}
+	}
+
+	// Only check up to longest characters from the end
+	for i := 1; i <= min(len(buf), len(longest)); i++ {
+		tail := buf[len(buf)-i:]
+		for _, t := range tools {
+			name := []byte(t.Function.Name)
+			if len(tail) < len(name) && bytes.HasPrefix(name, tail) {
+				return nil, 0
+			}
+		}
+	}
+
+	// find first occurrence of the longest tool name
+	var found *api.Tool
+	start := -1
+	end := -1
+
+	for i := range tools {
+		name := []byte(tools[i].Function.Name)
+		pos := bytes.Index(buf, name)
+		if pos == -1 {
+			continue
+		}
+
+		// Skip if we have a better match already
+		if start != -1 {
+			if pos > start {
+				continue
+			}
+			if pos == start && len(name) <= len(found.Function.Name) {
+				continue
+			}
+		}
+
+		found = &tools[i]
+		start = pos
+		end = pos + len(name)
+	}
+
+	if found != nil {
+		return found, end
+	}
+
+	return nil, 0
+}
+
 // findArguments returns the first object that appears to be
 // arguments for the provided tool in the provided buffer,
-// returning nil if no arguments are found.
+// returning nil if no arguments are found and the end position
 // TODO (jmorganca): this does not support parsing omitted arguments
 // objects for functions that have all-optional parameters
 // e.g. `{"name": "get_conditions", "arguments": {}}` will work but

tools/tools_test.go

@@ -112,6 +112,81 @@ func TestParser(t *testing.T) {
 				Description: "Say hello",
 			},
 		},
+		{
+			Type: "function",
+			Function: api.ToolFunction{
+				Name:        "say_hello_world",
+				Description: "Say hello world",
+			},
+		},
+		{
+			Type: "function",
+			Function: api.ToolFunction{
+				Name:        "get_address",
+				Description: "Get the address of a given location",
+				Parameters: struct {
+					Type       string   `json:"type"`
+					Defs       any      `json:"$defs,omitempty"`
+					Items      any      `json:"items,omitempty"`
+					Required   []string `json:"required"`
+					Properties map[string]struct {
+						Type        api.PropertyType `json:"type"`
+						Items       any              `json:"items,omitempty"`
+						Description string           `json:"description"`
+						Enum        []any            `json:"enum,omitempty"`
+					} `json:"properties"`
+				}{
+					Type: "object",
+					Properties: map[string]struct {
+						Type        api.PropertyType `json:"type"`
+						Items       any              `json:"items,omitempty"`
+						Description string           `json:"description"`
+						Enum        []any            `json:"enum,omitempty"`
+					}{
+						"location": {
+							Type:        api.PropertyType{"string"},
+							Description: "The location to get the address for",
+						},
+					},
+				},
+			},
+		},
+		{
+			Type: "function",
+			Function: api.ToolFunction{
+				Name:        "add",
+				Description: "Add two numbers",
+				Parameters: struct {
+					Type       string   `json:"type"`
+					Defs       any      `json:"$defs,omitempty"`
+					Items      any      `json:"items,omitempty"`
+					Required   []string `json:"required"`
+					Properties map[string]struct {
+						Type        api.PropertyType `json:"type"`
+						Items       any              `json:"items,omitempty"`
+						Description string           `json:"description"`
+						Enum        []any            `json:"enum,omitempty"`
+					} `json:"properties"`
+				}{
+					Type: "object",
+					Properties: map[string]struct {
+						Type        api.PropertyType `json:"type"`
+						Items       any              `json:"items,omitempty"`
+						Description string           `json:"description"`
+						Enum        []any            `json:"enum,omitempty"`
+					}{
+						"a": {
+							Type:        api.PropertyType{"string"},
+							Description: "The first number to add",
+						},
+						"b": {
+							Type:        api.PropertyType{"string"},
+							Description: "The second number to add",
+						},
+					},
+				},
+			},
+		},
 	}
 
 	tests := []struct {
@@ -629,6 +704,173 @@ func TestParser(t *testing.T) {
 				},
 			},
 		},
+		{
+			name: "tool name with collision",
+			inputs: []string{
+				"<tool_call>",
+				"{",
+				"\"name\": \"say_hello",
+				"_world\",",
+				"}",
+				"}",
+			},
+			content: "",
+			tmpl:    qwen,
+			calls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Index:     0,
+						Name:      "say_hello_world",
+						Arguments: api.ToolCallFunctionArguments{},
+					},
+				},
+			},
+		},
+		{
+			name: "tool name with collision multiple",
+			inputs: []string{
+				"<tool_call>",
+				"{",
+				"\"name\": \"say_hello",
+				"_world\",",
+				"}",
+				"</tool_call>",
+				"<tool_call>",
+				"{",
+				"\"name\": \"say_hello",
+				"\",",
+				"}",
+				"</tool_call>",
+			},
+			content: "",
+			tmpl:    qwen,
+			calls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Index:     0,
+						Name:      "say_hello_world",
+						Arguments: api.ToolCallFunctionArguments{},
+					},
+				},
+				{
+					Function: api.ToolCallFunction{
+						Index:     1,
+						Name:      "say_hello",
+						Arguments: api.ToolCallFunctionArguments{},
+					},
+				},
+			},
+		},
+		{
+			name: "tool name with collision non streaming",
+			inputs: []string{
+				`<tool_call>{"name": "say_hello`,
+			},
+			content: "",
+			tmpl:    qwen,
+			calls:   nil,
+		},
+		{
+			name: "tool name with collision non streaming multiple",
+			inputs: []string{
+				`<tool_call>{"name": "say_hello"}</tool_call><tool_call>{"name": "say_hello_world"}`,
+			},
+			content: "",
+			tmpl:    qwen,
+			calls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Index:     0,
+						Name:      "say_hello",
+						Arguments: api.ToolCallFunctionArguments{},
+					},
+				},
+				{
+					Function: api.ToolCallFunction{
+						Index:     1,
+						Name:      "say_hello_world",
+						Arguments: api.ToolCallFunctionArguments{},
+					},
+				},
+			},
+		},
+		{
+			name: "tool name with collision non streaming shorter",
+			inputs: []string{
+				`<tool_call>{"name": "say_hello"}</tool_call>`,
+			},
+			content: "",
+			tmpl:    qwen,
+			calls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Index:     0,
+						Name:      "say_hello",
+						Arguments: api.ToolCallFunctionArguments{},
+					},
+				},
+			},
+		},
+		{
+			name: "tool name with collision non streaming longer",
+			inputs: []string{
+				`<tool_call>{"name": "say_hello_world"}</tool_call>`,
+			},
+			content: "",
+			tmpl:    qwen,
+			calls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Index:     0,
+						Name:      "say_hello_world",
+						Arguments: api.ToolCallFunctionArguments{},
+					},
+				},
+			},
+		},
+		{
+			name: "tool name with substring of another",
+			inputs: []string{
+				"{",
+				"\"name\": \"get_address\",",
+				"\"arguments\": {",
+				"\"location\": \"London\"",
+				"}",
+				"}",
+			},
+			content: "",
+			tmpl:    json,
+			calls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Index: 0,
+						Name:  "get_address",
+						Arguments: api.ToolCallFunctionArguments{
+							"location": "London",
+						},
+					},
+				},
+			},
+		},
+		{
+			name: "tool name with substring of another",
+			inputs: []string{
+				`<tool_call>{"name": "get_address", "arguments": {"location": "London"}}</tool_call>`,
+			},
+			content: "",
+			tmpl:    qwen,
+			calls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Index: 0,
+						Name:  "get_address",
+						Arguments: api.ToolCallFunctionArguments{
+							"location": "London",
+						},
+					},
+				},
+			},
+		},
 	}
 
 	for _, tt := range tests {