Reincluding Numbers

Enabling ollama to run on Intel GPUs with SYCL backend

.github/workflows/release.yaml

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

README.md

@@ -51,15 +51,17 @@ Here are some example models that can be downloaded:
 | ------------------ | ---------- | ----- | ------------------------------ |
 | Llama 3            | 8B         | 4.7GB | `ollama run llama3`            |
 | Llama 3            | 70B        | 40GB  | `ollama run llama3:70b`        |
-| Phi-3              | 3.8B       | 2.3GB | `ollama run phi3`              |
+| Phi 3 Mini         | 3.8B       | 2.3GB | `ollama run phi3`              |
+| Phi 3 Medium       | 14B        | 7.9GB | `ollama run phi3:medium`       |
+| Gemma              | 2B         | 1.4GB | `ollama run gemma:2b`          |
+| Gemma              | 7B         | 4.8GB | `ollama run gemma:7b`          |
 | Mistral            | 7B         | 4.1GB | `ollama run mistral`           |
+| Moondream 2        | 1.4B       | 829MB | `ollama run moondream`         |
 | Neural Chat        | 7B         | 4.1GB | `ollama run neural-chat`       |
 | Starling           | 7B         | 4.1GB | `ollama run starling-lm`       |
 | Code Llama         | 7B         | 3.8GB | `ollama run codellama`         |
 | Llama 2 Uncensored | 7B         | 3.8GB | `ollama run llama2-uncensored` |
 | LLaVA              | 7B         | 4.5GB | `ollama run llava`             |
-| Gemma              | 2B         | 1.4GB | `ollama run gemma:2b`          |
-| Gemma              | 7B         | 4.8GB | `ollama run gemma:7b`          |
 | Solar              | 10.7B      | 6.1GB | `ollama run solar`             |
 
 > Note: You should have at least 8 GB of RAM available to run the 7B models, 16 GB to run the 13B models, and 32 GB to run the 33B models.
@@ -192,25 +194,7 @@ ollama list
 
 ## Building
 
-Install `cmake` and `go`:
-
-```
-brew install cmake go
-```
-
-Then generate dependencies:
-
-```
-go generate ./...
-```
-
-Then build the binary:
-
-```
-go build .
-```
-
-More detailed instructions can be found in the [developer guide](https://github.com/ollama/ollama/blob/main/docs/development.md)
+See the [developer guide](https://github.com/ollama/ollama/blob/main/docs/development.md)
 
 ### Running local builds
 
@@ -258,6 +242,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 
 - [Open WebUI](https://github.com/open-webui/open-webui)
 - [Enchanted (macOS native)](https://github.com/AugustDev/enchanted)
+- [Hollama](https://github.com/fmaclen/hollama)
 - [Lollms-Webui](https://github.com/ParisNeo/lollms-webui)
 - [LibreChat](https://github.com/danny-avila/LibreChat)
 - [Bionic GPT](https://github.com/bionic-gpt/bionic-gpt)
@@ -297,6 +282,9 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Lobe Chat](https://github.com/lobehub/lobe-chat) with [Integrating Doc](https://lobehub.com/docs/self-hosting/examples/ollama)
 - [Ollama RAG Chatbot](https://github.com/datvodinh/rag-chatbot.git) (Local Chat with multiple PDFs using Ollama and RAG)
 - [BrainSoup](https://www.nurgo-software.com/products/brainsoup) (Flexible native client with RAG & multi-agent automation)
+- [macai](https://github.com/Renset/macai) (macOS client for Ollama, ChatGPT, and other compatible API back-ends)
+- [Olpaka](https://github.com/Otacon/olpaka) (User-friendly Flutter Web App for Ollama)
+- [OllamaSpring](https://github.com/CrazyNeil/OllamaSpring) (Ollama Client for macOS)
 
 ### Terminal
 
@@ -329,6 +317,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 
 - [Pacman](https://archlinux.org/packages/extra/x86_64/ollama/)
 - [Helm Chart](https://artifacthub.io/packages/helm/ollama-helm/ollama)
+- [Guix channel](https://codeberg.org/tusharhero/ollama-guix)
 
 ### Libraries
 
@@ -355,7 +344,8 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Ollama Connector for SAP ABAP](https://github.com/b-tocs/abap_btocs_ollama)
 - [Testcontainers](https://testcontainers.com/modules/ollama/)
 - [Portkey](https://portkey.ai/docs/welcome/integration-guides/ollama)
-
+- [PromptingTools.jl](https://github.com/svilupp/PromptingTools.jl) with an [example](https://svilupp.github.io/PromptingTools.jl/dev/examples/working_with_ollama)
+- [LlamaScript](https://github.com/Project-Llama/llamascript)
 ### Mobile
 
 - [Enchanted](https://github.com/AugustDev/enchanted)
@@ -387,6 +377,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [AI Telegram Bot](https://github.com/tusharhero/aitelegrambot) (Telegram bot using Ollama in backend)
 - [AI ST Completion](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (Sublime Text 4 AI assistant plugin with Ollama support)
 - [Discord-Ollama Chat Bot](https://github.com/kevinthedang/discord-ollama) (Generalized TypeScript Discord Bot w/ Tuning Documentation)
+- [Discord AI chat/moderation bot](https://github.com/rapmd73/Companion) Chat/moderation bot written in python. Uses Ollama to create personalities.
 
 ### Supported backends 
 - [llama.cpp](https://github.com/ggerganov/llama.cpp) project founded by Georgi Gerganov.

api/client.go

@@ -1,9 +1,16 @@
 // Package api implements the client-side API for code wishing to interact
 // with the ollama service. The methods of the [Client] type correspond to
-// the ollama REST API as described in https://github.com/ollama/ollama/blob/main/docs/api.md
-//
+// the ollama REST API as described in [the API documentation].
 // The ollama command-line client itself uses this package to interact with
 // the backend service.
+//
+// # Examples
+//
+// Several examples of using this package are available [in the GitHub
+// repository].
+//
+// [the API documentation]: https://github.com/ollama/ollama/blob/main/docs/api.md
+// [in the GitHub repository]: https://github.com/ollama/ollama/tree/main/examples
 package api
 
 import (
@@ -299,8 +306,14 @@ func (c *Client) Pull(ctx context.Context, req *PullRequest, fn PullProgressFunc
 	})
 }
 
+// PushProgressFunc is a function that [Client.Push] invokes when progress is
+// made.
+// It's similar to other progress function types like [PullProgressFunc].
 type PushProgressFunc func(ProgressResponse) error
 
+// Push uploads a model to the model library; requires registering for ollama.ai
+// and adding a public key first. fn is called each time progress is made on
+// the request and can be used to display a progress bar, etc.
 func (c *Client) Push(ctx context.Context, req *PushRequest, fn PushProgressFunc) error {
 	return c.stream(ctx, http.MethodPost, "/api/push", req, func(bts []byte) error {
 		var resp ProgressResponse
@@ -312,8 +325,15 @@ func (c *Client) Push(ctx context.Context, req *PushRequest, fn PushProgressFunc
 	})
 }
 
+// CreateProgressFunc is a function that [Client.Create] invokes when progress
+// is made.
+// It's similar to other progress function types like [PullProgressFunc].
 type CreateProgressFunc func(ProgressResponse) error
 
+// Create creates a model from a [Modelfile]. fn is a progress function that
+// behaves similarly to other methods (see [Client.Pull]).
+//
+// [Modelfile]: https://github.com/ollama/ollama/blob/main/docs/modelfile.md
 func (c *Client) Create(ctx context.Context, req *CreateRequest, fn CreateProgressFunc) error {
 	return c.stream(ctx, http.MethodPost, "/api/create", req, func(bts []byte) error {
 		var resp ProgressResponse
@@ -325,6 +345,7 @@ func (c *Client) Create(ctx context.Context, req *CreateRequest, fn CreateProgre
 	})
 }
 
+// List lists models that are available locally.
 func (c *Client) List(ctx context.Context) (*ListResponse, error) {
 	var lr ListResponse
 	if err := c.do(ctx, http.MethodGet, "/api/tags", nil, &lr); err != nil {
@@ -333,6 +354,17 @@ func (c *Client) List(ctx context.Context) (*ListResponse, error) {
 	return &lr, nil
 }
 
+// List running models.
+func (c *Client) ListRunning(ctx context.Context) (*ListResponse, error) {
+	var lr ListResponse
+	if err := c.do(ctx, http.MethodGet, "/api/ps", nil, &lr); err != nil {
+		return nil, err
+	}
+	return &lr, nil
+}
+
+// Copy copies a model - creating a model with another name from an existing
+// model.
 func (c *Client) Copy(ctx context.Context, req *CopyRequest) error {
 	if err := c.do(ctx, http.MethodPost, "/api/copy", req, nil); err != nil {
 		return err
@@ -340,6 +372,7 @@ func (c *Client) Copy(ctx context.Context, req *CopyRequest) error {
 	return nil
 }
 
+// Delete deletes a model and its data.
 func (c *Client) Delete(ctx context.Context, req *DeleteRequest) error {
 	if err := c.do(ctx, http.MethodDelete, "/api/delete", req, nil); err != nil {
 		return err
@@ -347,6 +380,7 @@ func (c *Client) Delete(ctx context.Context, req *DeleteRequest) error {
 	return nil
 }
 
+// Show obtains model information, including details, modelfile, license etc.
 func (c *Client) Show(ctx context.Context, req *ShowRequest) (*ShowResponse, error) {
 	var resp ShowResponse
 	if err := c.do(ctx, http.MethodPost, "/api/show", req, &resp); err != nil {
@@ -355,12 +389,16 @@ func (c *Client) Show(ctx context.Context, req *ShowRequest) (*ShowResponse, err
 	return &resp, nil
 }
 
+// Hearbeat checks if the server has started and is responsive; if yes, it
+// returns nil, otherwise an error.
 func (c *Client) Heartbeat(ctx context.Context) error {
 	if err := c.do(ctx, http.MethodHead, "/", nil, nil); err != nil {
 		return err
 	}
 	return nil
 }
+
+// Embeddings generates embeddings from a model.
 func (c *Client) Embeddings(ctx context.Context, req *EmbeddingRequest) (*EmbeddingResponse, error) {
 	var resp EmbeddingResponse
 	if err := c.do(ctx, http.MethodPost, "/api/embeddings", req, &resp); err != nil {
@@ -369,10 +407,13 @@ func (c *Client) Embeddings(ctx context.Context, req *EmbeddingRequest) (*Embedd
 	return &resp, nil
 }
 
+// CreateBlob creates a blob from a file on the server. digest is the
+// expected SHA256 digest of the file, and r represents the file.
 func (c *Client) CreateBlob(ctx context.Context, digest string, r io.Reader) error {
 	return c.do(ctx, http.MethodPost, fmt.Sprintf("/api/blobs/%s", digest), r, nil)
 }
 
+// Version returns the Ollama server version as a string.
 func (c *Client) Version(ctx context.Context) (string, error) {
 	var version struct {
 		Version string `json:"version"`

api/types.go

@@ -4,6 +4,7 @@ import (
 	"encoding/json"
 	"errors"
 	"fmt"
+	"log/slog"
 	"math"
 	"os"
 	"reflect"
@@ -12,6 +13,7 @@ import (
 	"time"
 )
 
+// StatusError is an error with and HTTP status code.
 type StatusError struct {
 	StatusCode   int
 	Status       string
@@ -32,6 +34,7 @@ func (e StatusError) Error() string {
 	}
 }
 
+// ImageData represents the raw binary data of an image file.
 type ImageData []byte
 
 // GenerateRequest describes a request sent by [Client.Generate]. While you
@@ -77,26 +80,44 @@ type GenerateRequest struct {
 	Options map[string]interface{} `json:"options"`
 }
 
+// ChatRequest describes a request sent by [Client.Chat].
 type ChatRequest struct {
-	Model     string    `json:"model"`
-	Messages  []Message `json:"messages"`
-	Stream    *bool     `json:"stream,omitempty"`
-	Format    string    `json:"format"`
+	// Model is the model name, as in [GenerateRequest].
+	Model string `json:"model"`
+
+	// Messages is the messages of the chat - can be used to keep a chat memory.
+	Messages []Message `json:"messages"`
+
+	// Stream enable streaming of returned response; true by default.
+	Stream *bool `json:"stream,omitempty"`
+
+	// Format is the format to return the response in (e.g. "json").
+	Format string `json:"format"`
+
+	// KeepAlive controls how long the model will stay loaded into memory
+	// followin the request.
 	KeepAlive *Duration `json:"keep_alive,omitempty"`
 
+	// Options lists model-specific options.
 	Options map[string]interface{} `json:"options"`
 }
 
+// Message is a single message in a chat sequence. The message contains the
+// role ("system", "user", or "assistant"), the content and an optional list
+// of images.
 type Message struct {
-	Role    string      `json:"role"` // one of ["system", "user", "assistant"]
+	Role    string      `json:"role"`
 	Content string      `json:"content"`
 	Images  []ImageData `json:"images,omitempty"`
 }
 
+// ChatResponse is the response returned by [Client.Chat]. Its fields are
+// similar to [GenerateResponse].
 type ChatResponse struct {
-	Model     string    `json:"model"`
-	CreatedAt time.Time `json:"created_at"`
-	Message   Message   `json:"message"`
+	Model      string    `json:"model"`
+	CreatedAt  time.Time `json:"created_at"`
+	Message    Message   `json:"message"`
+	DoneReason string    `json:"done_reason,omitempty"`
 
 	Done bool `json:"done"`
 
@@ -112,7 +133,8 @@ type Metrics struct {
 	EvalDuration       time.Duration `json:"eval_duration,omitempty"`
 }
 
-// Options specified in GenerateRequest, if you add a new option here add it to the API docs also
+// Options specified in [GenerateRequest], if you add a new option here add it
+// to the API docs also.
 type Options struct {
 	Runner
 
@@ -141,7 +163,6 @@ type Runner struct {
 	UseNUMA   bool `json:"numa,omitempty"`
 	NumCtx    int  `json:"num_ctx,omitempty"`
 	NumBatch  int  `json:"num_batch,omitempty"`
-	NumGQA    int  `json:"num_gqa,omitempty"`
 	NumGPU    int  `json:"num_gpu,omitempty"`
 	MainGPU   int  `json:"main_gpu,omitempty"`
 	LowVRAM   bool `json:"low_vram,omitempty"`
@@ -151,36 +172,45 @@ type Runner struct {
 	UseMMap   bool `json:"use_mmap,omitempty"`
 	UseMLock  bool `json:"use_mlock,omitempty"`
 	NumThread int  `json:"num_thread,omitempty"`
-
-	// Unused: RopeFrequencyBase is ignored. Instead the value in the model will be used
-	RopeFrequencyBase float32 `json:"rope_frequency_base,omitempty"`
-	// Unused: RopeFrequencyScale is ignored. Instead the value in the model will be used
-	RopeFrequencyScale float32 `json:"rope_frequency_scale,omitempty"`
 }
 
+// EmbeddingRequest is the request passed to [Client.Embeddings].
 type EmbeddingRequest struct {
-	Model     string    `json:"model"`
-	Prompt    string    `json:"prompt"`
+	// Model is the model name.
+	Model string `json:"model"`
+
+	// Prompt is the textual prompt to embed.
+	Prompt string `json:"prompt"`
+
+	// KeepAlive controls how long the model will stay loaded in memory following
+	// this request.
 	KeepAlive *Duration `json:"keep_alive,omitempty"`
 
+	// Options lists model-specific options.
 	Options map[string]interface{} `json:"options"`
 }
 
+// EmbeddingResponse is the response from [Client.Embeddings].
 type EmbeddingResponse struct {
 	Embedding []float64 `json:"embedding"`
 }
 
+// CreateRequest is the request passed to [Client.Create].
 type CreateRequest struct {
-	Model        string `json:"model"`
-	Path         string `json:"path"`
-	Modelfile    string `json:"modelfile"`
-	Stream       *bool  `json:"stream,omitempty"`
-	Quantization string `json:"quantization,omitempty"`
+	Model     string `json:"model"`
+	Path      string `json:"path"`
+	Modelfile string `json:"modelfile"`
+	Stream    *bool  `json:"stream,omitempty"`
+	Quantize  string `json:"quantize,omitempty"`
 
 	// Name is deprecated, see Model
 	Name string `json:"name"`
+
+	// Quantization is deprecated, see Quantize
+	Quantization string `json:"quantization,omitempty"`
 }
 
+// DeleteRequest is the request passed to [Client.Delete].
 type DeleteRequest struct {
 	Model string `json:"model"`
 
@@ -188,6 +218,7 @@ type DeleteRequest struct {
 	Name string `json:"name"`
 }
 
+// ShowRequest is the request passed to [Client.Show].
 type ShowRequest struct {
 	Model    string `json:"model"`
 	System   string `json:"system"`
@@ -199,6 +230,7 @@ type ShowRequest struct {
 	Name string `json:"name"`
 }
 
+// ShowResponse is the response returned from [Client.Show].
 type ShowResponse struct {
 	License    string       `json:"license,omitempty"`
 	Modelfile  string       `json:"modelfile,omitempty"`
@@ -209,11 +241,13 @@ type ShowResponse struct {
 	Messages   []Message    `json:"messages,omitempty"`
 }
 
+// CopyRequest is the request passed to [Client.Copy].
 type CopyRequest struct {
 	Source      string `json:"source"`
 	Destination string `json:"destination"`
 }
 
+// PullRequest is the request passed to [Client.Pull].
 type PullRequest struct {
 	Model    string `json:"model"`
 	Insecure bool   `json:"insecure,omitempty"`
@@ -225,6 +259,8 @@ type PullRequest struct {
 	Name string `json:"name"`
 }
 
+// ProgressResponse is the response passed to progress functions like
+// [PullProgressFunc] and [PushProgressFunc].
 type ProgressResponse struct {
 	Status    string `json:"status"`
 	Digest    string `json:"digest,omitempty"`
@@ -232,6 +268,7 @@ type ProgressResponse struct {
 	Completed int64  `json:"completed,omitempty"`
 }
 
+// PushRequest is the request passed to [Client.Push].
 type PushRequest struct {
 	Model    string `json:"model"`
 	Insecure bool   `json:"insecure,omitempty"`
@@ -243,34 +280,52 @@ type PushRequest struct {
 	Name string `json:"name"`
 }
 
+// ListResponse is the response from [Client.List].
 type ListResponse struct {
 	Models []ModelResponse `json:"models"`
 }
 
+// ModelResponse is a single model description in [ListResponse].
 type ModelResponse struct {
 	Name       string       `json:"name"`
 	Model      string       `json:"model"`
-	ModifiedAt time.Time    `json:"modified_at"`
+	ModifiedAt time.Time    `json:"modified_at,omitempty"`
 	Size       int64        `json:"size"`
 	Digest     string       `json:"digest"`
 	Details    ModelDetails `json:"details,omitempty"`
+	ExpiresAt  time.Time    `json:"expires_at,omitempty"`
+	SizeVRAM   int64        `json:"size_vram,omitempty"`
 }
 
 type TokenResponse struct {
 	Token string `json:"token"`
 }
 
+// GenerateResponse is the response passed into [GenerateResponseFunc].
 type GenerateResponse struct {
-	Model     string    `json:"model"`
-	CreatedAt time.Time `json:"created_at"`
-	Response  string    `json:"response"`
+	// Model is the model name that generated the response.
+	Model string `json:"model"`
 
-	Done    bool  `json:"done"`
+	//CreatedAt is the timestamp of the response.
+	CreatedAt time.Time `json:"created_at"`
+
+	// Response is the textual response itself.
+	Response string `json:"response"`
+
+	// Done specifies if the response is complete.
+	Done bool `json:"done"`
+
+	// DoneReason is the reason the model stopped generating text.
+	DoneReason string `json:"done_reason,omitempty"`
+
+	// Context is an encoding of the conversation used in this response; this
+	// can be sent in the next request to keep a conversational memory.
 	Context []int `json:"context,omitempty"`
 
 	Metrics
 }
 
+// ModelDetails provides details about a model.
 type ModelDetails struct {
 	ParentModel       string   `json:"parent_model"`
 	Format            string   `json:"format"`
@@ -308,7 +363,6 @@ func (m *Metrics) Summary() {
 	}
 }
 
-var ErrInvalidOpts = errors.New("invalid options")
 var ErrInvalidHostPort = errors.New("invalid port specified in OLLAMA_HOST")
 
 func (opts *Options) FromMap(m map[string]interface{}) error {
@@ -324,76 +378,76 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
 		}
 	}
 
-	invalidOpts := []string{}
 	for key, val := range m {
-		if opt, ok := jsonOpts[key]; ok {
-			field := valueOpts.FieldByName(opt.Name)
-			if field.IsValid() && field.CanSet() {
-				if val == nil {
-					continue
-				}
+		opt, ok := jsonOpts[key]
+		if !ok {
+			slog.Warn("invalid option provided", "option", opt.Name)
+			continue
+		}
 
-				switch field.Kind() {
-				case reflect.Int:
-					switch t := val.(type) {
-					case int64:
-						field.SetInt(t)
-					case float64:
-						// when JSON unmarshals numbers, it uses float64, not int
-						field.SetInt(int64(t))
-					default:
-						return fmt.Errorf("option %q must be of type integer", key)
-					}
-				case reflect.Bool:
-					val, ok := val.(bool)
-					if !ok {
-						return fmt.Errorf("option %q must be of type boolean", key)
-					}
-					field.SetBool(val)
-				case reflect.Float32:
-					// JSON unmarshals to float64
-					val, ok := val.(float64)
-					if !ok {
-						return fmt.Errorf("option %q must be of type float32", key)
-					}
-					field.SetFloat(val)
-				case reflect.String:
-					val, ok := val.(string)
-					if !ok {
-						return fmt.Errorf("option %q must be of type string", key)
-					}
-					field.SetString(val)
-				case reflect.Slice:
-					// JSON unmarshals to []interface{}, not []string
-					val, ok := val.([]interface{})
-					if !ok {
-						return fmt.Errorf("option %q must be of type array", key)
-					}
-					// convert []interface{} to []string
-					slice := make([]string, len(val))
-					for i, item := range val {
-						str, ok := item.(string)
-						if !ok {
-							return fmt.Errorf("option %q must be of an array of strings", key)
-						}
-						slice[i] = str
-					}
-					field.Set(reflect.ValueOf(slice))
-				default:
-					return fmt.Errorf("unknown type loading config params: %v", field.Kind())
-				}
+		field := valueOpts.FieldByName(opt.Name)
+		if field.IsValid() && field.CanSet() {
+			if val == nil {
+				continue
+			}
+
+			switch field.Kind() {
+			case reflect.Int:
+				switch t := val.(type) {
+				case int64:
+					field.SetInt(t)
+				case float64:
+					// when JSON unmarshals numbers, it uses float64, not int
+					field.SetInt(int64(t))
+				default:
+					return fmt.Errorf("option %q must be of type integer", key)
+				}
+			case reflect.Bool:
+				val, ok := val.(bool)
+				if !ok {
+					return fmt.Errorf("option %q must be of type boolean", key)
+				}
+				field.SetBool(val)
+			case reflect.Float32:
+				// JSON unmarshals to float64
+				val, ok := val.(float64)
+				if !ok {
+					return fmt.Errorf("option %q must be of type float32", key)
+				}
+				field.SetFloat(val)
+			case reflect.String:
+				val, ok := val.(string)
+				if !ok {
+					return fmt.Errorf("option %q must be of type string", key)
+				}
+				field.SetString(val)
+			case reflect.Slice:
+				// JSON unmarshals to []interface{}, not []string
+				val, ok := val.([]interface{})
+				if !ok {
+					return fmt.Errorf("option %q must be of type array", key)
+				}
+				// convert []interface{} to []string
+				slice := make([]string, len(val))
+				for i, item := range val {
+					str, ok := item.(string)
+					if !ok {
+						return fmt.Errorf("option %q must be of an array of strings", key)
+					}
+					slice[i] = str
+				}
+				field.Set(reflect.ValueOf(slice))
+			default:
+				return fmt.Errorf("unknown type loading config params: %v", field.Kind())
 			}
-		} else {
-			invalidOpts = append(invalidOpts, key)
 		}
 	}
 
-	if len(invalidOpts) > 0 {
-		return fmt.Errorf("%w: %v", ErrInvalidOpts, strings.Join(invalidOpts, ", "))
-	}
 	return nil
 }
 
+// DefaultOptions is the default set of options for [GenerateRequest]; these
+// values are used unless the user specifies other values explicitly.
 func DefaultOptions() Options {
 	return Options{
 		// options set on request to runner
@@ -421,8 +475,7 @@ func DefaultOptions() Options {
 			NumCtx:    2048,
 			NumBatch:  512,
 			NumGPU:    -1, // -1 here indicates that NumGPU should be set dynamically
-			NumGQA:    1,
-			NumThread: 0, // let the runtime decide
+			NumThread: 0,  // let the runtime decide
 			LowVRAM:   false,
 			F16KV:     true,
 			UseMLock:  false,

app/lifecycle/logging.go

@@ -6,7 +6,7 @@ import (
 	"os"
 	"path/filepath"
 
-	"github.com/ollama/ollama/server/envconfig"
+	"github.com/ollama/ollama/envconfig"
 )
 
 func InitLogging() {

cmd/cmd.go

@@ -12,6 +12,7 @@ import (
 	"fmt"
 	"io"
 	"log"
+	"math"
 	"net"
 	"net/http"
 	"os"
@@ -24,7 +25,7 @@ import (
 	"time"
 
 	"github.com/containerd/console"
-
+	"github.com/mattn/go-runewidth"
 	"github.com/olekukonko/tablewriter"
 	"github.com/spf13/cobra"
 	"golang.org/x/crypto/ssh"
@@ -33,7 +34,9 @@ import (
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/auth"
+	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
+	"github.com/ollama/ollama/parser"
 	"github.com/ollama/ollama/progress"
 	"github.com/ollama/ollama/server"
 	"github.com/ollama/ollama/types/errtypes"
@@ -62,7 +65,7 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	}
 	defer f.Close()
 
-	modelfile, err := model.ParseFile(f)
+	modelfile, err := parser.ParseFile(f)
 	if err != nil {
 		return err
 	}
@@ -142,9 +145,9 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 		return nil
 	}
 
-	quantization, _ := cmd.Flags().GetString("quantization")
+	quantize, _ := cmd.Flags().GetString("quantize")
 
-	request := api.CreateRequest{Name: args[0], Modelfile: modelfile.String(), Quantization: quantization}
+	request := api.CreateRequest{Name: args[0], Modelfile: modelfile.String(), Quantize: quantize}
 	if err := client.Create(cmd.Context(), &request, fn); err != nil {
 		return err
 	}
@@ -206,7 +209,7 @@ func tempZipFiles(path string) (string, error) {
 		// pytorch files might also be unresolved git lfs references; skip if they are
 		// covers pytorch_model-x-of-y.bin, pytorch_model.fp32-x-of-y.bin, pytorch_model.bin
 		files = append(files, pt...)
-	} else if pt, _ := glob(filepath.Join(path, "consolidated*.pth"), "application/octet-stream"); len(pt) > 0 {
+	} else if pt, _ := glob(filepath.Join(path, "consolidated*.pth"), "application/zip"); len(pt) > 0 {
 		// pytorch files might also be unresolved git lfs references; skip if they are
 		// covers consolidated.x.pth, consolidated.pth
 		files = append(files, pt...)
@@ -324,6 +327,18 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	}
 	opts.Format = format
 
+	keepAlive, err := cmd.Flags().GetString("keepalive")
+	if err != nil {
+		return err
+	}
+	if keepAlive != "" {
+		d, err := time.ParseDuration(keepAlive)
+		if err != nil {
+			return err
+		}
+		opts.KeepAlive = &api.Duration{Duration: d}
+	}
+
 	prompts := args[1:]
 	// prepend stdin to the prompt if provided
 	if !term.IsTerminal(int(os.Stdin.Fd())) {
@@ -496,6 +511,52 @@ func ListHandler(cmd *cobra.Command, args []string) error {
 	return nil
 }
 
+func ListRunningHandler(cmd *cobra.Command, args []string) error {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return err
+	}
+
+	models, err := client.ListRunning(cmd.Context())
+	if err != nil {
+		return err
+	}
+
+	var data [][]string
+
+	for _, m := range models.Models {
+		if len(args) == 0 || strings.HasPrefix(m.Name, args[0]) {
+			var procStr string
+			switch {
+			case m.SizeVRAM == 0:
+				procStr = "100% CPU"
+			case m.SizeVRAM == m.Size:
+				procStr = "100% GPU"
+			case m.SizeVRAM > m.Size || m.Size == 0:
+				procStr = "Unknown"
+			default:
+				sizeCPU := m.Size - m.SizeVRAM
+				cpuPercent := math.Round(float64(sizeCPU) / float64(m.Size) * 100)
+				procStr = fmt.Sprintf("%d%%/%d%% CPU/GPU", int(cpuPercent), int(100-cpuPercent))
+			}
+			data = append(data, []string{m.Name, m.Digest[:12], format.HumanBytes(m.Size), procStr, format.HumanTime(m.ExpiresAt, "Never")})
+		}
+	}
+
+	table := tablewriter.NewWriter(os.Stdout)
+	table.SetHeader([]string{"NAME", "ID", "SIZE", "PROCESSOR", "UNTIL"})
+	table.SetHeaderAlignment(tablewriter.ALIGN_LEFT)
+	table.SetAlignment(tablewriter.ALIGN_LEFT)
+	table.SetHeaderLine(false)
+	table.SetBorder(false)
+	table.SetNoWhiteSpace(true)
+	table.SetTablePadding("\t")
+	table.AppendBulk(data)
+	table.Render()
+
+	return nil
+}
+
 func DeleteHandler(cmd *cobra.Command, args []string) error {
 	client, err := api.ClientFromEnvironment()
 	if err != nil {
@@ -672,6 +733,7 @@ type runOptions struct {
 	Images      []api.ImageData
 	Options     map[string]interface{}
 	MultiModal  bool
+	KeepAlive   *api.Duration
 }
 
 type displayResponseState struct {
@@ -684,7 +746,8 @@ func displayResponse(content string, wordWrap bool, state *displayResponseState)
 	if wordWrap && termWidth >= 10 {
 		for _, ch := range content {
 			if state.lineLength+1 > termWidth-5 {
-				if len(state.wordBuffer) > termWidth-10 {
+
+				if runewidth.StringWidth(state.wordBuffer) > termWidth-10 {
 					fmt.Printf("%s%c", state.wordBuffer, ch)
 					state.wordBuffer = ""
 					state.lineLength = 0
@@ -692,12 +755,18 @@ func displayResponse(content string, wordWrap bool, state *displayResponseState)
 				}
 
 				// backtrack the length of the last word and clear to the end of the line
-				fmt.Printf("\x1b[%dD\x1b[K\n", len(state.wordBuffer))
+				fmt.Printf("\x1b[%dD\x1b[K\n", runewidth.StringWidth(state.wordBuffer))
 				fmt.Printf("%s%c", state.wordBuffer, ch)
-				state.lineLength = len(state.wordBuffer) + 1
+				chWidth := runewidth.RuneWidth(ch)
+
+				state.lineLength = runewidth.StringWidth(state.wordBuffer) + chWidth
 			} else {
 				fmt.Print(string(ch))
-				state.lineLength += 1
+				state.lineLength += runewidth.RuneWidth(ch)
+				if runewidth.RuneWidth(ch) >= 2 {
+					state.wordBuffer = ""
+					continue
+				}
 
 				switch ch {
 				case ' ':
@@ -766,6 +835,10 @@ func chat(cmd *cobra.Command, opts runOptions) (*api.Message, error) {
 		Options:  opts.Options,
 	}
 
+	if opts.KeepAlive != nil {
+		req.KeepAlive = opts.KeepAlive
+	}
+
 	if err := client.Chat(cancelCtx, req, fn); err != nil {
 		if errors.Is(err, context.Canceled) {
 			return nil, nil
@@ -841,14 +914,15 @@ func generate(cmd *cobra.Command, opts runOptions) error {
 	}
 
 	request := api.GenerateRequest{
-		Model:    opts.Model,
-		Prompt:   opts.Prompt,
-		Context:  generateContext,
-		Images:   opts.Images,
-		Format:   opts.Format,
-		System:   opts.System,
-		Template: opts.Template,
-		Options:  opts.Options,
+		Model:     opts.Model,
+		Prompt:    opts.Prompt,
+		Context:   generateContext,
+		Images:    opts.Images,
+		Format:    opts.Format,
+		System:    opts.System,
+		Template:  opts.Template,
+		Options:   opts.Options,
+		KeepAlive: opts.KeepAlive,
 	}
 
 	if err := client.Generate(ctx, &request, fn); err != nil {
@@ -1006,12 +1080,19 @@ func versionHandler(cmd *cobra.Command, _ []string) {
 	}
 }
 
-func appendHostEnvDocs(cmd *cobra.Command) {
-	const hostEnvDocs = `
+func appendEnvDocs(cmd *cobra.Command, envs []envconfig.EnvVar) {
+	if len(envs) == 0 {
+		return
+	}
+
+	envUsage := `
 Environment Variables:
-      OLLAMA_HOST        The host:port or base URL of the Ollama server (e.g. http://localhost:11434)
 `
-	cmd.SetUsageTemplate(cmd.UsageTemplate() + hostEnvDocs)
+	for _, e := range envs {
+		envUsage += fmt.Sprintf("      %-24s   %s\n", e.Name, e.Description)
+	}
+
+	cmd.SetUsageTemplate(cmd.UsageTemplate() + envUsage)
 }
 
 func NewCLI() *cobra.Command {
@@ -1050,8 +1131,8 @@ func NewCLI() *cobra.Command {
 		RunE:    CreateHandler,
 	}
 
-	createCmd.Flags().StringP("file", "f", "Modelfile", "Name of the Modelfile (default \"Modelfile\")")
-	createCmd.Flags().StringP("quantization", "q", "", "Quantization level.")
+	createCmd.Flags().StringP("file", "f", "Modelfile", "Name of the Modelfile")
+	createCmd.Flags().StringP("quantize", "q", "", "Quantize model to this level (e.g. q4_0)")
 
 	showCmd := &cobra.Command{
 		Use:     "show MODEL",
@@ -1075,6 +1156,7 @@ func NewCLI() *cobra.Command {
 		RunE:    RunHandler,
 	}
 
+	runCmd.Flags().String("keepalive", "", "Duration to keep a model loaded (e.g. 5m)")
 	runCmd.Flags().Bool("verbose", false, "Show timings for response")
 	runCmd.Flags().Bool("insecure", false, "Use an insecure registry")
 	runCmd.Flags().Bool("nowordwrap", false, "Don't wrap words to the next line automatically")
@@ -1086,15 +1168,6 @@ func NewCLI() *cobra.Command {
 		Args:    cobra.ExactArgs(0),
 		RunE:    RunServer,
 	}
-	serveCmd.SetUsageTemplate(serveCmd.UsageTemplate() + `
-Environment Variables:
-
-    OLLAMA_HOST         The host:port to bind to (default "127.0.0.1:11434")
-    OLLAMA_ORIGINS      A comma separated list of allowed origins.
-    OLLAMA_MODELS       The path to the models directory (default is "~/.ollama/models")
-    OLLAMA_KEEP_ALIVE   The duration that models stay loaded in memory (default is "5m")
-    OLLAMA_DEBUG        Set to 1 to enable additional debug logging
-`)
 
 	pullCmd := &cobra.Command{
 		Use:     "pull MODEL",
@@ -1123,6 +1196,14 @@ Environment Variables:
 		PreRunE: checkServerHeartbeat,
 		RunE:    ListHandler,
 	}
+
+	psCmd := &cobra.Command{
+		Use:     "ps",
+		Short:   "List running models",
+		PreRunE: checkServerHeartbeat,
+		RunE:    ListRunningHandler,
+	}
+
 	copyCmd := &cobra.Command{
 		Use:     "cp SOURCE DESTINATION",
 		Short:   "Copy a model",
@@ -1139,6 +1220,10 @@ Environment Variables:
 		RunE:    DeleteHandler,
 	}
 
+	envVars := envconfig.AsMap()
+
+	envs := []envconfig.EnvVar{envVars["OLLAMA_HOST"]}
+
 	for _, cmd := range []*cobra.Command{
 		createCmd,
 		showCmd,
@@ -1146,10 +1231,30 @@ Environment Variables:
 		pullCmd,
 		pushCmd,
 		listCmd,
+		psCmd,
 		copyCmd,
 		deleteCmd,
+		serveCmd,
 	} {
-		appendHostEnvDocs(cmd)
+		switch cmd {
+		case runCmd:
+			appendEnvDocs(cmd, []envconfig.EnvVar{envVars["OLLAMA_HOST"], envVars["OLLAMA_NOHISTORY"]})
+		case serveCmd:
+			appendEnvDocs(cmd, []envconfig.EnvVar{
+				envVars["OLLAMA_DEBUG"],
+				envVars["OLLAMA_HOST"],
+				envVars["OLLAMA_KEEP_ALIVE"],
+				envVars["OLLAMA_MAX_LOADED_MODELS"],
+				envVars["OLLAMA_MAX_QUEUE"],
+				envVars["OLLAMA_MODELS"],
+				envVars["OLLAMA_NUM_PARALLEL"],
+				envVars["OLLAMA_NOPRUNE"],
+				envVars["OLLAMA_ORIGINS"],
+				envVars["OLLAMA_TMPDIR"],
+			})
+		default:
+			appendEnvDocs(cmd, envs)
+		}
 	}
 
 	rootCmd.AddCommand(
@@ -1160,6 +1265,7 @@ Environment Variables:
 		pullCmd,
 		pushCmd,
 		listCmd,
+		psCmd,
 		copyCmd,
 		deleteCmd,
 	)

cmd/interactive.go

@@ -15,8 +15,10 @@ import (
 	"golang.org/x/exp/slices"
 
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/progress"
 	"github.com/ollama/ollama/readline"
+	"github.com/ollama/ollama/types/errtypes"
 )
 
 type MultilineState int
@@ -56,6 +58,11 @@ func loadModel(cmd *cobra.Command, opts *runOptions) error {
 		Model:    opts.Model,
 		Messages: []api.Message{},
 	}
+
+	if opts.KeepAlive != nil {
+		chatReq.KeepAlive = opts.KeepAlive
+	}
+
 	err = client.Chat(cmd.Context(), chatReq, func(resp api.ChatResponse) error {
 		p.StopAndClear()
 		if len(opts.Messages) > 0 {
@@ -132,6 +139,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 		fmt.Fprintln(os.Stderr, "   Alt + f            Move forward (right) one word")
 		fmt.Fprintln(os.Stderr, "  Ctrl + k            Delete the sentence after the cursor")
 		fmt.Fprintln(os.Stderr, "  Ctrl + u            Delete the sentence before the cursor")
+		fmt.Fprintln(os.Stderr, "  Ctrl + w            Delete the word before the cursor")
 		fmt.Fprintln(os.Stderr, "")
 		fmt.Fprintln(os.Stderr, "  Ctrl + l            Clear the screen")
 		fmt.Fprintln(os.Stderr, "  Ctrl + c            Stop the model from responding")
@@ -162,7 +170,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 		fmt.Fprintln(os.Stderr, "  /set parameter repeat_penalty <float> How strongly to penalize repetitions")
 		fmt.Fprintln(os.Stderr, "  /set parameter repeat_last_n <int>    Set how far back to look for repetitions")
 		fmt.Fprintln(os.Stderr, "  /set parameter num_gpu <int>          The number of layers to send to the GPU")
-		fmt.Fprintln(os.Stderr, "  /set parameter stop \"<string>\", ...   Set the stop parameters")
+		fmt.Fprintln(os.Stderr, "  /set parameter stop <string> <string> ...   Set the stop parameters")
 		fmt.Fprintln(os.Stderr, "")
 	}
 
@@ -176,6 +184,10 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 		return err
 	}
 
+	if envconfig.NoHistory {
+		scanner.HistoryDisable()
+	}
+
 	fmt.Print(readline.StartBracketedPaste)
 	defer fmt.Printf(readline.EndBracketedPaste)
 
@@ -276,13 +288,20 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 			fn := func(resp api.ProgressResponse) error { return nil }
 			err = client.Create(cmd.Context(), req, fn)
 			if err != nil {
-				fmt.Println("error: couldn't save model")
+				if strings.Contains(err.Error(), errtypes.InvalidModelNameErrMsg) {
+					fmt.Printf("error: The model name '%s' is invalid\n", args[1])
+					continue
+				}
 				return err
 			}
 			fmt.Printf("Created new model '%s'\n", args[1])
 			continue
 		case strings.HasPrefix(line, "/clear"):
 			opts.Messages = []api.Message{}
+			if opts.System != "" {
+				newMessage := api.Message{Role: "system", Content: opts.System}
+				opts.Messages = append(opts.Messages, newMessage)
+			}
 			fmt.Println("Cleared session context")
 			continue
 		case strings.HasPrefix(line, "/set"):

convert/convert.go

@@ -5,6 +5,7 @@ import (
 	"encoding/binary"
 	"encoding/json"
 	"fmt"
+	"io"
 	"log/slog"
 	"os"
 	"path/filepath"
@@ -17,6 +18,16 @@ import (
 	"github.com/ollama/ollama/llm"
 )
 
+const (
+	_ int32 = iota
+	tokenTypeNormal
+	tokenTypeUnknown
+	tokenTypeControl
+	tokenTypeUserDefined
+	tokenTypeUnused
+	tokenTypeByte
+)
+
 type Params struct {
 	Architectures     []string `json:"architectures"`
 	VocabSize         int      `json:"vocab_size"`
@@ -36,6 +47,8 @@ type Params struct {
 	Experts     int `json:"num_local_experts"`
 	ExpertsUsed int `json:"num_experts_per_tok"`
 
+	PreTokenizer string
+
 	ByteOrder
 }
 
@@ -47,7 +60,7 @@ type ByteOrder interface {
 type ModelArch interface {
 	GetTensors() error
 	LoadVocab() error
-	WriteGGUF() (string, error)
+	WriteGGUF(io.WriteSeeker) error
 }
 
 type ModelFormat interface {
@@ -73,10 +86,9 @@ func GetModelFormat(dirname string) (ModelFormat, error) {
 	}
 
 	for _, fn := range files {
-		slog.Debug(fmt.Sprintf("file = %s", fn))
 		if strings.HasSuffix(fn, ".safetensors") {
 			return &SafetensorFormat{}, nil
-		} else if strings.HasSuffix(fn, ".bin") {
+		} else if strings.HasSuffix(fn, ".bin") || strings.HasSuffix(fn, ".pth") {
 			slog.Debug("model is torch")
 			return &TorchFormat{}, nil
 		}
@@ -91,6 +103,7 @@ type Vocab struct {
 	Tokens []string
 	Scores []float32
 	Types  []int32
+	Merges []string
 }
 
 func LoadSentencePieceTokens(dirpath string, params *Params) (*Vocab, error) {
@@ -169,7 +182,7 @@ func LoadSentencePieceTokens(dirpath string, params *Params) (*Vocab, error) {
 		}
 		v.Tokens = append(v.Tokens, t.key)
 		v.Scores = append(v.Scores, -1000.0)
-		v.Types = append(v.Types, int32(llm.GGUFTokenUserDefined))
+		v.Types = append(v.Types, tokenTypeUserDefined)
 	}
 	slog.Info(fmt.Sprintf("vocab size w/ extra tokens: %d", len(v.Tokens)))
 
@@ -179,7 +192,7 @@ func LoadSentencePieceTokens(dirpath string, params *Params) (*Vocab, error) {
 		for cnt := 0; cnt < missingTokens; cnt++ {
 			v.Tokens = append(v.Tokens, fmt.Sprintf("<dummy%05d>", cnt+1))
 			v.Scores = append(v.Scores, -1)
-			v.Types = append(v.Types, int32(llm.GGUFTokenUserDefined))
+			v.Types = append(v.Types, tokenTypeUserDefined)
 		}
 	}
 

convert/convert_test.go

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

convert/gemma.go

@@ -1,14 +1,11 @@
 package convert
 
 import (
-	"encoding/binary"
 	"fmt"
 	"io"
 	"log/slog"
-	"os"
 	"strings"
 
-	"github.com/d4l3k/go-bfloat16"
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
 
@@ -19,49 +16,27 @@ type GemmaModel struct {
 	ModelData
 }
 
-func gemmaLayerHandler(w io.Writer, r safetensorWriterTo, f *os.File) error {
-	slog.Debug(fmt.Sprintf("converting '%s'", r.t.Name))
-
-	data := make([]byte, r.end-r.start)
-	if err := binary.Read(f, r.bo, data); err != nil {
-		return err
-	}
-
-	tDataF32 := bfloat16.DecodeFloat32(data)
-
-	var err error
-	tDataF32, err = addOnes(tDataF32, int(r.t.Shape[0]))
-	if err != nil {
-		return err
-	}
-
-	if err := binary.Write(w, r.bo, tDataF32); err != nil {
-		return err
-	}
-	return nil
-}
-
 func addOnes(data []float32, vectorSize int) ([]float32, error) {
 	n := tensor.New(tensor.WithShape(vectorSize), tensor.WithBacking(data))
 	ones := tensor.Ones(tensor.Float32, vectorSize)
 
-	var err error
-	n, err = n.Add(ones)
+	n, err := n.Add(ones)
 	if err != nil {
-		return []float32{}, err
+		return nil, err
 	}
 
-	newN, err := native.SelectF32(n, 0)
+	ts, err := native.SelectF32(n, 0)
 	if err != nil {
-		return []float32{}, err
+		return nil, err
 	}
 
-	var fullTensor []float32
-	for _, v := range newN {
-		fullTensor = append(fullTensor, v...)
+	var f32s []float32
+	for _, t := range ts {
+		f32s = append(f32s, t...)
 	}
 
-	return fullTensor, nil
+
+	return f32s, nil
 }
 
 func (m *GemmaModel) GetTensors() error {
@@ -71,12 +46,10 @@ func (m *GemmaModel) GetTensors() error {
 	}
 
 	slog.Debug(fmt.Sprintf("Total tensors: %d", len(t)))
-
-	m.Tensors = []llm.Tensor{}
 	for _, l := range t {
 		if strings.HasSuffix(l.Name, "norm.weight") {
 			wt := l.WriterTo.(safetensorWriterTo)
-			wt.handler = gemmaLayerHandler
+			wt.repacker = m.Repack
 			l.WriterTo = wt
 		}
 		m.Tensors = append(m.Tensors, l)
@@ -94,7 +67,11 @@ func (m *GemmaModel) LoadVocab() error {
 	return nil
 }
 
-func (m *GemmaModel) WriteGGUF() (string, error) {
+func (m *GemmaModel) Repack(_ string, data []float32, shape []uint64) ([]float32, error) {
+	return addOnes(data, int(shape[0]))
+}
+
+func (m *GemmaModel) WriteGGUF(ws io.WriteSeeker) error {
 	kv := llm.KV{
 		"general.architecture":                   "gemma",
 		"general.name":                           m.Name,
@@ -122,16 +99,5 @@ 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

@@ -1,18 +1,17 @@
 package convert
 
 import (
-	"encoding/binary"
+	"cmp"
+	"errors"
 	"fmt"
 	"io"
-	"log/slog"
 	"os"
+	"path/filepath"
 	"regexp"
 	"strings"
 
-	"github.com/nlpodyssey/gopickle/pytorch"
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
-	"github.com/x448/float16"
 
 	"github.com/ollama/ollama/llm"
 )
@@ -21,81 +20,12 @@ type LlamaModel struct {
 	ModelData
 }
 
-func llamaLayerHandler(w io.Writer, r torchWriterTo) error {
-	slog.Debug(fmt.Sprintf("repacking layer '%s'", r.t.Name))
-
-	data := r.storage.(*pytorch.HalfStorage).Data
-	tData := make([]uint16, len(data))
-	for cnt, v := range data {
-		tData[cnt] = uint16(float16.Fromfloat32(v))
-	}
-
-	var err error
-	var heads uint32
-	if strings.Contains(r.t.Name, "attn_q") {
-		heads = uint32(r.params.AttentionHeads)
-	} else if strings.Contains(r.t.Name, "attn_k") {
-		heads = uint32(r.params.KeyValHeads)
-		if heads == 0 {
-			heads = uint32(r.params.AttentionHeads)
-		}
-	} else {
-		return fmt.Errorf("unknown layer type")
-	}
-
-	slog.Debug(fmt.Sprintf("heads = %d", heads))
-
-	tData, err = llamaRepack(tData, int(heads), r.t.Shape)
-	if err != nil {
-		return err
-	}
-
-	if err = binary.Write(w, r.bo, tData); err != nil {
-		return err
-	}
-	return nil
-}
-
-func llamaRepack(data []uint16, heads int, shape []uint64) ([]uint16, error) {
-	n := tensor.New(tensor.WithShape(int(shape[0]), int(shape[1])), tensor.WithBacking(data))
-	origShape := n.Shape().Clone()
-
-	// reshape the tensor and swap axes 1 and 2 to unpack the layer for gguf
-	if err := n.Reshape(heads, 2, origShape[0]/heads/2, origShape[1]); err != nil {
-		return nil, err
-	}
-
-	if err := n.T(0, 2, 1, 3); err != nil {
-		return nil, err
-	}
-
-	if err := n.Reshape(origShape...); err != nil {
-		return nil, err
-	}
-
-	if err := n.Transpose(); err != nil {
-		return nil, err
-	}
-	newN, err := native.SelectU16(n, 1)
-	if err != nil {
-		return nil, err
-	}
-
-	var fullTensor []uint16
-	for _, v := range newN {
-		fullTensor = append(fullTensor, v...)
-	}
-	return fullTensor, nil
-}
-
 func (m *LlamaModel) GetTensors() error {
 	t, err := m.Format.GetTensors(m.Path, m.Params)
 	if err != nil {
 		return err
 	}
 
-	m.Tensors = []llm.Tensor{}
-
 	pattern := `^blk\.[0-9]+\.attn_(?P<layer>q|k)\.weight$`
 	re, err := regexp.Compile(pattern)
 	if err != nil {
@@ -105,10 +35,16 @@ func (m *LlamaModel) GetTensors() error {
 	for _, l := range t {
 		matches := re.FindAllStringSubmatch(l.Name, -1)
 		if len(matches) > 0 {
-			slog.Debug(fmt.Sprintf("setting handler for: %s", l.Name))
-			wt := l.WriterTo.(torchWriterTo)
-			wt.handler = llamaLayerHandler
-			l.WriterTo = wt
+			switch m.Format.(type) {
+			case *TorchFormat:
+				wt := l.WriterTo.(torchWriterTo)
+				wt.repacker = m.Repack
+				l.WriterTo = wt
+			case *SafetensorFormat:
+				wt := l.WriterTo.(safetensorWriterTo)
+				wt.repacker = m.Repack
+				l.WriterTo = wt
+			}
 		}
 		m.Tensors = append(m.Tensors, l)
 	}
@@ -116,23 +52,26 @@ func (m *LlamaModel) GetTensors() error {
 	return nil
 }
 
-func (m *LlamaModel) LoadVocab() error {
-	var v *Vocab
-	var err error
-
-	slog.Debug("loading vocab")
-	v, err = LoadSentencePieceTokens(m.Path, m.Params)
-	if err != nil {
+func (m *LlamaModel) LoadVocab() (err error) {
+	pre, ts, merges, err := parseTokens(filepath.Join(m.Path, "tokenizer.json"))
+	if errors.Is(err, os.ErrNotExist) {
+		return nil
+	} else if err != nil {
 		return err
 	}
 
-	slog.Debug("vocab loaded")
+	m.Vocab = &Vocab{}
+	for _, t := range ts {
+		m.Vocab.Tokens = append(m.Vocab.Tokens, t.Content)
+		m.Vocab.Types = append(m.Vocab.Types, t.Type())
+	}
 
-	m.Vocab = v
+	m.Vocab.Merges = merges
+	m.Params.PreTokenizer = pre
 	return nil
 }
 
-func (m *LlamaModel) WriteGGUF() (string, error) {
+func (m *LlamaModel) WriteGGUF(ws io.WriteSeeker) error {
 	kv := llm.KV{
 		"general.architecture":                   "llama",
 		"general.name":                           m.Name,
@@ -141,36 +80,79 @@ func (m *LlamaModel) WriteGGUF() (string, error) {
 		"llama.embedding_length":                 uint32(m.Params.HiddenSize),
 		"llama.block_count":                      uint32(m.Params.HiddenLayers),
 		"llama.feed_forward_length":              uint32(m.Params.IntermediateSize),
+		"llama.rope.freq_base":                   float32(m.Params.RopeFrequencyBase),
 		"llama.rope.dimension_count":             uint32(m.Params.HiddenSize / m.Params.AttentionHeads),
 		"llama.attention.head_count":             uint32(m.Params.AttentionHeads),
 		"llama.attention.head_count_kv":          uint32(m.Params.KeyValHeads),
 		"llama.attention.layer_norm_rms_epsilon": float32(m.Params.NormEPS),
 		"general.file_type":                      uint32(1),
-		"tokenizer.ggml.model":                   "llama",
+		"tokenizer.ggml.model":                   "gpt2",
 
+		"tokenizer.ggml.pre":        m.Params.PreTokenizer,
 		"tokenizer.ggml.tokens":     m.Vocab.Tokens,
-		"tokenizer.ggml.scores":     m.Vocab.Scores,
 		"tokenizer.ggml.token_type": m.Vocab.Types,
 
 		"tokenizer.ggml.bos_token_id":     uint32(m.Params.BoSTokenID),
 		"tokenizer.ggml.eos_token_id":     uint32(m.Params.EoSTokenID),
 		"tokenizer.ggml.unknown_token_id": uint32(0),
-		"tokenizer.ggml.add_bos_token":    true,
-		"tokenizer.ggml.add_eos_token":    false,
 	}
 
-	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
+	if len(m.Vocab.Merges) > 0 {
+		kv["tokenizer.ggml.merges"] = m.Vocab.Merges
+	} else {
+		kv["tokenizer.ggml.scores"] = m.Vocab.Scores
 	}
 
-	slog.Debug(fmt.Sprintf("gguf file = %s", f.Name()))
-
-	return f.Name(), nil
+	return llm.NewGGUFV3(m.Params.ByteOrder).Encode(ws, kv, m.Tensors)
+}
+
+func (m *LlamaModel) Repack(name string, data []float32, shape []uint64) ([]float32, error) {
+	return llamaRepack(name, m.Params, data, shape)
+}
+
+func llamaRepack(name string, params *Params, data []float32, shape []uint64) ([]float32, error) {
+	var dims []int
+	for _, dim := range shape {
+		if dim != 0 {
+			dims = append(dims, int(dim))
+		}
+	}
+
+	var heads int
+	if strings.HasSuffix(name, "attn_q.weight") {
+		heads = params.AttentionHeads
+	} else if strings.HasSuffix(name, "attn_k.weight") {
+		heads = cmp.Or(params.KeyValHeads, params.AttentionHeads)
+	} else {
+		return nil, fmt.Errorf("unknown tensor name: %s", name)
+	}
+
+	n := tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+	if err := n.Reshape(append([]int{heads, 2, dims[0] / heads / 2}, dims[1:]...)...); err != nil {
+		return nil, err
+	}
+
+	if err := n.T(0, 2, 1, 3); err != nil {
+		return nil, err
+	}
+
+	if err := n.Reshape(dims...); err != nil {
+		return nil, err
+	}
+
+	if err := n.Transpose(); err != nil {
+		return nil, err
+	}
+
+	ts, err := native.SelectF32(n, 1)
+	if err != nil {
+		return nil, err
+	}
+
+	var f32s []float32
+	for _, t := range ts {
+		f32s = append(f32s, t...)
+	}
+
+	return f32s, nil
 }

convert/mistral.go

@@ -1,17 +1,8 @@
 package convert
 
 import (
-	"encoding/binary"
-	"fmt"
 	"io"
-	"os"
 	"regexp"
-	"strings"
-
-	"github.com/d4l3k/go-bfloat16"
-	"github.com/pdevine/tensor"
-	"github.com/pdevine/tensor/native"
-	"github.com/x448/float16"
 
 	"github.com/ollama/ollama/llm"
 )
@@ -20,90 +11,12 @@ type MistralModel struct {
 	ModelData
 }
 
-func mistralLayerHandler(w io.Writer, r safetensorWriterTo, f *os.File) error {
-	layerSize := r.end - r.start
-
-	var err error
-	tData := make([]uint16, layerSize/2)
-	if err = binary.Read(f, r.bo, tData); err != nil {
-		return err
-	}
-
-	var heads uint32
-	if strings.Contains(r.t.Name, "attn_q") {
-		heads = uint32(r.params.AttentionHeads)
-	} else if strings.Contains(r.t.Name, "attn_k") {
-		heads = uint32(r.params.KeyValHeads)
-		if heads == 0 {
-			heads = uint32(r.params.AttentionHeads)
-		}
-	} else {
-		return fmt.Errorf("unknown layer type")
-	}
-
-	tData, err = repack(tData, int(heads), r.t.Shape)
-	if err != nil {
-		return err
-	}
-
-	var buf []byte
-	for _, n := range tData {
-		buf = r.bo.AppendUint16(buf, n)
-	}
-
-	tempBuf := make([]uint16, len(tData))
-	tDataF32 := bfloat16.DecodeFloat32(buf)
-	for cnt, v := range tDataF32 {
-		tDataF16 := float16.Fromfloat32(v)
-		tempBuf[cnt] = uint16(tDataF16)
-	}
-
-	if err = binary.Write(w, r.bo, tempBuf); err != nil {
-		return err
-	}
-	return nil
-}
-
-func repack(data []uint16, heads int, shape []uint64) ([]uint16, error) {
-	n := tensor.New(tensor.WithShape(int(shape[0]), int(shape[1])), tensor.WithBacking(data))
-	origShape := n.Shape().Clone()
-
-	// reshape the tensor and swap axes 1 and 2 to unpack the layer for gguf
-	if err := n.Reshape(heads, 2, origShape[0]/heads/2, origShape[1]); err != nil {
-		return nil, err
-	}
-
-	if err := n.T(0, 2, 1, 3); err != nil {
-		return nil, err
-	}
-
-	if err := n.Reshape(origShape...); err != nil {
-		return nil, err
-	}
-
-	if err := n.Transpose(); err != nil {
-		return nil, err
-	}
-	newN, err := native.SelectU16(n, 1)
-	if err != nil {
-		return nil, err
-	}
-
-	var fullTensor []uint16
-	for _, v := range newN {
-		fullTensor = append(fullTensor, v...)
-	}
-	return fullTensor, nil
-}
-
 func (m *MistralModel) GetTensors() error {
 	t, err := m.Format.GetTensors(m.Path, m.Params)
 	if err != nil {
 		return err
 	}
 
-	m.Tensors = []llm.Tensor{}
-
 	pattern := `^blk\.[0-9]+\.attn_(?P<layer>q|k)\.weight$`
 	re, err := regexp.Compile(pattern)
 	if err != nil {
@@ -114,7 +27,7 @@ func (m *MistralModel) GetTensors() error {
 		matches := re.FindAllStringSubmatch(l.Name, -1)
 		if len(matches) > 0 {
 			wt := l.WriterTo.(safetensorWriterTo)
-			wt.handler = mistralLayerHandler
+			wt.repacker = m.Repack
 			l.WriterTo = wt
 		}
 		m.Tensors = append(m.Tensors, l)
@@ -132,7 +45,7 @@ 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 +71,9 @@ 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)
+}
+
+func (m *MistralModel) Repack(name string, data []float32, shape []uint64) ([]float32, error) {
+	return llamaRepack(name, m.Params, data, shape)
 }

convert/mixtral.go

@@ -1,7 +1,7 @@
 package convert
 
 import (
-	"os"
+	"io"
 	"regexp"
 
 	"github.com/ollama/ollama/llm"
@@ -17,8 +17,6 @@ func (m *MixtralModel) GetTensors() error {
 		return err
 	}
 
-	m.Tensors = []llm.Tensor{}
-
 	pattern := `^blk\.[0-9]+\.attn_(?P<layer>q|k)\.weight$`
 	re, err := regexp.Compile(pattern)
 	if err != nil {
@@ -29,7 +27,7 @@ func (m *MixtralModel) GetTensors() error {
 		matches := re.FindAllStringSubmatch(l.Name, -1)
 		if len(matches) > 0 {
 			wt := l.WriterTo.(safetensorWriterTo)
-			wt.handler = mistralLayerHandler
+			wt.repacker = m.Repack
 			l.WriterTo = wt
 		}
 		m.Tensors = append(m.Tensors, l)
@@ -47,7 +45,7 @@ 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 +79,9 @@ 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)
+}
+
+func (m *MixtralModel) Repack(name string, data []float32, shape []uint64) ([]float32, error) {
+	return llamaRepack(name, m.Params, data, shape)
 }

convert/safetensors.go

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

convert/tokenizer.go

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

convert/torch.go

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

docs/README.md

@@ -6,7 +6,7 @@
 * [Importing models](./import.md)
 * [Linux Documentation](./linux.md)
 * [Windows Documentation](./windows.md)
-* [Docker Documentation](https://hub.docker.com/r/ollama/ollama)
+* [Docker Documentation](./docker.md)
 
 ### Reference
 

docs/api.md

@@ -313,7 +313,6 @@ curl http://localhost:11434/api/generate -d '{
     "numa": false,
     "num_ctx": 1024,
     "num_batch": 2,
-    "num_gqa": 1,
     "num_gpu": 1,
     "main_gpu": 0,
     "low_vram": false,
@@ -321,8 +320,6 @@ curl http://localhost:11434/api/generate -d '{
     "vocab_only": false,
     "use_mmap": true,
     "use_mlock": false,
-    "rope_frequency_base": 1.1,
-    "rope_frequency_scale": 0.8,
     "num_thread": 8
   }
 }'
@@ -800,9 +797,9 @@ curl http://localhost:11434/api/show -d '{
 
 ```json
 {
-  "modelfile": "# Modelfile generated by \"ollama show\"\n# To build a new Modelfile based on this one, replace the FROM line with:\n# FROM llava:latest\n\nFROM /Users/matt/.ollama/models/blobs/sha256:200765e1283640ffbd013184bf496e261032fa75b99498a9613be4e94d63ad52\nTEMPLATE \"\"\"{{ .System }}\nUSER: {{ .Prompt }}\nASSSISTANT: \"\"\"\nPARAMETER num_ctx 4096\nPARAMETER stop \"\u003c/s\u003e\"\nPARAMETER stop \"USER:\"\nPARAMETER stop \"ASSSISTANT:\"",
-  "parameters": "num_ctx                        4096\nstop                           \u003c/s\u003e\nstop                           USER:\nstop                           ASSSISTANT:",
-  "template": "{{ .System }}\nUSER: {{ .Prompt }}\nASSSISTANT: ",
+  "modelfile": "# Modelfile generated by \"ollama show\"\n# To build a new Modelfile based on this one, replace the FROM line with:\n# FROM llava:latest\n\nFROM /Users/matt/.ollama/models/blobs/sha256:200765e1283640ffbd013184bf496e261032fa75b99498a9613be4e94d63ad52\nTEMPLATE \"\"\"{{ .System }}\nUSER: {{ .Prompt }}\nASSISTANT: \"\"\"\nPARAMETER num_ctx 4096\nPARAMETER stop \"\u003c/s\u003e\"\nPARAMETER stop \"USER:\"\nPARAMETER stop \"ASSISTANT:\"",
+  "parameters": "num_ctx                        4096\nstop                           \u003c/s\u003e\nstop                           USER:\nstop                           ASSISTANT:",
+  "template": "{{ .System }}\nUSER: {{ .Prompt }}\nASSISTANT: ",
   "details": {
     "format": "gguf",
     "family": "llama",

docs/development.md

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

docs/docker.md

@@ -0,0 +1,71 @@
+# Ollama Docker image
+
+### CPU only
+
+```bash
+docker run -d -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama
+```
+
+### Nvidia GPU
+Install the [NVIDIA Container Toolkit](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html#installation).
+
+#### Install with Apt
+1.  Configure the repository
+```bash
+curl -fsSL https://nvidia.github.io/libnvidia-container/gpgkey \
+    | sudo gpg --dearmor -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg
+curl -s -L https://nvidia.github.io/libnvidia-container/stable/deb/nvidia-container-toolkit.list \
+    | sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' \
+    | sudo tee /etc/apt/sources.list.d/nvidia-container-toolkit.list
+sudo apt-get update
+```
+2.  Install the NVIDIA Container Toolkit packages
+```bash
+sudo apt-get install -y nvidia-container-toolkit
+```
+
+#### Install with Yum or Dnf
+1.  Configure the repository
+    
+```bash
+curl -s -L https://nvidia.github.io/libnvidia-container/stable/rpm/nvidia-container-toolkit.repo \
+    | sudo tee /etc/yum.repos.d/nvidia-container-toolkit.repo
+```
+    
+2. Install the NVIDIA Container Toolkit packages
+    
+```bash
+sudo yum install -y nvidia-container-toolkit
+```
+
+#### Configure Docker to use Nvidia driver 
+```
+sudo nvidia-ctk runtime configure --runtime=docker
+sudo systemctl restart docker
+```
+
+#### Start the container
+
+```bash
+docker run -d --gpus=all -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama
+```
+
+### AMD GPU
+
+To run Ollama using Docker with AMD GPUs, use the `rocm` tag and the following command:
+
+```
+docker run -d --device /dev/kfd --device /dev/dri -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama:rocm
+```
+
+### Run model locally
+
+Now you can run a model:
+
+```
+docker exec -it ollama ollama run llama3
+```
+
+### Try different models
+
+More models can be found on the [Ollama library](https://ollama.com/library).

docs/faq.md

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

docs/troubleshooting.md

@@ -1,85 +1,86 @@
-# How to troubleshoot issues
-
-Sometimes Ollama may not perform as expected. One of the best ways to figure out what happened is to take a look at the logs. Find the logs on **Mac** by running the command:
-
-```shell
-cat ~/.ollama/logs/server.log
-```
-
-On **Linux** systems with systemd, the logs can be found with this command:
-
-```shell
-journalctl -u ollama
-```
-
-When you run Ollama in a **container**, the logs go to stdout/stderr in the container:
-
-```shell
-docker logs <container-name>
-```
-(Use `docker ps` to find the container name)
-
-If manually running `ollama serve` in a terminal, the logs will be on that terminal.
-
-When you run Ollama on **Windows**, there are a few different locations.  You can view them in the explorer window by hitting `<cmd>+R` and type in:
-- `explorer %LOCALAPPDATA%\Ollama` to view logs
-- `explorer %LOCALAPPDATA%\Programs\Ollama` to browse the binaries (The installer adds this to your user PATH)
-- `explorer %HOMEPATH%\.ollama` to browse where models and configuration is stored
-- `explorer %TEMP%` where temporary executable files are stored in one or more `ollama*` directories
-
-To enable additional debug logging to help troubleshoot problems, first **Quit the running app from the tray menu** then in a powershell terminal
-```powershell
-$env:OLLAMA_DEBUG="1"
-& "ollama app.exe"
-```
-
-Join the [Discord](https://discord.gg/ollama) for help interpreting the logs.
-
-## LLM libraries
-
-Ollama includes multiple LLM libraries compiled for different GPUs and CPU
-vector features.  Ollama tries to pick the best one based on the capabilities of
-your system.  If this autodetection has problems, or you run into other problems
-(e.g. crashes in your GPU) you can workaround this by forcing a specific LLM
-library.  `cpu_avx2` will perform the best, followed by `cpu_avx` an the slowest
-but most compatible is `cpu`.  Rosetta emulation under MacOS will work with the
-`cpu` library. 
-
-In the server log, you will see a message that looks something like this (varies
-from release to release):
-
-```
-Dynamic LLM libraries [rocm_v6 cpu cpu_avx cpu_avx2 cuda_v11 rocm_v5]
-```
-
-**Experimental LLM Library Override**
-
-You can set OLLAMA_LLM_LIBRARY to any of the available LLM libraries to bypass
-autodetection, so for example, if you have a CUDA card, but want to force the
-CPU LLM library with AVX2 vector support, use:
-
-```
-OLLAMA_LLM_LIBRARY="cpu_avx2" ollama serve
-```
-
-You can see what features your CPU has with the following.  
-```
-cat /proc/cpuinfo| grep flags  | head -1
-```
-
-## Installing older or pre-release versions on Linux
-
-If you run into problems on Linux and want to install an older version, or you'd
-like to try out a pre-release before it's officially released, you can tell the
-install script which version to install.
-
-```sh
-curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION="0.1.29" sh
-```
-
-## Linux tmp noexec 
-
-If your system is configured with the "noexec" flag where Ollama stores its
-temporary executable files, you can specify an alternate location by setting
-OLLAMA_TMPDIR to a location writable by the user ollama runs as.  For example
-OLLAMA_TMPDIR=/usr/share/ollama/
+# How to troubleshoot issues
+
+Sometimes Ollama may not perform as expected. One of the best ways to figure out what happened is to take a look at the logs. Find the logs on **Mac** by running the command:
+
+```shell
+cat ~/.ollama/logs/server.log
+```
+
+On **Linux** systems with systemd, the logs can be found with this command:
+
+```shell
+journalctl -u ollama
+```
+
+When you run Ollama in a **container**, the logs go to stdout/stderr in the container:
+
+```shell
+docker logs <container-name>
+```
+(Use `docker ps` to find the container name)
+
+If manually running `ollama serve` in a terminal, the logs will be on that terminal.
+
+When you run Ollama on **Windows**, there are a few different locations. You can view them in the explorer window by hitting `<cmd>+R` and type in:
+- `explorer %LOCALAPPDATA%\Ollama` to view logs
+- `explorer %LOCALAPPDATA%\Programs\Ollama` to browse the binaries (The installer adds this to your user PATH)
+- `explorer %HOMEPATH%\.ollama` to browse where models and configuration is stored
+- `explorer %TEMP%` where temporary executable files are stored in one or more `ollama*` directories
+
+To enable additional debug logging to help troubleshoot problems, first **Quit the running app from the tray menu** then in a powershell terminal
+```powershell
+$env:OLLAMA_DEBUG="1"
+& "ollama app.exe"
+```
+
+Join the [Discord](https://discord.gg/ollama) for help interpreting the logs.
+
+## LLM libraries
+
+Ollama includes multiple LLM libraries compiled for different GPUs and CPU vector features. Ollama tries to pick the best one based on the capabilities of your system. If this autodetection has problems, or you run into other problems (e.g. crashes in your GPU) you can workaround this by forcing a specific LLM library. `cpu_avx2` will perform the best, followed by `cpu_avx` an the slowest but most compatible is `cpu`. Rosetta emulation under MacOS will work with the `cpu` library. 
+
+In the server log, you will see a message that looks something like this (varies from release to release):
+
+```
+Dynamic LLM libraries [rocm_v6 cpu cpu_avx cpu_avx2 cuda_v11 rocm_v5]
+```
+
+**Experimental LLM Library Override**
+
+You can set OLLAMA_LLM_LIBRARY to any of the available LLM libraries to bypass autodetection, so for example, if you have a CUDA card, but want to force the CPU LLM library with AVX2 vector support, use:
+
+```
+OLLAMA_LLM_LIBRARY="cpu_avx2" ollama serve
+```
+
+You can see what features your CPU has with the following.
+```
+cat /proc/cpuinfo| grep flags | head -1
+```
+
+## Installing older or pre-release versions on Linux
+
+If you run into problems on Linux and want to install an older version, or you'd like to try out a pre-release before it's officially released, you can tell the install script which version to install.
+
+```sh
+curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION="0.1.29" sh
+```
+
+## Linux tmp noexec 
+
+If your system is configured with the "noexec" flag where Ollama stores its temporary executable files, you can specify an alternate location by setting OLLAMA_TMPDIR to a location writable by the user ollama runs as. For example OLLAMA_TMPDIR=/usr/share/ollama/
+
+## Container fails to run on NVIDIA GPU
+
+Make sure you've set up the container runtime first as described in [docker.md](./docker.md)
+
+Sometimes the container runtime can have difficulties initializing the GPU. When you check the server logs, this can show up as various error codes, such as "3" (not initialized), "46" (device unavailable), "100" (no device), "999" (unknown), or others. The following troubleshooting techniques may help resolve the problem
+
+- Is the uvm driver not loaded? `sudo nvidia-modprobe -u`
+- Try reloading the nvidia_uvm driver - `sudo rmmod nvidia_uvm` then `sudo modprobe nvidia_uvm`
+- Try rebooting
+- Make sure you're running the latest nvidia drivers
+
+If none of those resolve the problem, gather additional information and file an issue:
+- Set `CUDA_ERROR_LEVEL=50` and try again to get more diagnostic logs
+- Check dmesg for any errors `sudo dmesg | grep -i nvrm` and `sudo dmesg | grep -i nvidia`

docs/tutorials/langchainjs.md

@@ -5,13 +5,13 @@ In this tutorial, we are going to use JavaScript with LangChain and Ollama to le
 To get started, let's just use **LangChain** to ask a simple question to a model. To do this with JavaScript, we need to install **LangChain**:
 
 ```bash
-npm install langchain
+npm install @langchain/community
 ```
 
 Now we can start building out our JavaScript:
 
 ```javascript
-import { Ollama } from "langchain/llms/ollama";
+import { Ollama } from "@langchain/community/llms/ollama";
 
 const ollama = new Ollama({
   baseUrl: "http://localhost:11434",

docs/tutorials/langchainpy.md

@@ -12,7 +12,7 @@ So let's figure out how we can use **LangChain** with Ollama to ask our question
 
 Let's start by asking a simple question that we can get an answer to from the **Llama2** model using **Ollama**. First, we need to install the **LangChain** package:
 
-`pip install langchain`
+`pip install langchain_community`
 
 Then we can create a model and ask the question:
 

docs/windows.md

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

envconfig/config.go

@@ -15,6 +15,10 @@ var (
 	AllowOrigins []string
 	// Set via OLLAMA_DEBUG in the environment
 	Debug bool
+	// Experimental flash attention
+	FlashAttention bool
+	// Set via OLLAMA_KEEP_ALIVE in the environment
+	KeepAlive string
 	// Set via OLLAMA_LLM_LIBRARY in the environment
 	LLMLibrary string
 	// Set via OLLAMA_MAX_LOADED_MODELS in the environment
@@ -23,6 +27,8 @@ var (
 	MaxQueuedRequests int
 	// Set via OLLAMA_MAX_VRAM in the environment
 	MaxVRAM uint64
+	// Set via OLLAMA_NOHISTORY in the environment
+	NoHistory bool
 	// Set via OLLAMA_NOPRUNE in the environment
 	NoPrune bool
 	// Set via OLLAMA_NUM_PARALLEL in the environment
@@ -33,21 +39,40 @@ var (
 	TmpDir string
 )
 
-func AsMap() map[string]string {
-	return map[string]string{
-		"OLLAMA_ORIGINS":           fmt.Sprintf("%v", AllowOrigins),
-		"OLLAMA_DEBUG":             fmt.Sprintf("%v", Debug),
-		"OLLAMA_LLM_LIBRARY":       fmt.Sprintf("%v", LLMLibrary),
-		"OLLAMA_MAX_LOADED_MODELS": fmt.Sprintf("%v", MaxRunners),
-		"OLLAMA_MAX_QUEUE":         fmt.Sprintf("%v", MaxQueuedRequests),
-		"OLLAMA_MAX_VRAM":          fmt.Sprintf("%v", MaxVRAM),
-		"OLLAMA_NOPRUNE":           fmt.Sprintf("%v", NoPrune),
-		"OLLAMA_NUM_PARALLEL":      fmt.Sprintf("%v", NumParallel),
-		"OLLAMA_RUNNERS_DIR":       fmt.Sprintf("%v", RunnersDir),
-		"OLLAMA_TMPDIR":            fmt.Sprintf("%v", TmpDir),
+type EnvVar struct {
+	Name        string
+	Value       any
+	Description string
+}
+
+func AsMap() map[string]EnvVar {
+	return map[string]EnvVar{
+		"OLLAMA_DEBUG":             {"OLLAMA_DEBUG", Debug, "Show additional debug information (e.g. OLLAMA_DEBUG=1)"},
+		"OLLAMA_FLASH_ATTENTION":   {"OLLAMA_FLASH_ATTENTION", FlashAttention, "Enabled flash attention"},
+		"OLLAMA_HOST":              {"OLLAMA_HOST", "", "IP Address for the ollama server (default 127.0.0.1:11434)"},
+		"OLLAMA_KEEP_ALIVE":        {"OLLAMA_KEEP_ALIVE", KeepAlive, "The duration that models stay loaded in memory (default \"5m\")"},
+		"OLLAMA_LLM_LIBRARY":       {"OLLAMA_ORIGINS", LLMLibrary, ""},
+		"OLLAMA_MAX_LOADED_MODELS": {"OLLAMA_MAX_LOADED_MODELS", MaxRunners, "Maximum number of loaded models (default 1)"},
+		"OLLAMA_MAX_QUEUE":         {"OLLAMA_MAX_QUEUE", MaxQueuedRequests, "Maximum number of queued requests"},
+		"OLLAMA_MAX_VRAM":          {"OLLAMA_MAX_VRAM", MaxVRAM, ""},
+		"OLLAMA_MODELS":            {"OLLAMA_MODELS", "", "The path to the models directory"},
+		"OLLAMA_NOHISTORY":         {"OLLAMA_NOHISTORY", NoHistory, "Do not preserve readline history"},
+		"OLLAMA_NOPRUNE":           {"OLLAMA_NOPRUNE", NoPrune, "Do not prune model blobs on startup"},
+		"OLLAMA_NUM_PARALLEL":      {"OLLAMA_NUM_PARALLEL", NumParallel, "Maximum number of parallel requests (default 1)"},
+		"OLLAMA_ORIGINS":           {"OLLAMA_ORIGINS", AllowOrigins, "A comma separated list of allowed origins"},
+		"OLLAMA_RUNNERS_DIR":       {"OLLAMA_RUNNERS_DIR", RunnersDir, ""},
+		"OLLAMA_TMPDIR":            {"OLLAMA_TMPDIR", TmpDir, "Location for temporary files"},
 	}
 }
 
+func Values() map[string]string {
+	vals := make(map[string]string)
+	for k, v := range AsMap() {
+		vals[k] = fmt.Sprintf("%v", v.Value)
+	}
+	return vals
+}
+
 var defaultAllowOrigins = []string{
 	"localhost",
 	"127.0.0.1",
@@ -78,6 +103,13 @@ func LoadConfig() {
 		}
 	}
 
+	if fa := clean("OLLAMA_FLASH_ATTENTION"); fa != "" {
+		d, err := strconv.ParseBool(fa)
+		if err == nil {
+			FlashAttention = d
+		}
+	}
+
 	RunnersDir = clean("OLLAMA_RUNNERS_DIR")
 	if runtime.GOOS == "windows" && RunnersDir == "" {
 		// On Windows we do not carry the payloads inside the main executable
@@ -137,6 +169,10 @@ func LoadConfig() {
 		}
 	}
 
+	if nohistory := clean("OLLAMA_NOHISTORY"); nohistory != "" {
+		NoHistory = true
+	}
+
 	if noprune := clean("OLLAMA_NOPRUNE"); noprune != "" {
 		NoPrune = true
 	}
@@ -171,4 +207,6 @@ func LoadConfig() {
 			MaxQueuedRequests = p
 		}
 	}
+
+	KeepAlive = clean("OLLAMA_KEEP_ALIVE")
 }

envconfig/config_test.go

@@ -0,0 +1,23 @@
+package envconfig
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+func TestConfig(t *testing.T) {
+	Debug = false // Reset whatever was loaded in init()
+	t.Setenv("OLLAMA_DEBUG", "")
+	LoadConfig()
+	require.False(t, Debug)
+	t.Setenv("OLLAMA_DEBUG", "false")
+	LoadConfig()
+	require.False(t, Debug)
+	t.Setenv("OLLAMA_DEBUG", "1")
+	LoadConfig()
+	require.True(t, Debug)
+	t.Setenv("OLLAMA_FLASH_ATTENTION", "1")
+	LoadConfig()
+	require.True(t, FlashAttention)
+}

examples/bash-comparemodels/README.md

@@ -1,10 +0,0 @@
-# Bash Shell examples
-
-When calling `ollama`, you can pass it a file to run all the prompts in the file, one after the other:
-
-`ollama run llama3 < sourcequestions.txt`
-
-This concept is used in the following example.
-
-## Compare Models
-`comparemodels.sh` is a script that runs all the questions in `sourcequestions.txt` using any 4 models you choose that you have already pulled from the Ollama library or have created locally.

examples/bash-comparemodels/comparemodels.sh

@@ -1,64 +0,0 @@
-#! /usr/bin/env bash
-# Compare multiple models by running them with the same questions
-
-NUMBEROFCHOICES=4
-SELECTIONS=()
-declare -a SUMS=()
-
-# Get the list of models
-CHOICES=$(ollama list | awk '{print $1}')
-
-# Select which models to run as a comparison
-echo "Select $NUMBEROFCHOICES models to compare:"
-select ITEM in $CHOICES; do
-    if [[ -n $ITEM ]]; then
-        echo "You have selected $ITEM"
-        SELECTIONS+=("$ITEM")
-        ((COUNT++))
-        if [[ $COUNT -eq $NUMBEROFCHOICES ]]; then
-            break
-        fi
-    else
-        echo "Invalid selection"
-    fi
-done
-
-# Loop through each of the selected models
-for ITEM in "${SELECTIONS[@]}"; do
-    echo "--------------------------------------------------------------"
-    echo "Loading the model $ITEM into memory"
-    ollama run "$ITEM" ""
-    echo "--------------------------------------------------------------"
-    echo "Running the questions through the model $ITEM"
-    COMMAND_OUTPUT=$(ollama run "$ITEM" --verbose < sourcequestions.txt 2>&1| tee /dev/stderr)
-
-    # eval duration is sometimes listed in seconds and sometimes in milliseconds. 
-    # Add up the values for each model
-    SUM=$(echo "$COMMAND_OUTPUT" | awk '
-    /eval duration:/ {
-        value = $3
-        if (index(value, "ms") > 0) {
-            gsub("ms", "", value)
-            value /= 1000
-        } else {
-            gsub("s", "", value)
-        }
-        sum += value
-    }
-    END { print sum }')
-
-
-    SUMS+=("All questions for $ITEM completed in $SUM seconds")
-done
-
-echo ""
-echo "--------------------------------------------------------------"
-echo -e "Sums of eval durations for each run:"
-for val in "${SUMS[@]}"; do
-    echo "$val"
-done
-
-echo "--------------------------------------------------------------"
-echo "Comparison complete. Now you can decide"
-echo "which model is best."
-echo "--------------------------------------------------------------"

examples/bash-comparemodels/sourcequestions.txt

@@ -1,7 +0,0 @@
-Why is the sky blue
-What is a black hole
-Explain the big bang theory like I am 5?
-What is the quickest way to win a game of Monopoly with 3 others?
-Why does a vacuum bottle keep my coffee hot and my milkshake cold?
-What is the difference between a meteor, a meteorite, and a meteoroid?
-Create an array with 5 items and print to the console. Do this in Python, C#, Typescript, and Rust.

examples/python-simplechat/client.py

@@ -9,6 +9,7 @@ def chat(messages):
     r = requests.post(
         "http://0.0.0.0:11434/api/chat",
         json={"model": model, "messages": messages, "stream": True},
+	stream=True
     )
     r.raise_for_status()
     output = ""

format/bytes.go

@@ -53,6 +53,8 @@ func HumanBytes(b int64) string {
 
 func HumanBytes2(b uint64) string {
 	switch {
+	case b >= GibiByte:
+		return fmt.Sprintf("%.1f GiB", float64(b)/GibiByte)
 	case b >= MebiByte:
 		return fmt.Sprintf("%.1f MiB", float64(b)/MebiByte)
 	case b >= KibiByte:

format/format.go

@@ -13,12 +13,20 @@ const (
 
 func HumanNumber(b uint64) string {
 	switch {
-	case b > Billion:
-		return fmt.Sprintf("%.0fB", math.Round(float64(b)/Billion))
-	case b > Million:
-		return fmt.Sprintf("%.0fM", math.Round(float64(b)/Million))
-	case b > Thousand:
-		return fmt.Sprintf("%.0fK", math.Round(float64(b)/Thousand))
+	case b >= Billion:
+		number := float64(b) / Billion
+		if number == math.Floor(number) {
+			return fmt.Sprintf("%.0fB", number) // no decimals if whole number
+		}
+		return fmt.Sprintf("%.1fB", number) // one decimal if not a whole number
+	case b >= Million:
+		number := float64(b) / Million
+		if number == math.Floor(number) {
+			return fmt.Sprintf("%.0fM", number) // no decimals if whole number
+		}
+		return fmt.Sprintf("%.2fM", number) // two decimals if not a whole number
+	case b >= Thousand:
+		return fmt.Sprintf("%.0fK", float64(b)/Thousand)
 	default:
 		return fmt.Sprintf("%d", b)
 	}

format/format_test.go

@@ -0,0 +1,34 @@
+package format
+
+import (
+	"testing"
+)
+
+func TestHumanNumber(t *testing.T) {
+
+	type testCase struct {
+		input    uint64
+		expected string
+	}
+
+	testCases := []testCase{
+		{0, "0"},
+		{1000000, "1M"},
+		{125000000, "125M"},
+		{500500000, "500.50M"},
+		{500550000, "500.55M"},
+		{1000000000, "1B"},
+		{2800000000, "2.8B"},
+		{2850000000, "2.9B"},
+		{1000000000000, "1000B"},
+	}
+
+	for _, tc := range testCases {
+		t.Run(tc.expected, func(t *testing.T) {
+			result := HumanNumber(tc.input)
+			if result != tc.expected {
+				t.Errorf("Expected %s, got %s", tc.expected, result)
+			}
+		})
+	}
+}

format/time.go

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

format/time_test.go

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

go.mod

@@ -1,77 +1,76 @@
 module github.com/ollama/ollama
 
-go 1.22
-
-toolchain go1.22.0
+go 1.22.0
 
 require (
 	github.com/containerd/console v1.0.3
-	github.com/d4l3k/go-bfloat16 v0.0.0-20211005043715-690c3bdd05f1
 	github.com/emirpasic/gods v1.18.1
-	github.com/gin-gonic/gin v1.9.1
-	github.com/golang/protobuf v1.5.0 // indirect
-	github.com/google/uuid v1.0.0
-	github.com/mitchellh/mapstructure v1.5.0
+	github.com/gin-gonic/gin v1.10.0
+	github.com/golang/protobuf v1.5.4 // indirect
+	github.com/google/uuid v1.1.2
 	github.com/olekukonko/tablewriter v0.0.5
 	github.com/spf13/cobra v1.7.0
-	github.com/stretchr/testify v1.8.4
+	github.com/stretchr/testify v1.9.0
 	github.com/x448/float16 v0.8.4
 	golang.org/x/sync v0.3.0
 )
 
 require (
+	github.com/d4l3k/go-bfloat16 v0.0.0-20211005043715-690c3bdd05f1
+	github.com/mattn/go-runewidth v0.0.14
 	github.com/nlpodyssey/gopickle v0.3.0
-	github.com/pdevine/tensor v0.0.0-20240228013915-64ccaa8d9ca9
+	github.com/pdevine/tensor v0.0.0-20240510204454-f88f4562727c
 )
 
 require (
-	github.com/apache/arrow/go/arrow v0.0.0-20201229220542-30ce2eb5d4dc // indirect
+	github.com/apache/arrow/go/arrow v0.0.0-20211112161151-bc219186db40 // indirect
+	github.com/bytedance/sonic/loader v0.1.1 // indirect
 	github.com/chewxy/hm v1.0.0 // indirect
-	github.com/chewxy/math32 v1.0.8 // indirect
+	github.com/chewxy/math32 v1.10.1 // indirect
+	github.com/cloudwego/base64x v0.1.4 // indirect
+	github.com/cloudwego/iasm v0.2.0 // indirect
 	github.com/davecgh/go-spew v1.1.1 // indirect
 	github.com/gogo/protobuf v1.3.2 // indirect
-	github.com/google/flatbuffers v1.12.0 // indirect
-	github.com/mattn/go-runewidth v0.0.14 // indirect
+	github.com/google/flatbuffers v24.3.25+incompatible // indirect
+	github.com/kr/text v0.2.0 // indirect
 	github.com/pkg/errors v0.9.1 // indirect
 	github.com/pmezard/go-difflib v1.0.0 // indirect
 	github.com/rivo/uniseg v0.2.0 // indirect
 	github.com/xtgo/set v1.0.0 // indirect
 	go4.org/unsafe/assume-no-moving-gc v0.0.0-20231121144256-b99613f794b6 // indirect
 	golang.org/x/xerrors v0.0.0-20200804184101-5ec99f83aff1 // indirect
-	gonum.org/v1/gonum v0.8.2 // indirect
+	gonum.org/v1/gonum v0.15.0 // indirect
 	gorgonia.org/vecf32 v0.9.0 // indirect
 	gorgonia.org/vecf64 v0.9.0 // indirect
 )
 
 require (
-	github.com/bytedance/sonic v1.9.1 // indirect
-	github.com/chenzhuoyu/base64x v0.0.0-20221115062448-fe3a3abad311 // indirect
-	github.com/gabriel-vasile/mimetype v1.4.2 // indirect
-	github.com/gin-contrib/cors v1.4.0
+	github.com/bytedance/sonic v1.11.6 // indirect
+	github.com/gabriel-vasile/mimetype v1.4.3 // indirect
+	github.com/gin-contrib/cors v1.7.2
 	github.com/gin-contrib/sse v0.1.0 // indirect
 	github.com/go-playground/locales v0.14.1 // indirect
 	github.com/go-playground/universal-translator v0.18.1 // indirect
-	github.com/go-playground/validator/v10 v10.14.0 // indirect
+	github.com/go-playground/validator/v10 v10.20.0 // indirect
 	github.com/goccy/go-json v0.10.2 // indirect
-	github.com/google/go-cmp v0.5.9 // indirect
 	github.com/inconshreveable/mousetrap v1.1.0 // indirect
 	github.com/json-iterator/go v1.1.12 // indirect
-	github.com/klauspost/cpuid/v2 v2.2.4 // indirect
-	github.com/leodido/go-urn v1.2.4 // indirect
-	github.com/mattn/go-isatty v0.0.19 // indirect
+	github.com/klauspost/cpuid/v2 v2.2.7 // indirect
+	github.com/leodido/go-urn v1.4.0 // indirect
+	github.com/mattn/go-isatty v0.0.20 // indirect
 	github.com/modern-go/concurrent v0.0.0-20180306012644-bacd9c7ef1dd // indirect
 	github.com/modern-go/reflect2 v1.0.2 // indirect
-	github.com/pelletier/go-toml/v2 v2.0.8 // indirect
+	github.com/pelletier/go-toml/v2 v2.2.2 // indirect
 	github.com/spf13/pflag v1.0.5 // indirect
 	github.com/twitchyliquid64/golang-asm v0.15.1 // indirect
-	github.com/ugorji/go/codec v1.2.11 // indirect
-	golang.org/x/arch v0.3.0 // indirect
-	golang.org/x/crypto v0.14.0
-	golang.org/x/exp v0.0.0-20230817173708-d852ddb80c63
-	golang.org/x/net v0.17.0 // indirect
-	golang.org/x/sys v0.13.0
-	golang.org/x/term v0.13.0
-	golang.org/x/text v0.14.0 // indirect
-	google.golang.org/protobuf v1.30.0
+	github.com/ugorji/go/codec v1.2.12 // indirect
+	golang.org/x/arch v0.8.0 // indirect
+	golang.org/x/crypto v0.23.0
+	golang.org/x/exp v0.0.0-20231110203233-9a3e6036ecaa
+	golang.org/x/net v0.25.0 // indirect
+	golang.org/x/sys v0.20.0
+	golang.org/x/term v0.20.0
+	golang.org/x/text v0.15.0 // indirect
+	google.golang.org/protobuf v1.34.1
 	gopkg.in/yaml.v3 v3.0.1 // indirect
 )

go.sum

@@ -1,22 +1,32 @@
 cloud.google.com/go v0.26.0/go.mod h1:aQUYkXzVsufM+DwF1aE+0xfcU+56JwCaLick0ClmMTw=
+cloud.google.com/go v0.34.0/go.mod h1:aQUYkXzVsufM+DwF1aE+0xfcU+56JwCaLick0ClmMTw=
+dmitri.shuralyov.com/gpu/mtl v0.0.0-20190408044501-666a987793e9/go.mod h1:H6x//7gZCb22OMCxBHrMx7a5I7Hp++hsVxbQ4BYO7hU=
+gioui.org v0.0.0-20210308172011-57750fc8a0a6/go.mod h1:RSH6KIUZ0p2xy5zHDxgAM4zumjgTw83q2ge/PI+yyw8=
 github.com/BurntSushi/toml v0.3.1/go.mod h1:xHWCNGjB5oqiDr8zfno3MHue2Ht5sIBksp03qcyfWMU=
+github.com/BurntSushi/xgb v0.0.0-20160522181843-27f122750802/go.mod h1:IVnqGOEym/WlBOVXweHU+Q+/VP0lqqI8lqeDx9IjBqo=
 github.com/ajstarks/svgo v0.0.0-20180226025133-644b8db467af/go.mod h1:K08gAheRH3/J6wwsYMMT4xOr94bZjxIelGM0+d/wbFw=
-github.com/apache/arrow/go/arrow v0.0.0-20201229220542-30ce2eb5d4dc h1:zvQ6w7KwtQWgMQiewOF9tFtundRMVZFSAksNV6ogzuY=
-github.com/apache/arrow/go/arrow v0.0.0-20201229220542-30ce2eb5d4dc/go.mod h1:c9sxoIT3YgLxH4UhLOCKaBlEojuMhVYpk4Ntv3opUTQ=
-github.com/bytedance/sonic v1.5.0/go.mod h1:ED5hyg4y6t3/9Ku1R6dU/4KyJ48DZ4jPhfY1O2AihPM=
-github.com/bytedance/sonic v1.9.1 h1:6iJ6NqdoxCDr6mbY8h18oSO+cShGSMRGCEo7F2h0x8s=
-github.com/bytedance/sonic v1.9.1/go.mod h1:i736AoUSYt75HyZLoJW9ERYxcy6eaN6h4BZXU064P/U=
+github.com/antihax/optional v1.0.0/go.mod h1:uupD/76wgC+ih3iEmQUL+0Ugr19nfwCT1kdvxnR2qWY=
+github.com/apache/arrow/go/arrow v0.0.0-20211112161151-bc219186db40 h1:q4dksr6ICHXqG5hm0ZW5IHyeEJXoIJSOZeBLmWPNeIQ=
+github.com/apache/arrow/go/arrow v0.0.0-20211112161151-bc219186db40/go.mod h1:Q7yQnSMnLvcXlZ8RV+jwz/6y1rQTqbX6C82SndT52Zs=
+github.com/boombuler/barcode v1.0.0/go.mod h1:paBWMcWSl3LHKBqUq+rly7CNSldXjb2rDl3JlRe0mD8=
+github.com/bytedance/sonic v1.11.6 h1:oUp34TzMlL+OY1OUWxHqsdkgC/Zfc85zGqw9siXjrc0=
+github.com/bytedance/sonic v1.11.6/go.mod h1:LysEHSvpvDySVdC2f87zGWf6CIKJcAvqab1ZaiQtds4=
+github.com/bytedance/sonic/loader v0.1.1 h1:c+e5Pt1k/cy5wMveRDyk2X4B9hF4g7an8N3zCYjJFNM=
+github.com/bytedance/sonic/loader v0.1.1/go.mod h1:ncP89zfokxS5LZrJxl5z0UJcsk4M4yY2JpfqGeCtNLU=
 github.com/census-instrumentation/opencensus-proto v0.2.1/go.mod h1:f6KPmirojxKA12rnyqOA5BBL4O983OfeGPqjHWSTneU=
-github.com/chenzhuoyu/base64x v0.0.0-20211019084208-fb5309c8db06/go.mod h1:DH46F32mSOjUmXrMHnKwZdA8wcEefY7UVqBKYGjpdQY=
-github.com/chenzhuoyu/base64x v0.0.0-20221115062448-fe3a3abad311 h1:qSGYFH7+jGhDF8vLC+iwCD4WpbV1EBDSzWkJODFLams=
-github.com/chenzhuoyu/base64x v0.0.0-20221115062448-fe3a3abad311/go.mod h1:b583jCggY9gE99b6G5LEC39OIiVsWj+R97kbl5odCEk=
 github.com/chewxy/hm v1.0.0 h1:zy/TSv3LV2nD3dwUEQL2VhXeoXbb9QkpmdRAVUFiA6k=
 github.com/chewxy/hm v1.0.0/go.mod h1:qg9YI4q6Fkj/whwHR1D+bOGeF7SniIP40VweVepLjg0=
 github.com/chewxy/math32 v1.0.0/go.mod h1:Miac6hA1ohdDUTagnvJy/q+aNnEk16qWUdb8ZVhvCN0=
-github.com/chewxy/math32 v1.0.8 h1:fU5E4Ec4Z+5RtRAi3TovSxUjQPkgRh+HbP7tKB2OFbM=
-github.com/chewxy/math32 v1.0.8/go.mod h1:dOB2rcuFrCn6UHrze36WSLVPKtzPMRAQvBvUwkSsLqs=
+github.com/chewxy/math32 v1.10.1 h1:LFpeY0SLJXeaiej/eIp2L40VYfscTvKh/FSEZ68uMkU=
+github.com/chewxy/math32 v1.10.1/go.mod h1:dOB2rcuFrCn6UHrze36WSLVPKtzPMRAQvBvUwkSsLqs=
 github.com/client9/misspell v0.3.4/go.mod h1:qj6jICC3Q7zFZvVWo7KLAzC3yx5G7kyvSDkc90ppPyw=
+github.com/cloudwego/base64x v0.1.4 h1:jwCgWpFanWmN8xoIUHa2rtzmkd5J2plF/dnLS6Xd/0Y=
+github.com/cloudwego/base64x v0.1.4/go.mod h1:0zlkT4Wn5C6NdauXdJRhSKRlJvmclQ1hhJgA0rcu/8w=
+github.com/cloudwego/iasm v0.2.0 h1:1KNIy1I1H9hNNFEEH3DVnI4UujN+1zjpuk6gwHLTssg=
+github.com/cloudwego/iasm v0.2.0/go.mod h1:8rXZaNYT2n95jn+zTI1sDr+IgcD2GVs0nlbbQPiEFhY=
 github.com/cncf/udpa/go v0.0.0-20191209042840-269d4d468f6f/go.mod h1:M8M6+tZqaGXZJjfX53e64911xZQV5JYwmTeXPW+k8Sc=
+github.com/cncf/udpa/go v0.0.0-20201120205902-5459f2c99403/go.mod h1:WmhPx2Nbnhtbo57+VJT5O0JRkEi1Wbu0z5j0R8u5Hbk=
+github.com/cncf/xds/go v0.0.0-20210312221358-fbca930ec8ed/go.mod h1:eXthEFrGJvWHgFFCl3hGmgk+/aYT6PnTQLykKQRLhEs=
 github.com/containerd/console v1.0.3 h1:lIr7SlA5PxZyMV30bDW0MGbiOPXwc63yRuCP0ARubLw=
 github.com/containerd/console v1.0.3/go.mod h1:7LqA/THxQ86k76b8c/EMSiaJ3h1eZkMkXar0TQ1gf3U=
 github.com/cpuguy83/go-md2man/v2 v2.0.2/go.mod h1:tgQtvFlXSQOSOSIRvRPT7W67SCa46tRHOmNcaadrF8o=
@@ -31,30 +41,35 @@ github.com/emirpasic/gods v1.18.1/go.mod h1:8tpGGwCnJ5H4r6BWwaV6OrWmMoPhUl5jm/FM
 github.com/envoyproxy/go-control-plane v0.9.0/go.mod h1:YTl/9mNaCwkRvm6d1a2C3ymFceY/DCBVvsKhRF0iEA4=
 github.com/envoyproxy/go-control-plane v0.9.1-0.20191026205805-5f8ba28d4473/go.mod h1:YTl/9mNaCwkRvm6d1a2C3ymFceY/DCBVvsKhRF0iEA4=
 github.com/envoyproxy/go-control-plane v0.9.4/go.mod h1:6rpuAdCZL397s3pYoYcLgu1mIlRU8Am5FuJP05cCM98=
+github.com/envoyproxy/go-control-plane v0.9.9-0.20201210154907-fd9021fe5dad/go.mod h1:cXg6YxExXjJnVBQHBLXeUAgxn2UodCpnH306RInaBQk=
+github.com/envoyproxy/go-control-plane v0.9.9-0.20210217033140-668b12f5399d/go.mod h1:cXg6YxExXjJnVBQHBLXeUAgxn2UodCpnH306RInaBQk=
+github.com/envoyproxy/go-control-plane v0.9.9-0.20210512163311-63b5d3c536b0/go.mod h1:hliV/p42l8fGbc6Y9bQ70uLwIvmJyVE5k4iMKlh8wCQ=
 github.com/envoyproxy/protoc-gen-validate v0.1.0/go.mod h1:iSmxcyjqTsJpI2R4NaDN7+kN2VEUnK/pcBlmesArF7c=
 github.com/fogleman/gg v1.2.1-0.20190220221249-0403632d5b90/go.mod h1:R/bRT+9gY/C5z7JzPU0zXsXHKM4/ayA+zqcVNZzPa1k=
-github.com/gabriel-vasile/mimetype v1.4.2 h1:w5qFW6JKBz9Y393Y4q372O9A7cUSequkh1Q7OhCmWKU=
-github.com/gabriel-vasile/mimetype v1.4.2/go.mod h1:zApsH/mKG4w07erKIaJPFiX0Tsq9BFQgN3qGY5GnNgA=
-github.com/gin-contrib/cors v1.4.0 h1:oJ6gwtUl3lqV0WEIwM/LxPF1QZ5qe2lGWdY2+bz7y0g=
-github.com/gin-contrib/cors v1.4.0/go.mod h1:bs9pNM0x/UsmHPBWT2xZz9ROh8xYjYkiURUfmBoMlcs=
+github.com/fogleman/gg v1.3.0/go.mod h1:R/bRT+9gY/C5z7JzPU0zXsXHKM4/ayA+zqcVNZzPa1k=
+github.com/gabriel-vasile/mimetype v1.4.3 h1:in2uUcidCuFcDKtdcBxlR0rJ1+fsokWf+uqxgUFjbI0=
+github.com/gabriel-vasile/mimetype v1.4.3/go.mod h1:d8uq/6HKRL6CGdk+aubisF/M5GcPfT7nKyLpA0lbSSk=
+github.com/ghodss/yaml v1.0.0/go.mod h1:4dBDuWmgqj2HViK6kFavaiC9ZROes6MMH2rRYeMEF04=
+github.com/gin-contrib/cors v1.7.2 h1:oLDHxdg8W/XDoN/8zamqk/Drgt4oVZDvaV0YmvVICQw=
+github.com/gin-contrib/cors v1.7.2/go.mod h1:SUJVARKgQ40dmrzgXEVxj2m7Ig1v1qIboQkPDTQ9t2E=
 github.com/gin-contrib/sse v0.1.0 h1:Y/yl/+YNO8GZSjAhjMsSuLt29uWRFHdHYUb5lYOV9qE=
 github.com/gin-contrib/sse v0.1.0/go.mod h1:RHrZQHXnP2xjPF+u1gW/2HnVO7nvIa9PG3Gm+fLHvGI=
-github.com/gin-gonic/gin v1.8.1/go.mod h1:ji8BvRH1azfM+SYow9zQ6SZMvR8qOMZHmsCuWR9tTTk=
-github.com/gin-gonic/gin v1.9.1 h1:4idEAncQnU5cB7BeOkPtxjfCSye0AAm1R0RVIqJ+Jmg=
-github.com/gin-gonic/gin v1.9.1/go.mod h1:hPrL7YrpYKXt5YId3A/Tnip5kqbEAP+KLuI3SUcPTeU=
-github.com/go-playground/assert/v2 v2.0.1/go.mod h1:VDjEfimB/XKnb+ZQfWdccd7VUvScMdVu0Titje2rxJ4=
+github.com/gin-gonic/gin v1.10.0 h1:nTuyha1TYqgedzytsKYqna+DfLos46nTv2ygFy86HFU=
+github.com/gin-gonic/gin v1.10.0/go.mod h1:4PMNQiOhvDRa013RKVbsiNwoyezlm2rm0uX/T7kzp5Y=
+github.com/go-fonts/dejavu v0.1.0/go.mod h1:4Wt4I4OU2Nq9asgDCteaAaWZOV24E+0/Pwo0gppep4g=
+github.com/go-fonts/latin-modern v0.2.0/go.mod h1:rQVLdDMK+mK1xscDwsqM5J8U2jrRa3T0ecnM9pNujks=
+github.com/go-fonts/liberation v0.1.1/go.mod h1:K6qoJYypsmfVjWg8KOVDQhLc8UDgIK2HYqyqAO9z7GY=
+github.com/go-fonts/stix v0.1.0/go.mod h1:w/c1f0ldAUlJmLBvlbkvVXLAD+tAMqobIIQpmnUIzUY=
+github.com/go-gl/glfw v0.0.0-20190409004039-e6da0acd62b1/go.mod h1:vR7hzQXu2zJy9AVAgeJqvqgH9Q5CA+iKCZ2gyEVpxRU=
+github.com/go-latex/latex v0.0.0-20210118124228-b3d85cf34e07/go.mod h1:CO1AlKB2CSIqUrmQPqA0gdRIlnLEY0gK5JGjh37zN5U=
 github.com/go-playground/assert/v2 v2.2.0 h1:JvknZsQTYeFEAhQwI4qEt9cyV5ONwRHC+lYKSsYSR8s=
 github.com/go-playground/assert/v2 v2.2.0/go.mod h1:VDjEfimB/XKnb+ZQfWdccd7VUvScMdVu0Titje2rxJ4=
-github.com/go-playground/locales v0.14.0/go.mod h1:sawfccIbzZTqEDETgFXqTho0QybSa7l++s0DH+LDiLs=
 github.com/go-playground/locales v0.14.1 h1:EWaQ/wswjilfKLTECiXz7Rh+3BjFhfDFKv/oXslEjJA=
 github.com/go-playground/locales v0.14.1/go.mod h1:hxrqLVvrK65+Rwrd5Fc6F2O76J/NuW9t0sjnWqG1slY=
-github.com/go-playground/universal-translator v0.18.0/go.mod h1:UvRDBj+xPUEGrFYl+lu/H90nyDXpg0fqeB/AQUGNTVA=
 github.com/go-playground/universal-translator v0.18.1 h1:Bcnm0ZwsGyWbCzImXv+pAJnYK9S473LQFuzCbDbfSFY=
 github.com/go-playground/universal-translator v0.18.1/go.mod h1:xekY+UJKNuX9WP91TpwSH2VMlDf28Uj24BCp08ZFTUY=
-github.com/go-playground/validator/v10 v10.10.0/go.mod h1:74x4gJWsvQexRdW8Pn3dXSGrTK4nAUsbPlLADvpJkos=
-github.com/go-playground/validator/v10 v10.14.0 h1:vgvQWe3XCz3gIeFDm/HnTIbj6UGmg/+t63MyGU2n5js=
-github.com/go-playground/validator/v10 v10.14.0/go.mod h1:9iXMNT7sEkjXb0I+enO7QXmzG6QCsPWY4zveKFVRSyU=
-github.com/goccy/go-json v0.9.7/go.mod h1:6MelG93GURQebXPDq3khkgXZkazVtN9CRI+MGFi0w8I=
+github.com/go-playground/validator/v10 v10.20.0 h1:K9ISHbSaI0lyB2eWMPJo+kOS/FBExVwjEviJTixqxL8=
+github.com/go-playground/validator/v10 v10.20.0/go.mod h1:dbuPbCMFw/DrkbEynArYaCwl3amGuJotoKCe95atGMM=
 github.com/goccy/go-json v0.10.2 h1:CrxCmQqYDkv1z7lO7Wbh2HN93uovUHgrECaO5ZrCXAU=
 github.com/goccy/go-json v0.10.2/go.mod h1:6MelG93GURQebXPDq3khkgXZkazVtN9CRI+MGFi0w8I=
 github.com/gogo/protobuf v1.3.2 h1:Ov1cvc58UF3b5XjBnZv7+opcTcQFZebYjWzi34vdm4Q=
@@ -72,51 +87,54 @@ github.com/golang/protobuf v1.4.0-rc.4.0.20200313231945-b860323f09d0/go.mod h1:W
 github.com/golang/protobuf v1.4.0/go.mod h1:jodUvKwWbYaEsadDk5Fwe5c77LiNKVO9IDvqG2KuDX0=
 github.com/golang/protobuf v1.4.1/go.mod h1:U8fpvMrcmy5pZrNK1lt4xCsGvpyWQ/VVv6QDs8UjoX8=
 github.com/golang/protobuf v1.4.2/go.mod h1:oDoupMAO8OvCJWAcko0GGGIgR6R6ocIYbsSw735rRwI=
-github.com/golang/protobuf v1.5.0 h1:LUVKkCeviFUMKqHa4tXIIij/lbhnMbP7Fn5wKdKkRh4=
+github.com/golang/protobuf v1.4.3/go.mod h1:oDoupMAO8OvCJWAcko0GGGIgR6R6ocIYbsSw735rRwI=
 github.com/golang/protobuf v1.5.0/go.mod h1:FsONVRAS9T7sI+LIUmWTfcYkHO4aIWwzhcaSAoJOfIk=
-github.com/google/flatbuffers v1.11.0/go.mod h1:1AeVuKshWv4vARoZatz6mlQ0JxURH0Kv5+zNeJKJCa8=
-github.com/google/flatbuffers v1.12.0 h1:/PtAHvnBY4Kqnx/xCQ3OIV9uYcSFGScBsWI3Oogeh6w=
-github.com/google/flatbuffers v1.12.0/go.mod h1:1AeVuKshWv4vARoZatz6mlQ0JxURH0Kv5+zNeJKJCa8=
+github.com/golang/protobuf v1.5.2/go.mod h1:XVQd3VNwM+JqD3oG2Ue2ip4fOMUkwXdXDdiuN0vRsmY=
+github.com/golang/protobuf v1.5.4 h1:i7eJL8qZTpSEXOPTxNKhASYpMn+8e5Q6AdndVa1dWek=
+github.com/golang/protobuf v1.5.4/go.mod h1:lnTiLA8Wa4RWRcIUkrtSVa5nRhsEGBg48fD6rSs7xps=
+github.com/golang/snappy v0.0.3 h1:fHPg5GQYlCeLIPB9BZqMVR5nR9A+IM5zcgeTdjMYmLA=
+github.com/golang/snappy v0.0.3/go.mod h1:/XxbfmMg8lxefKM7IXC3fBNl/7bRcc72aCRzEWrmP2Q=
+github.com/google/flatbuffers v2.0.0+incompatible/go.mod h1:1AeVuKshWv4vARoZatz6mlQ0JxURH0Kv5+zNeJKJCa8=
+github.com/google/flatbuffers v24.3.25+incompatible h1:CX395cjN9Kke9mmalRoL3d81AtFUxJM+yDthflgJGkI=
+github.com/google/flatbuffers v24.3.25+incompatible/go.mod h1:1AeVuKshWv4vARoZatz6mlQ0JxURH0Kv5+zNeJKJCa8=
 github.com/google/go-cmp v0.2.0/go.mod h1:oXzfMopK8JAjlY9xF4vHSVASa0yLyX7SntLO5aqRK0M=
 github.com/google/go-cmp v0.3.0/go.mod h1:8QqcDgzrUqlUb/G2PQTWiueGozuR1884gddMywk6iLU=
 github.com/google/go-cmp v0.3.1/go.mod h1:8QqcDgzrUqlUb/G2PQTWiueGozuR1884gddMywk6iLU=
 github.com/google/go-cmp v0.4.0/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/gNBxE=
 github.com/google/go-cmp v0.5.0/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/gNBxE=
 github.com/google/go-cmp v0.5.5/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/gNBxE=
-github.com/google/go-cmp v0.5.9 h1:O2Tfq5qg4qc4AmwVlvv0oLiVAGB7enBSJ2x2DqQFi38=
-github.com/google/go-cmp v0.5.9/go.mod h1:17dUlkBOakJ0+DkrSSNjCkIjxS6bF9zb3elmeNGIjoY=
+github.com/google/go-cmp v0.5.6/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/gNBxE=
+github.com/google/go-cmp v0.6.0 h1:ofyhxvXcZhMsU5ulbFiLKl/XBFqE1GSq7atu8tAmTRI=
+github.com/google/go-cmp v0.6.0/go.mod h1:17dUlkBOakJ0+DkrSSNjCkIjxS6bF9zb3elmeNGIjoY=
 github.com/google/gofuzz v1.0.0/go.mod h1:dBl0BpW6vV/+mYPU4Po3pmUjxk6FQPldtuIdl/M65Eg=
-github.com/google/uuid v1.0.0 h1:b4Gk+7WdP/d3HZH8EJsZpvV7EtDOgaZLtnaNGIu1adA=
-github.com/google/uuid v1.0.0/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+yHo=
+github.com/google/uuid v1.1.2 h1:EVhdT+1Kseyi1/pUmXKaFxYsDNy9RQYkMWRH68J/W7Y=
+github.com/google/uuid v1.1.2/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+yHo=
+github.com/grpc-ecosystem/grpc-gateway v1.16.0/go.mod h1:BDjrQk3hbvj6Nolgz8mAMFbcEtjT1g+wF4CSlocrBnw=
 github.com/inconshreveable/mousetrap v1.1.0 h1:wN+x4NVGpMsO7ErUn/mUI3vEoE6Jt13X2s0bqwp9tc8=
 github.com/inconshreveable/mousetrap v1.1.0/go.mod h1:vpF70FUmC8bwa3OWnCshd2FqLfsEA9PFc4w1p2J65bw=
 github.com/json-iterator/go v1.1.12 h1:PV8peI4a0ysnczrg+LtxykD8LfKY9ML6u2jnxaEnrnM=
 github.com/json-iterator/go v1.1.12/go.mod h1:e30LSqwooZae/UwlEbR2852Gd8hjQvJoHmT4TnhNGBo=
+github.com/jung-kurt/gofpdf v1.0.0/go.mod h1:7Id9E/uU8ce6rXgefFLlgrJj/GYY22cpxn+r32jIOes=
 github.com/jung-kurt/gofpdf v1.0.3-0.20190309125859-24315acbbda5/go.mod h1:7Id9E/uU8ce6rXgefFLlgrJj/GYY22cpxn+r32jIOes=
 github.com/kisielk/errcheck v1.5.0/go.mod h1:pFxgyoBC7bSaBwPgfKdkLd5X25qrDl4LWUI2bnpBCr8=
 github.com/kisielk/gotool v1.0.0/go.mod h1:XhKaO+MFFWcvkIS/tQcRk01m1F5IRFswLeQ+oQHNcck=
+github.com/klauspost/compress v1.13.1 h1:wXr2uRxZTJXHLly6qhJabee5JqIhTRoLBhDOA74hDEQ=
+github.com/klauspost/compress v1.13.1/go.mod h1:8dP1Hq4DHOhN9w426knH3Rhby4rFm6D8eO+e+Dq5Gzg=
 github.com/klauspost/cpuid/v2 v2.0.9/go.mod h1:FInQzS24/EEf25PyTYn52gqo7WaD8xa0213Md/qVLRg=
-github.com/klauspost/cpuid/v2 v2.2.4 h1:acbojRNwl3o09bUq+yDCtZFc1aiwaAAxtcn8YkZXnvk=
-github.com/klauspost/cpuid/v2 v2.2.4/go.mod h1:RVVoqg1df56z8g3pUjL/3lE5UfnlrJX8tyFgg4nqhuY=
-github.com/kr/pretty v0.1.0/go.mod h1:dAy3ld7l9f0ibDNOQOHHMYYIIbhfbHSm3C4ZsoJORNo=
-github.com/kr/pretty v0.2.1/go.mod h1:ipq/a2n7PKx3OHsz4KJII5eveXtPO4qwEXGdVfWzfnI=
+github.com/klauspost/cpuid/v2 v2.2.7 h1:ZWSB3igEs+d0qvnxR/ZBzXVmxkgt8DdzP6m9pfuVLDM=
+github.com/klauspost/cpuid/v2 v2.2.7/go.mod h1:Lcz8mBdAVJIBVzewtcLocK12l3Y+JytZYpaMropDUws=
+github.com/knz/go-libedit v1.10.1/go.mod h1:MZTVkCWyz0oBc7JOWP3wNAzd002ZbM/5hgShxwh4x8M=
 github.com/kr/pretty v0.3.0 h1:WgNl7dwNpEZ6jJ9k1snq4pZsg7DOEN8hP9Xw0Tsjwk0=
 github.com/kr/pretty v0.3.0/go.mod h1:640gp4NfQd8pI5XOwp5fnNeVWj67G7CFk/SaSQn7NBk=
-github.com/kr/pty v1.1.1/go.mod h1:pFQYn66WHrOpPYNljwOMqo10TkYh1fy3cYio2l3bCsQ=
-github.com/kr/text v0.1.0/go.mod h1:4Jbv+DJW3UT/LiOwJeYQe1efqtUx/iVham/4vfdArNI=
 github.com/kr/text v0.2.0 h1:5Nx0Ya0ZqY2ygV366QzturHI13Jq95ApcVaJBhpS+AY=
 github.com/kr/text v0.2.0/go.mod h1:eLer722TekiGuMkidMxC/pM04lWEeraHUUmBw8l2grE=
-github.com/leodido/go-urn v1.2.1/go.mod h1:zt4jvISO2HfUBqxjfIshjdMTYS56ZS/qv49ictyFfxY=
-github.com/leodido/go-urn v1.2.4 h1:XlAE/cm/ms7TE/VMVoduSpNBoyc2dOxHs5MZSwAN63Q=
-github.com/leodido/go-urn v1.2.4/go.mod h1:7ZrI8mTSeBSHl/UaRyKQW1qZeMgak41ANeCNaVckg+4=
-github.com/mattn/go-isatty v0.0.14/go.mod h1:7GGIvUiUoEMVVmxf/4nioHXj79iQHKdU27kJ6hsGG94=
-github.com/mattn/go-isatty v0.0.19 h1:JITubQf0MOLdlGRuRq+jtsDlekdYPia9ZFsB8h/APPA=
-github.com/mattn/go-isatty v0.0.19/go.mod h1:W+V8PltTTMOvKvAeJH7IuucS94S2C6jfK/D7dTCTo3Y=
+github.com/leodido/go-urn v1.4.0 h1:WT9HwE9SGECu3lg4d/dIA+jxlljEa1/ffXKmRjqdmIQ=
+github.com/leodido/go-urn v1.4.0/go.mod h1:bvxc+MVxLKB4z00jd1z+Dvzr47oO32F/QSNjSBOlFxI=
+github.com/mattn/go-isatty v0.0.20 h1:xfD0iDuEKnDkl03q4limB+vH+GxLEtL/jb4xVJSWWEY=
+github.com/mattn/go-isatty v0.0.20/go.mod h1:W+V8PltTTMOvKvAeJH7IuucS94S2C6jfK/D7dTCTo3Y=
 github.com/mattn/go-runewidth v0.0.9/go.mod h1:H031xJmbD/WCDINGzjvQ9THkh0rPKHF+m2gUSrubnMI=
 github.com/mattn/go-runewidth v0.0.14 h1:+xnbZSEeDbOIg5/mE6JF0w6n9duR1l3/WmbinWVwUuU=
 github.com/mattn/go-runewidth v0.0.14/go.mod h1:Jdepj2loyihRzMpdS35Xk/zdY8IAYHsh153qUoGf23w=
-github.com/mitchellh/mapstructure v1.5.0 h1:jeMsZIYE/09sWLaz43PL7Gy6RuMjD2eJVyuac5Z2hdY=
-github.com/mitchellh/mapstructure v1.5.0/go.mod h1:bFUtVrKA4DC2yAKiSyO/QUcy7e+RRV2QTWOzhPopBRo=
 github.com/modern-go/concurrent v0.0.0-20180228061459-e0a39a4cb421/go.mod h1:6dJC0mAP4ikYIbvyc7fijjWJddQyLn8Ig3JB5CqoB9Q=
 github.com/modern-go/concurrent v0.0.0-20180306012644-bacd9c7ef1dd h1:TRLaZ9cD/w8PVh93nsPXa1VrQ6jlwL5oN8l14QlcNfg=
 github.com/modern-go/concurrent v0.0.0-20180306012644-bacd9c7ef1dd/go.mod h1:6dJC0mAP4ikYIbvyc7fijjWJddQyLn8Ig3JB5CqoB9Q=
@@ -126,12 +144,15 @@ github.com/nlpodyssey/gopickle v0.3.0 h1:BLUE5gxFLyyNOPzlXxt6GoHEMMxD0qhsE4p0CIQ
 github.com/nlpodyssey/gopickle v0.3.0/go.mod h1:f070HJ/yR+eLi5WmM1OXJEGaTpuJEUiib19olXgYha0=
 github.com/olekukonko/tablewriter v0.0.5 h1:P2Ga83D34wi1o9J6Wh1mRuqd4mF/x/lgBS7N7AbDhec=
 github.com/olekukonko/tablewriter v0.0.5/go.mod h1:hPp6KlRPjbx+hW8ykQs1w3UBbZlj6HuIJcUGPhkA7kY=
-github.com/pdevine/tensor v0.0.0-20240228013915-64ccaa8d9ca9 h1:DV4iXjNn6fGeDl1AkZ1I0QB/0DBjrc7kPpxHrmuDzW4=
-github.com/pdevine/tensor v0.0.0-20240228013915-64ccaa8d9ca9/go.mod h1:nR7l3gM6ubiOm+mCkmmUyIBUcBAyiUmW6dQrDZhugFE=
-github.com/pelletier/go-toml/v2 v2.0.1/go.mod h1:r9LEWfGN8R5k0VXJ+0BkIe7MYkRdwZOjgMj2KwnJFUo=
-github.com/pelletier/go-toml/v2 v2.0.8 h1:0ctb6s9mE31h0/lhu+J6OPmVeDxJn+kYnJc2jZR9tGQ=
-github.com/pelletier/go-toml/v2 v2.0.8/go.mod h1:vuYfssBdrU2XDZ9bYydBu6t+6a6PYNcZljzZR9VXg+4=
-github.com/pkg/diff v0.0.0-20210226163009-20ebb0f2a09e/go.mod h1:pJLUxLENpZxwdsKMEsNbx1VGcRFpLqf3715MtcvvzbA=
+github.com/pdevine/tensor v0.0.0-20240510204454-f88f4562727c h1:GwiUUjKefgvSNmv3NCvI/BL0kDebW6Xa+kcdpdc1mTY=
+github.com/pdevine/tensor v0.0.0-20240510204454-f88f4562727c/go.mod h1:PSojXDXF7TbgQiD6kkd98IHOS0QqTyUEaWRiS8+BLu8=
+github.com/pelletier/go-toml/v2 v2.2.2 h1:aYUidT7k73Pcl9nb2gScu7NSrKCSHIDE89b3+6Wq+LM=
+github.com/pelletier/go-toml/v2 v2.2.2/go.mod h1:1t835xjRzz80PqgE6HHgN2JOsmgYu/h4qDAS4n929Rs=
+github.com/phpdave11/gofpdf v1.4.2/go.mod h1:zpO6xFn9yxo3YLyMvW8HcKWVdbNqgIfOOp2dXMnm1mY=
+github.com/phpdave11/gofpdi v1.0.12/go.mod h1:vBmVV0Do6hSBHC8uKUQ71JGW+ZGQq74llk/7bXwjDoI=
+github.com/pierrec/lz4/v4 v4.1.8 h1:ieHkV+i2BRzngO4Wd/3HGowuZStgq6QkPsD1eolNAO4=
+github.com/pierrec/lz4/v4 v4.1.8/go.mod h1:gZWDp/Ze/IJXGXf23ltt2EXimqmTUXEy0GFuRQyBid4=
+github.com/pkg/errors v0.8.1/go.mod h1:bwawxfHBFNV+L2hUp1rHADufV3IMtnDRdf1r5NINEl0=
 github.com/pkg/errors v0.9.1 h1:FEBLx1zS214owpjy7qsBeixbURkuhQAwrK5UwLGTwt4=
 github.com/pkg/errors v0.9.1/go.mod h1:bwawxfHBFNV+L2hUp1rHADufV3IMtnDRdf1r5NINEl0=
 github.com/pmezard/go-difflib v1.0.0 h1:4DBwDE0NGyQoBHbLQYPwSUPoCMWR5BEzIk/f1lZbAQM=
@@ -139,10 +160,11 @@ github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZN
 github.com/prometheus/client_model v0.0.0-20190812154241-14fe0d1b01d4/go.mod h1:xMI15A0UPsDsEKsMN9yxemIoYk6Tm2C1GtYGdfGttqA=
 github.com/rivo/uniseg v0.2.0 h1:S1pD9weZBuJdFmowNwbpi7BJ8TNftyUImj/0WQi72jY=
 github.com/rivo/uniseg v0.2.0/go.mod h1:J6wj4VEh+S6ZtnVlnTBMWIodfgj8LQOQFoIToxlJtxc=
-github.com/rogpeppe/go-internal v1.6.1/go.mod h1:xXDCJY+GAPziupqXw64V24skbSoqbTEfhy4qGm1nDQc=
+github.com/rogpeppe/fastuuid v1.2.0/go.mod h1:jVj6XXZzXRy/MSR5jhDC/2q6DgLz+nrA6LYCDYWNEvQ=
 github.com/rogpeppe/go-internal v1.8.0 h1:FCbCCtXNOY3UtUuHUYaghJg4y7Fd14rXifAYUAtL9R8=
 github.com/rogpeppe/go-internal v1.8.0/go.mod h1:WmiCO8CzOY8rg0OYDC4/i/2WRWAB6poM+XZ2dLUbcbE=
 github.com/russross/blackfriday/v2 v2.1.0/go.mod h1:+Rmxgy9KzJVeS9/2gXHxylqXiyQDYRxCVz55jmeOWTM=
+github.com/ruudk/golang-pdf417 v0.0.0-20181029194003-1af4ab5afa58/go.mod h1:6lfFZQK844Gfx8o5WFuvpxWRwnSoipWe/p622j1v06w=
 github.com/spf13/cobra v1.7.0 h1:hyqWnYt1ZQShIddO5kBpj3vu05/++x6tJ6dg8EC572I=
 github.com/spf13/cobra v1.7.0/go.mod h1:uLxZILRyS/50WlhOIKD7W6V5bgeIt+4sICxh6uRMrb0=
 github.com/spf13/pflag v1.0.5 h1:iy+VFUOCP1a+8yFto/drg2CJ5u0yRoB7fZw3DKv/JXA=
@@ -150,96 +172,119 @@ github.com/spf13/pflag v1.0.5/go.mod h1:McXfInJRrz4CZXVZOBLb0bTZqETkiAhM9Iw0y3An
 github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+wExME=
 github.com/stretchr/objx v0.4.0/go.mod h1:YvHI0jy2hoMjB+UWwv71VJQ9isScKT/TqJzVSSt89Yw=
 github.com/stretchr/objx v0.5.0/go.mod h1:Yh+to48EsGEfYuaHDzXPcE3xhTkx73EhmCGUpEOglKo=
+github.com/stretchr/objx v0.5.2/go.mod h1:FRsXN1f5AsAjCGJKqEizvkpNtU+EGNCLh3NxZ/8L+MA=
 github.com/stretchr/testify v1.1.4/go.mod h1:a8OnRcib4nhh0OaRAV+Yts87kKdq0PP7pXfy6kDkUVs=
-github.com/stretchr/testify v1.2.0/go.mod h1:a8OnRcib4nhh0OaRAV+Yts87kKdq0PP7pXfy6kDkUVs=
+github.com/stretchr/testify v1.2.2/go.mod h1:a8OnRcib4nhh0OaRAV+Yts87kKdq0PP7pXfy6kDkUVs=
 github.com/stretchr/testify v1.3.0/go.mod h1:M5WIy9Dh21IEIfnGCwXGc5bZfKNJtfHm1UVUgZn+9EI=
-github.com/stretchr/testify v1.6.1/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=
+github.com/stretchr/testify v1.5.1/go.mod h1:5W2xD1RspED5o8YsWQXVCued0rvSQ+mT+I5cxcmMvtA=
 github.com/stretchr/testify v1.7.0/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=
 github.com/stretchr/testify v1.7.1/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=
 github.com/stretchr/testify v1.8.0/go.mod h1:yNjHg4UonilssWZ8iaSj1OCr/vHnekPRkoO+kdMU+MU=
 github.com/stretchr/testify v1.8.1/go.mod h1:w2LPCIKwWwSfY2zedu0+kehJoqGctiVI29o6fzry7u4=
-github.com/stretchr/testify v1.8.2/go.mod h1:w2LPCIKwWwSfY2zedu0+kehJoqGctiVI29o6fzry7u4=
-github.com/stretchr/testify v1.8.3/go.mod h1:sz/lmYIOXD/1dqDmKjjqLyZ2RngseejIcXlSw2iwfAo=
-github.com/stretchr/testify v1.8.4 h1:CcVxjf3Q8PM0mHUKJCdn+eZZtm5yQwehR5yeSVQQcUk=
 github.com/stretchr/testify v1.8.4/go.mod h1:sz/lmYIOXD/1dqDmKjjqLyZ2RngseejIcXlSw2iwfAo=
+github.com/stretchr/testify v1.9.0 h1:HtqpIVDClZ4nwg75+f6Lvsy/wHu+3BoSGCbBAcpTsTg=
+github.com/stretchr/testify v1.9.0/go.mod h1:r2ic/lqez/lEtzL7wO/rwa5dbSLXVDPFyf8C91i36aY=
 github.com/twitchyliquid64/golang-asm v0.15.1 h1:SU5vSMR7hnwNxj24w34ZyCi/FmDZTkS4MhqMhdFk5YI=
 github.com/twitchyliquid64/golang-asm v0.15.1/go.mod h1:a1lVb/DtPvCB8fslRZhAngC2+aY1QWCk3Cedj/Gdt08=
-github.com/ugorji/go v1.2.7/go.mod h1:nF9osbDWLy6bDVv/Rtoh6QgnvNDpmCalQV5urGCCS6M=
-github.com/ugorji/go/codec v1.2.7/go.mod h1:WGN1fab3R1fzQlVQTkfxVtIBhWDRqOviHU95kRgeqEY=
-github.com/ugorji/go/codec v1.2.11 h1:BMaWp1Bb6fHwEtbplGBGJ498wD+LKlNSl25MjdZY4dU=
-github.com/ugorji/go/codec v1.2.11/go.mod h1:UNopzCgEMSXjBc6AOMqYvWC1ktqTAfzJZUZgYf6w6lg=
+github.com/ugorji/go/codec v1.2.12 h1:9LC83zGrHhuUA9l16C9AHXAqEV/2wBQ4nkvumAE65EE=
+github.com/ugorji/go/codec v1.2.12/go.mod h1:UNopzCgEMSXjBc6AOMqYvWC1ktqTAfzJZUZgYf6w6lg=
 github.com/x448/float16 v0.8.4 h1:qLwI1I70+NjRFUR3zs1JPUCgaCXSh3SW62uAKT1mSBM=
 github.com/x448/float16 v0.8.4/go.mod h1:14CWIYCyZA/cWjXOioeEpHeN/83MdbZDRQHoFcYsOfg=
 github.com/xtgo/set v1.0.0 h1:6BCNBRv3ORNDQ7fyoJXRv+tstJz3m1JVFQErfeZz2pY=
 github.com/xtgo/set v1.0.0/go.mod h1:d3NHzGzSa0NmB2NhFyECA+QdRp29oEn2xbT+TpeFoM8=
 github.com/yuin/goldmark v1.1.27/go.mod h1:3hX8gzYuyVAZsxl0MRgGTJEmQBFcNTphYh9decYSb74=
 github.com/yuin/goldmark v1.2.1/go.mod h1:3hX8gzYuyVAZsxl0MRgGTJEmQBFcNTphYh9decYSb74=
+github.com/yuin/goldmark v1.3.5/go.mod h1:mwnBkeHKe2W/ZEtQ+71ViKU8L12m81fl3OWwC1Zlc8k=
+go.opentelemetry.io/proto/otlp v0.7.0/go.mod h1:PqfVotwruBrMGOCsRd/89rSnXhoiJIqeYNgFYFoEGnI=
 go4.org/unsafe/assume-no-moving-gc v0.0.0-20231121144256-b99613f794b6 h1:lGdhQUN/cnWdSH3291CUuxSEqc+AsGTiDxPP3r2J0l4=
 go4.org/unsafe/assume-no-moving-gc v0.0.0-20231121144256-b99613f794b6/go.mod h1:FftLjUGFEDu5k8lt0ddY+HcrH/qU/0qk+H8j9/nTl3E=
 golang.org/x/arch v0.0.0-20210923205945-b76863e36670/go.mod h1:5om86z9Hs0C8fWVUuoMHwpExlXzs5Tkyp9hOrfG7pp8=
-golang.org/x/arch v0.3.0 h1:02VY4/ZcO/gBOH6PUaoiptASxtXU10jazRCP865E97k=
-golang.org/x/arch v0.3.0/go.mod h1:5om86z9Hs0C8fWVUuoMHwpExlXzs5Tkyp9hOrfG7pp8=
+golang.org/x/arch v0.8.0 h1:3wRIsP3pM4yUptoR96otTUOXI367OS0+c9eeRi9doIc=
+golang.org/x/arch v0.8.0/go.mod h1:FEVrYAQjsQXMVJ1nsMoVVXPZg6p2JE2mx8psSWTDQys=
 golang.org/x/crypto v0.0.0-20190308221718-c2843e01d9a2/go.mod h1:djNgcEr1/C05ACkg1iLfiJU5Ep61QUkGW8qpdssI0+w=
+golang.org/x/crypto v0.0.0-20190510104115-cbcb75029529/go.mod h1:yigFU9vqHzYiE8UmvKecakEJjdnWj3jj499lnFckfCI=
 golang.org/x/crypto v0.0.0-20191011191535-87dc89f01550/go.mod h1:yigFU9vqHzYiE8UmvKecakEJjdnWj3jj499lnFckfCI=
 golang.org/x/crypto v0.0.0-20200622213623-75b288015ac9/go.mod h1:LzIPMQfyMNhhGPhUkYOs5KpL4U8rLKemX1yGLhDgUto=
-golang.org/x/crypto v0.0.0-20210711020723-a769d52b0f97/go.mod h1:GvvjBRRGRdwPK5ydBHafDWAxML/pGHZbMvKqRZ5+Abc=
-golang.org/x/crypto v0.14.0 h1:wBqGXzWJW6m1XrIKlAH0Hs1JJ7+9KBwnIO8v66Q9cHc=
-golang.org/x/crypto v0.14.0/go.mod h1:MVFd36DqK4CsrnJYDkBA3VC4m2GkXAM0PvzMCn4JQf4=
+golang.org/x/crypto v0.23.0 h1:dIJU/v2J8Mdglj/8rJ6UUOM3Zc9zLZxVZwwxMooUSAI=
+golang.org/x/crypto v0.23.0/go.mod h1:CKFgDieR+mRhux2Lsu27y0fO304Db0wZe70UKqHu0v8=
 golang.org/x/exp v0.0.0-20180321215751-8460e604b9de/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
 golang.org/x/exp v0.0.0-20180807140117-3d87b88a115f/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
 golang.org/x/exp v0.0.0-20190121172915-509febef88a4/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
 golang.org/x/exp v0.0.0-20190125153040-c74c464bbbf2/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
-golang.org/x/exp v0.0.0-20230817173708-d852ddb80c63 h1:m64FZMko/V45gv0bNmrNYoDEq8U5YUhetc9cBWKS1TQ=
-golang.org/x/exp v0.0.0-20230817173708-d852ddb80c63/go.mod h1:0v4NqG35kSWCMzLaMeX+IQrlSnVE/bqGSyC2cz/9Le8=
+golang.org/x/exp v0.0.0-20190306152737-a1d7652674e8/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
+golang.org/x/exp v0.0.0-20191002040644-a1355ae1e2c3/go.mod h1:NOZ3BPKG0ec/BKJQgnvsSFpcKLM5xXVWnvZS97DWHgE=
+golang.org/x/exp v0.0.0-20231110203233-9a3e6036ecaa h1:FRnLl4eNAQl8hwxVVC17teOw8kdjVDVAiFMtgUdTSRQ=
+golang.org/x/exp v0.0.0-20231110203233-9a3e6036ecaa/go.mod h1:zk2irFbV9DP96SEBUUAy67IdHUaZuSnrz1n472HUCLE=
 golang.org/x/image v0.0.0-20180708004352-c73c2afc3b81/go.mod h1:ux5Hcp/YLpHSI86hEcLt0YII63i6oz57MZXIpbrjZUs=
+golang.org/x/image v0.0.0-20190227222117-0694c2d4d067/go.mod h1:kZ7UVZpmo3dzQBMxlp+ypCbDeSB+sBbTgSJuh5dn5js=
+golang.org/x/image v0.0.0-20190802002840-cff245a6509b/go.mod h1:FeLwcggjj3mMvU+oOTbSwawSJRM1uh48EjtB4UJZlP0=
+golang.org/x/image v0.0.0-20190910094157-69e4b8554b2a/go.mod h1:FeLwcggjj3mMvU+oOTbSwawSJRM1uh48EjtB4UJZlP0=
+golang.org/x/image v0.0.0-20200119044424-58c23975cae1/go.mod h1:FeLwcggjj3mMvU+oOTbSwawSJRM1uh48EjtB4UJZlP0=
+golang.org/x/image v0.0.0-20200430140353-33d19683fad8/go.mod h1:FeLwcggjj3mMvU+oOTbSwawSJRM1uh48EjtB4UJZlP0=
+golang.org/x/image v0.0.0-20200618115811-c13761719519/go.mod h1:FeLwcggjj3mMvU+oOTbSwawSJRM1uh48EjtB4UJZlP0=
+golang.org/x/image v0.0.0-20201208152932-35266b937fa6/go.mod h1:FeLwcggjj3mMvU+oOTbSwawSJRM1uh48EjtB4UJZlP0=
+golang.org/x/image v0.0.0-20210216034530-4410531fe030/go.mod h1:FeLwcggjj3mMvU+oOTbSwawSJRM1uh48EjtB4UJZlP0=
 golang.org/x/lint v0.0.0-20181026193005-c67002cb31c3/go.mod h1:UVdnD1Gm6xHRNCYTkRU2/jEulfH38KcIWyp/GAMgvoE=
 golang.org/x/lint v0.0.0-20190227174305-5b3e6a55c961/go.mod h1:wehouNa3lNwaWXcvxsM5YxQ5yQlVC4a0KAMCusXpPoU=
 golang.org/x/lint v0.0.0-20190313153728-d0100b6bd8b3/go.mod h1:6SW0HCj/g11FgYtHlgUYUwCkIfeOF89ocIRzGO/8vkc=
+golang.org/x/lint v0.0.0-20210508222113-6edffad5e616/go.mod h1:3xt1FjdF8hUf6vQPIChWIBhFzV8gjjsPE/fR3IyQdNY=
+golang.org/x/mobile v0.0.0-20190719004257-d2bd2a29d028/go.mod h1:E/iHnbuqvinMTCcRqshq8CkpyQDoeVncDDYHnLhea+o=
+golang.org/x/mod v0.1.0/go.mod h1:0QHyrYULN0/3qlju5TqG8bIK38QM8yzMo5ekMj3DlcY=
+golang.org/x/mod v0.1.1-0.20191105210325-c90efee705ee/go.mod h1:QqPTAvyqsEbceGzBzNggFXnrqF1CaUcvgkdR5Ot7KZg=
 golang.org/x/mod v0.2.0/go.mod h1:s0Qsj1ACt9ePp/hMypM3fl4fZqREWJwdYDEqhRiZZUA=
 golang.org/x/mod v0.3.0/go.mod h1:s0Qsj1ACt9ePp/hMypM3fl4fZqREWJwdYDEqhRiZZUA=
+golang.org/x/mod v0.4.2/go.mod h1:s0Qsj1ACt9ePp/hMypM3fl4fZqREWJwdYDEqhRiZZUA=
 golang.org/x/net v0.0.0-20180724234803-3673e40ba225/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
 golang.org/x/net v0.0.0-20180826012351-8a410e7b638d/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
+golang.org/x/net v0.0.0-20190108225652-1e06a53dbb7e/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
 golang.org/x/net v0.0.0-20190213061140-3a22650c66bd/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
 golang.org/x/net v0.0.0-20190311183353-d8887717615a/go.mod h1:t9HGtf8HONx5eT2rtn7q6eTqICYqUVnKs3thJo3Qplg=
 golang.org/x/net v0.0.0-20190404232315-eb5bcb51f2a3/go.mod h1:t9HGtf8HONx5eT2rtn7q6eTqICYqUVnKs3thJo3Qplg=
 golang.org/x/net v0.0.0-20190620200207-3b0461eec859/go.mod h1:z5CRVTTTmAJ677TzLLGU+0bjPO0LkuOLi4/5GtJWs/s=
 golang.org/x/net v0.0.0-20200226121028-0de0cce0169b/go.mod h1:z5CRVTTTmAJ677TzLLGU+0bjPO0LkuOLi4/5GtJWs/s=
-golang.org/x/net v0.0.0-20200904194848-62affa334b73/go.mod h1:/O7V0waA8r7cgGh81Ro3o1hOxt32SMVPicZroKQ2sZA=
+golang.org/x/net v0.0.0-20200822124328-c89045814202/go.mod h1:/O7V0waA8r7cgGh81Ro3o1hOxt32SMVPicZroKQ2sZA=
 golang.org/x/net v0.0.0-20201021035429-f5854403a974/go.mod h1:sp8m0HH+o8qH0wwXwYZr8TS3Oi6o0r6Gce1SSxlDquU=
-golang.org/x/net v0.0.0-20210226172049-e18ecbb05110/go.mod h1:m0MpNAwzfU5UDzcl9v0D8zg8gWTRqZa9RBIspLL5mdg=
-golang.org/x/net v0.17.0 h1:pVaXccu2ozPjCXewfr1S7xza/zcXTity9cCdXQYSjIM=
-golang.org/x/net v0.17.0/go.mod h1:NxSsAGuq816PNPmqtQdLE42eU2Fs7NoRIZrHJAlaCOE=
+golang.org/x/net v0.0.0-20210405180319-a5a99cb37ef4/go.mod h1:p54w0d4576C0XHj96bSt6lcn1PtDYWL6XObtHCRCNQM=
+golang.org/x/net v0.0.0-20210614182718-04defd469f4e/go.mod h1:9nx3DQGgdP8bBQD5qxJ1jj9UTztislL4KSBs9R2vV5Y=
+golang.org/x/net v0.25.0 h1:d/OCCoBEUq33pjydKrGQhw7IlUPI2Oylr+8qLx49kac=
+golang.org/x/net v0.25.0/go.mod h1:JkAGAh7GEvH74S6FOH42FLoXpXbE/aqXSrIQjXgsiwM=
 golang.org/x/oauth2 v0.0.0-20180821212333-d2e6202438be/go.mod h1:N/0e6XlmueqKjAGxoOufVs8QHGRruUQn6yWY3a++T0U=
+golang.org/x/oauth2 v0.0.0-20200107190931-bf48bf16ab8d/go.mod h1:gOpvHmFTYa4IltrdGE7lF6nIHvwfUNPOp7c8zoXwtLw=
 golang.org/x/sync v0.0.0-20180314180146-1d60e4601c6f/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20181108010431-42b317875d0f/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
+golang.org/x/sync v0.0.0-20181221193216-37e7f081c4d4/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20190423024810-112230192c58/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20190911185100-cd5d95a43a6e/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20201020160332-67f06af15bc9/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
+golang.org/x/sync v0.0.0-20210220032951-036812b2e83c/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.3.0 h1:ftCYgMx6zT/asHUrPw8BLLscYtGznsLAnjq5RH9P66E=
 golang.org/x/sync v0.3.0/go.mod h1:FU7BRWz2tNW+3quACPkgCx/L+uEAv1htQ0V83Z9Rj+Y=
 golang.org/x/sys v0.0.0-20180830151530-49385e6e1522/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20190215142949-d0b11bdaac8a/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
+golang.org/x/sys v0.0.0-20190312061237-fead79001313/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
 golang.org/x/sys v0.0.0-20190412213103-97732733099d/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
 golang.org/x/sys v0.0.0-20200323222414-85ca7c5b95cd/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
-golang.org/x/sys v0.0.0-20200909081042-eff7692f9009/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
 golang.org/x/sys v0.0.0-20200930185726-fdedc70b468f/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
 golang.org/x/sys v0.0.0-20201119102817-f84b799fce68/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
 golang.org/x/sys v0.0.0-20210124154548-22da62e12c0c/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
-golang.org/x/sys v0.0.0-20210615035016-665e8c7367d1/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
+golang.org/x/sys v0.0.0-20210304124612-50617c2ba197/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
+golang.org/x/sys v0.0.0-20210330210617-4fbd30eecc44/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
+golang.org/x/sys v0.0.0-20210423082822-04245dca01da/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
+golang.org/x/sys v0.0.0-20210510120138-977fb7262007/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.0.0-20210630005230-0f9fa26af87c/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
-golang.org/x/sys v0.0.0-20210806184541-e5e7981a1069/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
-golang.org/x/sys v0.0.0-20220704084225-05e143d24a9e/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
+golang.org/x/sys v0.5.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.6.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
-golang.org/x/sys v0.13.0 h1:Af8nKPmuFypiUBjVoU9V20FiaFXOcuZI21p0ycVYYGE=
-golang.org/x/sys v0.13.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
+golang.org/x/sys v0.20.0 h1:Od9JTbYCk261bKm4M/mw7AklTlFYIa0bIp9BgSm1S8Y=
+golang.org/x/sys v0.20.0/go.mod h1:/VUhepiaJMQUp4+oa/7Zr1D23ma6VTLIYjOOTFZPUcA=
 golang.org/x/term v0.0.0-20201126162022-7de9c90e9dd1/go.mod h1:bj7SfCRtBDWHUb9snDiAeCFNEtKQo2Wmx5Cou7ajbmo=
-golang.org/x/term v0.13.0 h1:bb+I9cTfFazGW51MZqBVmZy7+JEJMouUHTUSKVQLBek=
-golang.org/x/term v0.13.0/go.mod h1:LTmsnFJwVN6bCy1rVCoS+qHT1HhALEFxKncY3WNNh4U=
+golang.org/x/term v0.20.0 h1:VnkxpohqXaOBYJtBmEppKUG6mXpi+4O6purfc2+sMhw=
+golang.org/x/term v0.20.0/go.mod h1:8UkIAJTvZgivsXaD6/pH6U9ecQzZ45awqEOzuCvwpFY=
 golang.org/x/text v0.3.0/go.mod h1:NqM8EUOU14njkJ3fqMW+pc6Ldnwhi/IjpwHt7yyuwOQ=
 golang.org/x/text v0.3.3/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
+golang.org/x/text v0.3.5/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
 golang.org/x/text v0.3.6/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
-golang.org/x/text v0.14.0 h1:ScX5w1eTa3QqT8oi6+ziP7dTV1S2+ALU0bI+0zXKWiQ=
-golang.org/x/text v0.14.0/go.mod h1:18ZOQIKpY8NJVqYksKHtTdi31H5itFRjB5/qKTNYzSU=
+golang.org/x/text v0.15.0 h1:h1V/4gjBv8v9cjcR6+AR5+/cIYK5N/WAgiv4xlsEtAk=
+golang.org/x/text v0.15.0/go.mod h1:18ZOQIKpY8NJVqYksKHtTdi31H5itFRjB5/qKTNYzSU=
 golang.org/x/tools v0.0.0-20180525024113-a5b4c53f6e8b/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
 golang.org/x/tools v0.0.0-20180917221912-90fa682c2a6e/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
 golang.org/x/tools v0.0.0-20190114222345-bf090417da8b/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
@@ -247,34 +292,40 @@ golang.org/x/tools v0.0.0-20190206041539-40960b6deb8e/go.mod h1:n7NCudcB/nEzxVGm
 golang.org/x/tools v0.0.0-20190226205152-f727befe758c/go.mod h1:9Yl7xja0Znq3iFh3HoIrodX9oNMXvdceNzlUR8zjMvY=
 golang.org/x/tools v0.0.0-20190311212946-11955173bddd/go.mod h1:LCzVGOaR6xXOjkQ3onu1FJEFr0SW1gC7cKk1uF8kGRs=
 golang.org/x/tools v0.0.0-20190524140312-2c0ae7006135/go.mod h1:RgjU9mgBXZiqYHBnxXauZ1Gv1EHHAz9KjViQ78xBX0Q=
+golang.org/x/tools v0.0.0-20190927191325-030b2cf1153e/go.mod h1:b+2E5dAYhXwXZwtnZ6UAqBI28+e2cm9otk0dWdXHAEo=
 golang.org/x/tools v0.0.0-20191119224855-298f0cb1881e/go.mod h1:b+2E5dAYhXwXZwtnZ6UAqBI28+e2cm9otk0dWdXHAEo=
+golang.org/x/tools v0.0.0-20200130002326-2f3ba24bd6e7/go.mod h1:TB2adYChydJhpapKDTa4BR/hXlZSLoq2Wpct/0txZ28=
 golang.org/x/tools v0.0.0-20200619180055-7c47624df98f/go.mod h1:EkVYQZoAsY45+roYkvgYkIh4xh/qjgUK9TdY2XT94GE=
 golang.org/x/tools v0.0.0-20210106214847-113979e3529a/go.mod h1:emZCQorbCU4vsT4fOWvOPXz4eW1wZW4PmDk9uLelYpA=
+golang.org/x/tools v0.1.4/go.mod h1:o0xws9oXOQQZyjljx8fwUC0k7L1pTE6eaCbjGeHmOkk=
 golang.org/x/xerrors v0.0.0-20190717185122-a985d3407aa7/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 golang.org/x/xerrors v0.0.0-20191011141410-1b5146add898/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 golang.org/x/xerrors v0.0.0-20191204190536-9bdfabe68543/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 golang.org/x/xerrors v0.0.0-20200804184101-5ec99f83aff1 h1:go1bK/D/BFZV2I8cIQd1NKEZ+0owSTG1fDTci4IqFcE=
 golang.org/x/xerrors v0.0.0-20200804184101-5ec99f83aff1/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 gonum.org/v1/gonum v0.0.0-20180816165407-929014505bf4/go.mod h1:Y+Yx5eoAFn32cQvJDxZx5Dpnq+c3wtXuadVZAcxbbBo=
-gonum.org/v1/gonum v0.8.2 h1:CCXrcPKiGGotvnN6jfUsKk4rRqm7q09/YbKb5xCEvtM=
 gonum.org/v1/gonum v0.8.2/go.mod h1:oe/vMfY3deqTw+1EZJhuvEW2iwGF1bW9wwu7XCu0+v0=
-gonum.org/v1/netlib v0.0.0-20190313105609-8cb42192e0e0 h1:OE9mWmgKkjJyEmDAAtGMPjXu+YNeGvK9VTSHY6+Qihc=
+gonum.org/v1/gonum v0.9.3/go.mod h1:TZumC3NeyVQskjXqmyWt4S3bINhy7B4eYwW69EbyX+0=
+gonum.org/v1/gonum v0.15.0 h1:2lYxjRbTYyxkJxlhC+LvJIx3SsANPdRybu1tGj9/OrQ=
+gonum.org/v1/gonum v0.15.0/go.mod h1:xzZVBJBtS+Mz4q0Yl2LJTk+OxOg4jiXZ7qBoM0uISGo=
 gonum.org/v1/netlib v0.0.0-20190313105609-8cb42192e0e0/go.mod h1:wa6Ws7BG/ESfp6dHfk7C6KdzKA7wR7u/rKwOGE66zvw=
 gonum.org/v1/plot v0.0.0-20190515093506-e2840ee46a6b/go.mod h1:Wt8AAjI+ypCyYX3nZBvf6cAIx93T+c/OS2HFAYskSZc=
+gonum.org/v1/plot v0.9.0/go.mod h1:3Pcqqmp6RHvJI72kgb8fThyUnav364FOsdDo2aGW5lY=
 google.golang.org/appengine v1.1.0/go.mod h1:EbEs0AVv82hx2wNQdGPgUI5lhzA/G0D9YwlJXL52JkM=
 google.golang.org/appengine v1.4.0/go.mod h1:xpcJRLb0r/rnEns0DIKYYv+WjYCduHsrkT7/EB5XEv4=
 google.golang.org/genproto v0.0.0-20180817151627-c66870c02cf8/go.mod h1:JiN7NxoALGmiZfu7CAH4rXhgtRTLTxftemlI0sWmxmc=
 google.golang.org/genproto v0.0.0-20190819201941-24fa4b261c55/go.mod h1:DMBHOl98Agz4BDEuKkezgsaosCRResVns1a3J2ZsMNc=
+google.golang.org/genproto v0.0.0-20200513103714-09dca8ec2884/go.mod h1:55QSHmfGQM9UVYDPBsyGGes0y52j32PQ3BqQfXhyH3c=
 google.golang.org/genproto v0.0.0-20200526211855-cb27e3aa2013/go.mod h1:NbSheEEYHJ7i3ixzK3sjbqSGDJWnxyFXZblF3eUsNvo=
-google.golang.org/genproto v0.0.0-20200911024640-645f7a48b24f h1:Yv4xsIx7HZOoyUGSJ2ksDyWE2qIBXROsZKt2ny3hCGM=
-google.golang.org/genproto v0.0.0-20200911024640-645f7a48b24f/go.mod h1:FWY/as6DDZQgahTzZj3fqbO1CbirC29ZNUFHwi0/+no=
+google.golang.org/genproto v0.0.0-20210630183607-d20f26d13c79/go.mod h1:yiaVoXHpRzHGyxV3o4DktVWY4mSUErTKaeEOq6C3t3U=
 google.golang.org/grpc v1.19.0/go.mod h1:mqu4LbDTu4XGKhr4mRzUsmM4RtVoemTSY81AxZiDr8c=
 google.golang.org/grpc v1.23.0/go.mod h1:Y5yQAOtifL1yxbo5wqy6BxZv8vAUGQwXBOALyacEbxg=
 google.golang.org/grpc v1.25.1/go.mod h1:c3i+UQWmh7LiEpx4sFZnkU36qjEYZ0imhYfXVyQciAY=
 google.golang.org/grpc v1.27.0/go.mod h1:qbnxyOmOxrQa7FizSgH+ReBfzJrCY1pSN7KXBS8abTk=
-google.golang.org/grpc v1.32.0 h1:zWTV+LMdc3kaiJMSTOFz2UgSBgx8RNQoTGiZu3fR9S0=
-google.golang.org/grpc v1.32.0/go.mod h1:N36X2cJ7JwdamYAgDz+s+rVMFjt3numwzf/HckM8pak=
-google.golang.org/grpc/cmd/protoc-gen-go-grpc v0.0.0-20200910201057-6591123024b3/go.mod h1:6Kw0yEErY5E/yWrBtf03jp27GLLJujG4z/JK95pnjjw=
+google.golang.org/grpc v1.33.1/go.mod h1:fr5YgcSWrqhRRxogOsw7RzIpsmvOZ6IcH4kBYTpR3n0=
+google.golang.org/grpc v1.36.0/go.mod h1:qjiiYl8FncCW8feJPdyg3v6XW24KsRHe+dy9BAGRRjU=
+google.golang.org/grpc v1.38.0/go.mod h1:NREThFqKR1f3iQ6oBuvc5LadQuXVGo9rkm5ZGrQdJfM=
+google.golang.org/grpc v1.39.0/go.mod h1:PImNr+rS9TWYb2O4/emRugxiyHZ5JyHW5F+RPnDzfrE=
 google.golang.org/protobuf v0.0.0-20200109180630-ec00e32a8dfd/go.mod h1:DFci5gLYBciE7Vtevhsrf46CRTquxDuWsQurQQe4oz8=
 google.golang.org/protobuf v0.0.0-20200221191635-4d8936d0db64/go.mod h1:kwYJMbMJ01Woi6D6+Kah6886xMZcty6N08ah7+eCXa0=
 google.golang.org/protobuf v0.0.0-20200228230310-ab0ca4ff8a60/go.mod h1:cfTl7dwQJ+fmap5saPgwCLgHXTUD7jkjRqWcaiX5VyM=
@@ -283,20 +334,18 @@ google.golang.org/protobuf v1.21.0/go.mod h1:47Nbq4nVaFHyn7ilMalzfO3qCViNmqZ2kzi
 google.golang.org/protobuf v1.22.0/go.mod h1:EGpADcykh3NcUnDUJcl1+ZksZNG86OlYog2l/sGQquU=
 google.golang.org/protobuf v1.23.0/go.mod h1:EGpADcykh3NcUnDUJcl1+ZksZNG86OlYog2l/sGQquU=
 google.golang.org/protobuf v1.23.1-0.20200526195155-81db48ad09cc/go.mod h1:EGpADcykh3NcUnDUJcl1+ZksZNG86OlYog2l/sGQquU=
-google.golang.org/protobuf v1.24.0/go.mod h1:r/3tXBNzIEhYS9I1OUVjXDlt8tc493IdKGjtUeSXeh4=
 google.golang.org/protobuf v1.25.0/go.mod h1:9JNX74DMeImyA3h4bdi1ymwjUzf21/xIlbajtzgsN7c=
 google.golang.org/protobuf v1.26.0-rc.1/go.mod h1:jlhhOSvTdKEhbULTjvd4ARK9grFBp09yW+WbY/TyQbw=
-google.golang.org/protobuf v1.28.0/go.mod h1:HV8QOd/L58Z+nl8r43ehVNZIU/HEI6OcFqwMG9pJV4I=
-google.golang.org/protobuf v1.30.0 h1:kPPoIgf3TsEvrm0PFe15JQ+570QVxYzEvvHqChK+cng=
-google.golang.org/protobuf v1.30.0/go.mod h1:HV8QOd/L58Z+nl8r43ehVNZIU/HEI6OcFqwMG9pJV4I=
+google.golang.org/protobuf v1.26.0/go.mod h1:9q0QmTI4eRPtz6boOQmLYwt+qCgq0jsYwAQnmE0givc=
+google.golang.org/protobuf v1.27.1/go.mod h1:9q0QmTI4eRPtz6boOQmLYwt+qCgq0jsYwAQnmE0givc=
+google.golang.org/protobuf v1.34.1 h1:9ddQBjfCyZPOHPUiPxpYESBLc+T8P3E+Vo4IbKZgFWg=
+google.golang.org/protobuf v1.34.1/go.mod h1:c6P6GXX6sHbq/GpV6MGZEdwhWPcYBgnhAHhKbcUYpos=
 gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
-gopkg.in/check.v1 v1.0.0-20180628173108-788fd7840127/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
 gopkg.in/check.v1 v1.0.0-20201130134442-10cb98267c6c h1:Hei/4ADfdWqJk1ZMxUNpqntNwaWcugrBjAiHlqqRiVk=
 gopkg.in/check.v1 v1.0.0-20201130134442-10cb98267c6c/go.mod h1:JHkPIbrfpd72SG/EVd6muEfDQjcINNoR0C8j2r3qZ4Q=
-gopkg.in/errgo.v2 v2.1.0/go.mod h1:hNsd1EY+bozCKY1Ytp96fpM3vjJbqLJn88ws8XvfDNI=
-gopkg.in/yaml.v2 v2.4.0/go.mod h1:RDklbk79AGWmwhnvt/jBztapEOGDOx6ZbXqjP6csGnQ=
+gopkg.in/yaml.v2 v2.2.2/go.mod h1:hI93XBmqTisBFMUTm0b8Fm+jr3Dg1NNxqwp+5A1VGuI=
+gopkg.in/yaml.v2 v2.2.3/go.mod h1:hI93XBmqTisBFMUTm0b8Fm+jr3Dg1NNxqwp+5A1VGuI=
 gopkg.in/yaml.v3 v3.0.0-20200313102051-9f266ea9e77c/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
-gopkg.in/yaml.v3 v3.0.0-20210107192922-496545a6307b/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
 gopkg.in/yaml.v3 v3.0.1 h1:fxVm/GzAzEWqLHuvctI91KS9hhNmmWOoWu0XTYJS7CA=
 gopkg.in/yaml.v3 v3.0.1/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
 gorgonia.org/vecf32 v0.9.0 h1:PClazic1r+JVJ1dEzRXgeiVl4g1/Hf/w+wUSqnco1Xg=
@@ -305,4 +354,5 @@ gorgonia.org/vecf64 v0.9.0 h1:bgZDP5x0OzBF64PjMGC3EvTdOoMEcmfAh1VCUnZFm1A=
 gorgonia.org/vecf64 v0.9.0/go.mod h1:hp7IOWCnRiVQKON73kkC/AUMtEXyf9kGlVrtPQ9ccVA=
 honnef.co/go/tools v0.0.0-20190102054323-c2f93a96b099/go.mod h1:rf3lG4BRIbNafJWhAfAdb/ePZxsR/4RtNHQocxwk9r4=
 honnef.co/go/tools v0.0.0-20190523083050-ea95bdfd59fc/go.mod h1:rf3lG4BRIbNafJWhAfAdb/ePZxsR/4RtNHQocxwk9r4=
+nullprogram.com/x/optparse v1.0.0/go.mod h1:KdyPE+Igbe0jQUrVfMqDMeJQIJZEuyV7pjYmp6pbG50=
 rsc.io/pdf v0.1.1/go.mod h1:n8OzWcQ6Sp37PL01nO98y4iUCRdTGarVfzxY20ICaU4=

gpu/amd_hip_windows.go

@@ -3,7 +3,6 @@ package gpu
 import (
 	"fmt"
 	"log/slog"
-	"strconv"
 	"syscall"
 	"unsafe"
 
@@ -74,16 +73,22 @@ func (hl *HipLib) Release() {
 	hl.dll = 0
 }
 
-func (hl *HipLib) AMDDriverVersion() (string, error) {
+func (hl *HipLib) AMDDriverVersion() (driverMajor, driverMinor int, err error) {
 	if hl.dll == 0 {
-		return "", fmt.Errorf("dll has been unloaded")
+		return 0, 0, fmt.Errorf("dll has been unloaded")
 	}
 	var version int
 	status, _, err := syscall.SyscallN(hl.hipDriverGetVersion, uintptr(unsafe.Pointer(&version)))
 	if status != hipSuccess {
-		return "", fmt.Errorf("failed call to hipDriverGetVersion: %d %s", status, err)
+		return 0, 0, fmt.Errorf("failed call to hipDriverGetVersion: %d %s", status, err)
 	}
-	return strconv.Itoa(version), nil
+
+	slog.Debug("hipDriverGetVersion", "version", version)
+	// TODO - this isn't actually right, but the docs claim hipDriverGetVersion isn't accurate anyway...
+	driverMajor = version / 1000
+	driverMinor = (version - (driverMajor * 1000)) / 10
+
+	return driverMajor, driverMinor, nil
 }
 
 func (hl *HipLib) HipGetDeviceCount() int {

gpu/amd_linux.go

@@ -8,6 +8,7 @@ import (
 	"log/slog"
 	"os"
 	"path/filepath"
+	"regexp"
 	"slices"
 	"strconv"
 	"strings"
@@ -41,10 +42,8 @@ func AMDGetGPUInfo() []GpuInfo {
 	}
 
 	// Opportunistic logging of driver version to aid in troubleshooting
-	ver, err := AMDDriverVersion()
-	if err == nil {
-		slog.Info("AMD Driver: " + ver)
-	} else {
+	driverMajor, driverMinor, err := AMDDriverVersion()
+	if err != nil {
 		// TODO - if we see users crash and burn with the upstreamed kernel this can be adjusted to hard-fail rocm support and fallback to CPU
 		slog.Warn("ollama recommends running the https://www.amd.com/en/support/linux-drivers", "error", err)
 	}
@@ -91,6 +90,7 @@ func AMDGetGPUInfo() []GpuInfo {
 		scanner := bufio.NewScanner(fp)
 		isCPU := false
 		var major, minor, patch uint64
+		var vendor, device uint64
 		for scanner.Scan() {
 			line := strings.TrimSpace(scanner.Text())
 			// Note: we could also use "cpu_cores_count X" where X is greater than zero to detect CPUs
@@ -118,6 +118,26 @@ func AMDGetGPUInfo() []GpuInfo {
 					slog.Debug("malformed int " + line)
 					continue
 				}
+			} else if strings.HasPrefix(line, "vendor_id") {
+				ver := strings.Fields(line)
+				if len(ver) != 2 {
+					slog.Debug("malformed vendor_id", "vendor_id", line)
+					continue
+				}
+				vendor, err = strconv.ParseUint(ver[1], 10, 32)
+				if err != nil {
+					slog.Debug("malformed vendor_id" + line)
+				}
+			} else if strings.HasPrefix(line, "device_id") {
+				ver := strings.Fields(line)
+				if len(ver) != 2 {
+					slog.Debug("malformed device_id", "device_id", line)
+					continue
+				}
+				device, err = strconv.ParseUint(ver[1], 10, 32)
+				if err != nil {
+					slog.Debug("malformed device_id" + line)
+				}
 			}
 
 			// TODO - any other properties we want to extract and record?
@@ -140,7 +160,7 @@ func AMDGetGPUInfo() []GpuInfo {
 		}
 
 		if int(major) < RocmComputeMin {
-			slog.Warn(fmt.Sprintf("amdgpu too old gfx%d%d%x", major, minor, patch), "gpu", gpuID)
+			slog.Warn(fmt.Sprintf("amdgpu too old gfx%d%x%x", major, minor, patch), "gpu", gpuID)
 			continue
 		}
 
@@ -210,24 +230,29 @@ func AMDGetGPUInfo() []GpuInfo {
 
 		// iGPU detection, remove this check once we can support an iGPU variant of the rocm library
 		if totalMemory < IGPUMemLimit {
-			slog.Info("amdgpu appears to be an iGPU, skipping", "gpu", gpuID, "total", format.HumanBytes2(totalMemory))
+			slog.Info("unsupported Radeon iGPU detected skipping", "id", gpuID, "total", format.HumanBytes2(totalMemory))
 			continue
 		}
+		var name string
+		// TODO - PCI ID lookup
+		if vendor > 0 && device > 0 {
+			name = fmt.Sprintf("%04x:%04x", vendor, device)
+		}
 
-		slog.Info("amdgpu memory", "gpu", gpuID, "total", format.HumanBytes2(totalMemory))
-		slog.Info("amdgpu memory", "gpu", gpuID, "available", format.HumanBytes2(totalMemory-usedMemory))
+		slog.Debug("amdgpu memory", "gpu", gpuID, "total", format.HumanBytes2(totalMemory))
+		slog.Debug("amdgpu memory", "gpu", gpuID, "available", format.HumanBytes2(totalMemory-usedMemory))
 		gpuInfo := GpuInfo{
 			Library: "rocm",
 			memInfo: memInfo{
 				TotalMemory: totalMemory,
 				FreeMemory:  (totalMemory - usedMemory),
 			},
-			ID: fmt.Sprintf("%d", gpuID),
-			// Name: not exposed in sysfs directly, would require pci device id lookup
-			Major:         int(major),
-			Minor:         int(minor),
-			Patch:         int(patch),
+			ID:            fmt.Sprintf("%d", gpuID),
+			Name:          name,
+			Compute:       fmt.Sprintf("gfx%d%x%x", major, minor, patch),
 			MinimumMemory: rocmMinimumMemory,
+			DriverMajor:   driverMajor,
+			DriverMinor:   driverMinor,
 		}
 
 		// If the user wants to filter to a subset of devices, filter out if we aren't a match
@@ -266,7 +291,7 @@ func AMDGetGPUInfo() []GpuInfo {
 				}
 				slog.Debug("rocm supported GPUs", "types", supported)
 			}
-			gfx := fmt.Sprintf("gfx%d%d%x", gpuInfo.Major, gpuInfo.Minor, gpuInfo.Patch)
+			gfx := gpuInfo.Compute
 			if !slices.Contains[[]string, string](supported, gfx) {
 				slog.Warn("amdgpu is not supported", "gpu", gpuInfo.ID, "gpu_type", gfx, "library", libDir, "supported_types", supported)
 				// TODO - consider discrete markdown just for ROCM troubleshooting?
@@ -276,7 +301,7 @@ func AMDGetGPUInfo() []GpuInfo {
 				slog.Info("amdgpu is supported", "gpu", gpuInfo.ID, "gpu_type", gfx)
 			}
 		} else {
-			slog.Debug("skipping rocm gfx compatibility check with HSA_OVERRIDE_GFX_VERSION=" + gfxOverride)
+			slog.Info("skipping rocm gfx compatibility check", "HSA_OVERRIDE_GFX_VERSION", gfxOverride)
 		}
 
 		// The GPU has passed all the verification steps and is supported
@@ -322,19 +347,34 @@ func AMDValidateLibDir() (string, error) {
 	return "", fmt.Errorf("no suitable rocm found, falling back to CPU")
 }
 
-func AMDDriverVersion() (string, error) {
-	_, err := os.Stat(DriverVersionFile)
+func AMDDriverVersion() (driverMajor, driverMinor int, err error) {
+	_, err = os.Stat(DriverVersionFile)
 	if err != nil {
-		return "", fmt.Errorf("amdgpu version file missing: %s %w", DriverVersionFile, err)
+		return 0, 0, fmt.Errorf("amdgpu version file missing: %s %w", DriverVersionFile, err)
 	}
 	fp, err := os.Open(DriverVersionFile)
 	if err != nil {
-		return "", err
+		return 0, 0, err
 	}
 	defer fp.Close()
 	verString, err := io.ReadAll(fp)
 	if err != nil {
-		return "", err
+		return 0, 0, err
 	}
-	return strings.TrimSpace(string(verString)), nil
+
+	pattern := `\A(\d+)\.(\d+).*`
+	regex := regexp.MustCompile(pattern)
+	match := regex.FindStringSubmatch(string(verString))
+	if len(match) < 2 {
+		return 0, 0, fmt.Errorf("malformed version string %s", string(verString))
+	}
+	driverMajor, err = strconv.Atoi(match[1])
+	if err != nil {
+		return 0, 0, err
+	}
+	driverMinor, err = strconv.Atoi(match[2])
+	if err != nil {
+		return 0, 0, err
+	}
+	return driverMajor, driverMinor, nil
 }

gpu/amd_windows.go

@@ -7,7 +7,6 @@ import (
 	"os"
 	"path/filepath"
 	"slices"
-	"strconv"
 	"strings"
 
 	"github.com/ollama/ollama/format"
@@ -34,13 +33,12 @@ func AMDGetGPUInfo() []GpuInfo {
 	}
 	defer hl.Release()
 
-	ver, err := hl.AMDDriverVersion()
-	if err == nil {
-		slog.Info("AMD Driver: " + ver)
-	} else {
-		// For now this is benign, but we may eventually need to fail compatibility checks
-		slog.Debug("error looking up amd driver version", "error", err)
-	}
+	// TODO - this reports incorrect version information, so omitting for now
+	// driverMajor, driverMinor, err := hl.AMDDriverVersion()
+	// if err != nil {
+	// 	// For now this is benign, but we may eventually need to fail compatibility checks
+	// 	slog.Debug("error looking up amd driver version", "error", err)
+	// }
 
 	// Note: the HIP library automatically handles subsetting to any HIP_VISIBLE_DEVICES the user specified
 	count := hl.HipGetDeviceCount()
@@ -62,10 +60,10 @@ func AMDGetGPUInfo() []GpuInfo {
 			return nil
 		}
 	} else {
-		slog.Debug("skipping rocm gfx compatibility check with HSA_OVERRIDE_GFX_VERSION=" + gfxOverride)
+		slog.Info("skipping rocm gfx compatibility check", "HSA_OVERRIDE_GFX_VERSION", gfxOverride)
 	}
 
-	slog.Info("detected hip devices", "count", count)
+	slog.Debug("detected hip devices", "count", count)
 	// TODO how to determine the underlying device ID when visible devices is causing this to subset?
 	for i := 0; i < count; i++ {
 		err = hl.HipSetDevice(i)
@@ -85,18 +83,11 @@ func AMDGetGPUInfo() []GpuInfo {
 		// Can luid be used on windows for setting visible devices (and is it actually set?)
 		n = bytes.IndexByte(props.GcnArchName[:], 0)
 		gfx := string(props.GcnArchName[:n])
-		slog.Info("hip device", "id", i, "name", name, "gfx", gfx)
-		var major, minor, patch string
-		switch len(gfx) {
-		case 6:
-			major, minor, patch = gfx[3:4], gfx[4:5], gfx[5:]
-		case 7:
-			major, minor, patch = gfx[3:5], gfx[5:6], gfx[6:]
-		}
+		slog.Debug("hip device", "id", i, "name", name, "gfx", gfx)
 		//slog.Info(fmt.Sprintf("[%d] Integrated: %d", i, props.iGPU)) // DOESN'T REPORT CORRECTLY!  Always 0
 		// TODO  Why isn't props.iGPU accurate!?
 		if strings.EqualFold(name, iGPUName) {
-			slog.Info("iGPU detected skipping", "id", i)
+			slog.Info("unsupported Radeon iGPU detected skipping", "id", i, "name", name, "gfx", gfx)
 			continue
 		}
 		if gfxOverride == "" {
@@ -106,7 +97,7 @@ func AMDGetGPUInfo() []GpuInfo {
 				slog.Warn("See https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md for HSA_OVERRIDE_GFX_VERSION usage")
 				continue
 			} else {
-				slog.Info("amdgpu is supported", "gpu", i, "gpu_type", gfx)
+				slog.Debug("amdgpu is supported", "gpu", i, "gpu_type", gfx)
 			}
 		}
 
@@ -124,8 +115,8 @@ func AMDGetGPUInfo() []GpuInfo {
 
 		// TODO revisit this once ROCm v6 is available on windows.
 		// v5.7 only reports VRAM used by this process, so it's completely wrong and unusable
-		slog.Info("amdgpu memory", "gpu", i, "total", format.HumanBytes2(totalMemory))
-		slog.Info("amdgpu memory", "gpu", i, "available", format.HumanBytes2(freeMemory))
+		slog.Debug("amdgpu memory", "gpu", i, "total", format.HumanBytes2(totalMemory))
+		slog.Debug("amdgpu memory", "gpu", i, "available", format.HumanBytes2(freeMemory))
 		gpuInfo := GpuInfo{
 			Library: "rocm",
 			memInfo: memInfo{
@@ -135,31 +126,12 @@ func AMDGetGPUInfo() []GpuInfo {
 			ID:             fmt.Sprintf("%d", i), // TODO this is probably wrong if we specify visible devices
 			DependencyPath: libDir,
 			MinimumMemory:  rocmMinimumMemory,
-		}
-		if major != "" {
-			gpuInfo.Major, err = strconv.Atoi(major)
-			if err != nil {
-				slog.Info("failed to parse version", "version", gfx, "error", err)
-			}
-		}
-		if minor != "" {
-			gpuInfo.Minor, err = strconv.Atoi(minor)
-			if err != nil {
-				slog.Info("failed to parse version", "version", gfx, "error", err)
-			}
-		}
-		if patch != "" {
-			// Patch rev is hex; e.g. gfx90a
-			p, err := strconv.ParseInt(patch, 16, 0)
-			if err != nil {
-				slog.Info("failed to parse version", "version", gfx, "error", err)
-			} else {
-				gpuInfo.Patch = int(p)
-			}
-		}
-		if gpuInfo.Major < RocmComputeMin {
-			slog.Warn(fmt.Sprintf("amdgpu [%s] too old gfx%d%d%x", gpuInfo.ID, gpuInfo.Major, gpuInfo.Minor, gpuInfo.Patch))
-			continue
+			Name:           name,
+			Compute:        gfx,
+
+			// TODO - this information isn't accurate on windows, so don't report it until we find the right way to retrieve
+			// DriverMajor:    driverMajor,
+			// DriverMinor:    driverMinor,
 		}
 
 		resp = append(resp, gpuInfo)

gpu/assets.go

@@ -13,7 +13,7 @@ import (
 	"syscall"
 	"time"
 
-	"github.com/ollama/ollama/server/envconfig"
+	"github.com/ollama/ollama/envconfig"
 )
 
 var (

gpu/cpu_common.go

@@ -8,14 +8,14 @@ import (
 
 func GetCPUVariant() string {
 	if cpu.X86.HasAVX2 {
-		slog.Info("CPU has AVX2")
+		slog.Debug("CPU has AVX2")
 		return "avx2"
 	}
 	if cpu.X86.HasAVX {
-		slog.Info("CPU has AVX")
+		slog.Debug("CPU has AVX")
 		return "avx"
 	}
-	slog.Info("CPU does not have vector extensions")
+	slog.Debug("CPU does not have vector extensions")
 	// else LCD
 	return ""
 }

gpu/gpu.go

@@ -16,23 +16,25 @@ import (
 	"os"
 	"path/filepath"
 	"runtime"
+	"strconv"
 	"strings"
 	"sync"
 	"unsafe"
 
 	"github.com/ollama/ollama/format"
-	"github.com/ollama/ollama/server/envconfig"
+	"github.com/ollama/ollama/envconfig"
 )
 
 type handles struct {
 	deviceCount int
 	cudart      *C.cudart_handle_t
 	nvcuda      *C.nvcuda_handle_t
+	oneapi      *C.oneapi_handle_t
 }
 
 const (
-	cudaMinimumMemory = 256 * format.MebiByte
-	rocmMinimumMemory = 256 * format.MebiByte
+	cudaMinimumMemory = 457 * format.MebiByte
+	rocmMinimumMemory = 457 * format.MebiByte
 )
 
 var gpuMutex sync.Mutex
@@ -80,6 +82,15 @@ var NvcudaWindowsGlobs = []string{
 	"c:\\windows\\system*\\nvcuda.dll",
 }
 
+var OneapiWindowsGlobs = []string{
+	"c:\\Windows\\System32\\DriverStore\\FileRepository\\*\\ze_intel_gpu64.dll",
+}
+
+var OneapiLinuxGlobs = []string{
+	"/usr/lib/x86_64-linux-gnu/libze_intel_gpu.so*",
+	"/usr/lib*/libze_intel_gpu.so*",
+}
+
 // Jetson devices have JETSON_JETPACK="x.y.z" factory set to the Jetpack version installed.
 // Included to drive logic for reducing Ollama-allocated overhead on L4T/Jetson devices.
 var CudaTegra string = os.Getenv("JETSON_JETPACK")
@@ -94,6 +105,8 @@ func initGPUHandles() *handles {
 	var cudartMgmtPatterns []string
 	var nvcudaMgmtName string
 	var nvcudaMgmtPatterns []string
+	var oneapiMgmtName string
+	var oneapiMgmtPatterns []string
 
 	tmpDir, _ := PayloadsDir()
 	switch runtime.GOOS {
@@ -105,6 +118,8 @@ func initGPUHandles() *handles {
 		// Aligned with driver, we can't carry as payloads
 		nvcudaMgmtName = "nvcuda.dll"
 		nvcudaMgmtPatterns = NvcudaWindowsGlobs
+		oneapiMgmtName = "ze_intel_gpu64.dll"
+		oneapiMgmtPatterns = OneapiWindowsGlobs
 	case "linux":
 		cudartMgmtName = "libcudart.so*"
 		if tmpDir != "" {
@@ -115,16 +130,18 @@ func initGPUHandles() *handles {
 		// Aligned with driver, we can't carry as payloads
 		nvcudaMgmtName = "libcuda.so*"
 		nvcudaMgmtPatterns = NvcudaLinuxGlobs
+		oneapiMgmtName = "libze_intel_gpu.so"
+		oneapiMgmtPatterns = OneapiLinuxGlobs
 	default:
 		return gpuHandles
 	}
 
-	slog.Info("Detecting GPUs")
+	slog.Debug("Detecting GPUs")
 	nvcudaLibPaths := FindGPULibs(nvcudaMgmtName, nvcudaMgmtPatterns)
 	if len(nvcudaLibPaths) > 0 {
 		deviceCount, nvcuda, libPath := LoadNVCUDAMgmt(nvcudaLibPaths)
 		if nvcuda != nil {
-			slog.Info("detected GPUs", "count", deviceCount, "library", libPath)
+			slog.Debug("detected GPUs", "count", deviceCount, "library", libPath)
 			gpuHandles.nvcuda = nvcuda
 			gpuHandles.deviceCount = deviceCount
 			return gpuHandles
@@ -135,12 +152,24 @@ func initGPUHandles() *handles {
 	if len(cudartLibPaths) > 0 {
 		deviceCount, cudart, libPath := LoadCUDARTMgmt(cudartLibPaths)
 		if cudart != nil {
-			slog.Info("detected GPUs", "library", libPath, "count", deviceCount)
+			slog.Debug("detected GPUs", "library", libPath, "count", deviceCount)
 			gpuHandles.cudart = cudart
 			gpuHandles.deviceCount = deviceCount
 			return gpuHandles
 		}
 	}
+
+	oneapiLibPaths := FindGPULibs(oneapiMgmtName, oneapiMgmtPatterns)
+	if len(oneapiLibPaths) > 0 {
+		deviceCount, oneapi, libPath := LoadOneapiMgmt(oneapiLibPaths)
+		if oneapi != nil {
+			slog.Debug("detected Intel GPUs", "library", libPath, "count", deviceCount)
+			gpuHandles.oneapi = oneapi
+			gpuHandles.deviceCount = deviceCount
+			return gpuHandles
+		}
+	}
+
 	return gpuHandles
 }
 
@@ -181,33 +210,53 @@ func GetGPUInfo() GpuInfoList {
 		if cpuVariant == "" && runtime.GOARCH == "amd64" {
 			continue
 		}
-		gpuInfo := GpuInfo{
-			Library: "cuda",
-		}
-		if gpuHandles.cudart != nil {
-			C.cudart_check_vram(*gpuHandles.cudart, C.int(i), &memInfo)
-		} else {
-			C.nvcuda_check_vram(*gpuHandles.nvcuda, C.int(i), &memInfo)
-		}
-		if memInfo.err != nil {
-			slog.Info("error looking up nvidia GPU memory", "error", C.GoString(memInfo.err))
-			C.free(unsafe.Pointer(memInfo.err))
-			continue
-		}
-		if memInfo.major < CudaComputeMin[0] || (memInfo.major == CudaComputeMin[0] && memInfo.minor < CudaComputeMin[1]) {
-			slog.Info(fmt.Sprintf("[%d] CUDA GPU is too old. Compute Capability detected: %d.%d", i, memInfo.major, memInfo.minor))
-			continue
-		}
-		gpuInfo.TotalMemory = uint64(memInfo.total)
-		gpuInfo.FreeMemory = uint64(memInfo.free)
-		gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
-		gpuInfo.Major = int(memInfo.major)
-		gpuInfo.Minor = int(memInfo.minor)
-		gpuInfo.MinimumMemory = cudaMinimumMemory
-		gpuInfo.DependencyPath = depPath
+		if gpuHandles.cudart != nil || gpuHandles.nvcuda != nil {
+			gpuInfo := GpuInfo{
+				Library: "cuda",
+			}
+			var driverMajor int
+			var driverMinor int
+			if gpuHandles.cudart != nil {
+				C.cudart_check_vram(*gpuHandles.cudart, C.int(i), &memInfo)
+			} else {
+				C.nvcuda_check_vram(*gpuHandles.nvcuda, C.int(i), &memInfo)
+				driverMajor = int(gpuHandles.nvcuda.driver_major)
+				driverMinor = int(gpuHandles.nvcuda.driver_minor)
+			}
+			if memInfo.err != nil {
+				slog.Info("error looking up nvidia GPU memory", "error", C.GoString(memInfo.err))
+				C.free(unsafe.Pointer(memInfo.err))
+				continue
+			}
+			if memInfo.major < CudaComputeMin[0] || (memInfo.major == CudaComputeMin[0] && memInfo.minor < CudaComputeMin[1]) {
+				slog.Info(fmt.Sprintf("[%d] CUDA GPU is too old. Compute Capability detected: %d.%d", i, memInfo.major, memInfo.minor))
+				continue
+			}
+			gpuInfo.TotalMemory = uint64(memInfo.total)
+			gpuInfo.FreeMemory = uint64(memInfo.free)
+			gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
+			gpuInfo.Compute = fmt.Sprintf("%d.%d", memInfo.major, memInfo.minor)
+			gpuInfo.MinimumMemory = cudaMinimumMemory
+			gpuInfo.DependencyPath = depPath
+			gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
+			gpuInfo.DriverMajor = int(driverMajor)
+			gpuInfo.DriverMinor = int(driverMinor)
 
-		// TODO potentially sort on our own algorithm instead of what the underlying GPU library does...
-		resp = append(resp, gpuInfo)
+			// TODO potentially sort on our own algorithm instead of what the underlying GPU library does...
+			resp = append(resp, gpuInfo)
+		}
+		if gpuHandles.oneapi != nil {
+			gpuInfo := GpuInfo{
+				Library: "oneapi",
+			}
+			C.oneapi_check_vram(*gpuHandles.oneapi, &memInfo)
+			var totalFreeMem float64 = float64(memInfo.free) * 0.95 // work-around: leave some reserve vram for mkl lib used in ggml-sycl backend.
+			memInfo.free = C.uint64_t(totalFreeMem)
+			gpuInfo.TotalMemory = uint64(memInfo.total)
+			gpuInfo.FreeMemory = uint64(memInfo.free)
+			gpuInfo.ID = strconv.Itoa(i)
+			resp = append(resp, gpuInfo)
+		}
 	}
 
 	// Then AMD
@@ -342,6 +391,23 @@ func LoadNVCUDAMgmt(nvcudaLibPaths []string) (int, *C.nvcuda_handle_t, string) {
 	return 0, nil, ""
 }
 
+func LoadOneapiMgmt(oneapiLibPaths []string) (int, *C.oneapi_handle_t, string) {
+	var resp C.oneapi_init_resp_t
+	resp.oh.verbose = getVerboseState()
+	for _, libPath := range oneapiLibPaths {
+		lib := C.CString(libPath)
+		defer C.free(unsafe.Pointer(lib))
+		C.oneapi_init(lib, &resp)
+		if resp.err != nil {
+			slog.Debug("Unable to load oneAPI management library", "library", libPath, "error", C.GoString(resp.err))
+			C.free(unsafe.Pointer(resp.err))
+		} else {
+			return int(resp.num_devices), &resp.oh, libPath
+		}
+	}
+	return 0, nil, ""
+}
+
 func getVerboseState() C.uint16_t {
 	if envconfig.Debug {
 		return C.uint16_t(1)
@@ -362,6 +428,8 @@ func (l GpuInfoList) GetVisibleDevicesEnv() (string, string) {
 		return cudaGetVisibleDevicesEnv(l)
 	case "rocm":
 		return rocmGetVisibleDevicesEnv(l)
+	case "oneapi":
+		return oneapiGetVisibleDevicesEnv(l)
 	default:
 		slog.Debug("no filter required for library " + l[0].Library)
 		return "", ""

gpu/gpu_darwin.go

@@ -15,7 +15,7 @@ import (
 )
 
 const (
-	metalMinimumMemory = 384 * format.MebiByte
+	metalMinimumMemory = 512 * format.MebiByte
 )
 
 func GetGPUInfo() GpuInfoList {

gpu/gpu_info.h

@@ -39,16 +39,19 @@ extern "C" {
 #endif
 
 #define GPU_ID_LEN 64
+#define GPU_NAME_LEN 96
 
 typedef struct mem_info {
   char *err;  // If non-nill, caller responsible for freeing
   char gpu_id[GPU_ID_LEN];
+  char gpu_name[GPU_NAME_LEN];
   uint64_t total;
   uint64_t free;
 
   // Compute Capability
   int major; 
   int minor;
+  int patch;
 } mem_info_t;
 
 void cpu_check_ram(mem_info_t *resp);
@@ -59,6 +62,7 @@ void cpu_check_ram(mem_info_t *resp);
 
 #include "gpu_info_cudart.h"
 #include "gpu_info_nvcuda.h"
+#include "gpu_info_oneapi.h"
 
 #endif  // __GPU_INFO_H__
 #endif  // __APPLE__

gpu/gpu_info_cpu.c

@@ -10,8 +10,6 @@ void cpu_check_ram(mem_info_t *resp) {
   if (GlobalMemoryStatusEx(&info) != 0) {
     resp->total = info.ullTotalPhys;
     resp->free = info.ullAvailPhys;
-    resp->major = 0;
-    resp->minor = 0;
     snprintf(&resp->gpu_id[0], GPU_ID_LEN, "0");
   } else {
     resp->err = LOAD_ERR();
@@ -31,8 +29,6 @@ void cpu_check_ram(mem_info_t *resp) {
   } else {
     resp->total = info.totalram * info.mem_unit;
     resp->free = info.freeram * info.mem_unit;
-    resp->major = 0;
-    resp->minor = 0;
     snprintf(&resp->gpu_id[0], GPU_ID_LEN, "0");
   }
   return;

gpu/gpu_info_nvcuda.c

@@ -22,6 +22,7 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
       {"cuDeviceGet", (void *)&resp->ch.cuDeviceGet},
       {"cuDeviceGetAttribute", (void *)&resp->ch.cuDeviceGetAttribute},
       {"cuDeviceGetUuid", (void *)&resp->ch.cuDeviceGetUuid},
+      {"cuDeviceGetName", (void *)&resp->ch.cuDeviceGetName},
       {"cuCtxCreate_v3", (void *)&resp->ch.cuCtxCreate_v3},
       {"cuMemGetInfo_v2", (void *)&resp->ch.cuMemGetInfo_v2},
       {"cuCtxDestroy", (void *)&resp->ch.cuCtxDestroy},
@@ -70,18 +71,17 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
   }
 
   int version = 0;
-  nvcudaDriverVersion_t driverVersion;
-  driverVersion.major = 0;
-  driverVersion.minor = 0;
+  resp->ch.driver_major = 0;
+  resp->ch.driver_minor = 0;
 
   // Report driver version if we're in verbose mode, ignore errors
   ret = (*resp->ch.cuDriverGetVersion)(&version);
   if (ret != CUDA_SUCCESS) {
     LOG(resp->ch.verbose, "cuDriverGetVersion failed: %d\n", ret);
   } else {
-    driverVersion.major = version / 1000;
-    driverVersion.minor = (version - (driverVersion.major * 1000)) / 10;
-    LOG(resp->ch.verbose, "CUDA driver version: %d-%d\n", driverVersion.major, driverVersion.minor);
+    resp->ch.driver_major = version / 1000;
+    resp->ch.driver_minor = (version - (resp->ch.driver_major * 1000)) / 10;
+    LOG(resp->ch.verbose, "CUDA driver version: %d.%d\n", resp->ch.driver_major, resp->ch.driver_minor);
   }
 
   ret = (*resp->ch.cuDeviceGetCount)(&resp->num_devices);
@@ -117,8 +117,6 @@ void nvcuda_check_vram(nvcuda_handle_t h, int i, mem_info_t *resp) {
     return;
   }
 
-  resp->major = 0;
-  resp->minor = 0;
   int major = 0;
   int minor = 0;
   ret = (*h.cuDeviceGetAttribute)(&major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, device);
@@ -161,6 +159,12 @@ void nvcuda_check_vram(nvcuda_handle_t h, int i, mem_info_t *resp) {
       );
   }
 
+  ret = (*h.cuDeviceGetName)(&resp->gpu_name[0], GPU_NAME_LEN, device);
+  if (ret != CUDA_SUCCESS) {
+    LOG(h.verbose, "[%d] device name lookup failure: %d\n", i, ret);
+    resp->gpu_name[0] = '\0';
+  }
+
   // To get memory we have to set (and release) a context
   ret = (*h.cuCtxCreate_v3)(&ctx, NULL, 0, 0, device);
   if (ret != CUDA_SUCCESS) {

gpu/gpu_info_nvcuda.h

@@ -44,12 +44,15 @@ typedef void* CUcontext;
 typedef struct nvcuda_handle {
   void *handle;
   uint16_t verbose;
+  int driver_major;
+  int driver_minor;
   CUresult (*cuInit)(unsigned int Flags);
   CUresult (*cuDriverGetVersion)(int *driverVersion);
   CUresult (*cuDeviceGetCount)(int *);
   CUresult (*cuDeviceGet)(CUdevice* device, int ordinal);
   CUresult (*cuDeviceGetAttribute)(int* pi, CUdevice_attribute attrib, CUdevice dev);
   CUresult (*cuDeviceGetUuid)(CUuuid* uuid, CUdevice dev); // signature compatible with cuDeviceGetUuid_v2
+  CUresult (*cuDeviceGetName)(char *name, int len, CUdevice dev);
 
   // Context specific aspects
   CUresult (*cuCtxCreate_v3)(CUcontext* pctx, void *params, int len, unsigned int flags, CUdevice dev);

gpu/gpu_info_oneapi.c

@@ -0,0 +1,214 @@
+#ifndef __APPLE__
+
+#include "gpu_info_oneapi.h"
+
+#include <string.h>
+
+void oneapi_init(char *oneapi_lib_path, oneapi_init_resp_t *resp)
+{
+  ze_result_t ret;
+  resp->err = NULL;
+  const int buflen = 256;
+  char buf[buflen + 1];
+  int i;
+  struct lookup
+  {
+    char *s;
+    void **p;
+  } l[] = {
+      {"zesInit", (void *)&resp->oh.zesInit},
+      {"zesDriverGet", (void *)&resp->oh.zesDriverGet},
+      {"zesDeviceGet", (void *)&resp->oh.zesDeviceGet},
+      {"zesDeviceGetProperties", (void *)&resp->oh.zesDeviceGetProperties},
+      {"zesDeviceEnumMemoryModules",
+       (void *)&resp->oh.zesDeviceEnumMemoryModules},
+      {"zesMemoryGetProperties", (void *)&resp->oh.zesMemoryGetProperties},
+      {"zesMemoryGetState", (void *)&resp->oh.zesMemoryGetState},
+      {NULL, NULL},
+  };
+
+  resp->oh.handle = LOAD_LIBRARY(oneapi_lib_path, RTLD_LAZY);
+  if (!resp->oh.handle)
+  {
+    char *msg = LOAD_ERR();
+    snprintf(buf, buflen,
+             "Unable to load %s library to query for Intel GPUs: %s\n",
+             oneapi_lib_path, msg);
+    free(msg);
+    resp->err = strdup(buf);
+    return;
+  }
+
+  // TODO once we've squashed the remaining corner cases remove this log
+  LOG(resp->oh.verbose,
+      "wiring Level-Zero management library functions in %s\n",
+      oneapi_lib_path);
+
+  for (i = 0; l[i].s != NULL; i++)
+  {
+    // TODO once we've squashed the remaining corner cases remove this log
+    LOG(resp->oh.verbose, "dlsym: %s\n", l[i].s);
+
+    *l[i].p = LOAD_SYMBOL(resp->oh.handle, l[i].s);
+    if (!l[i].p)
+    {
+      resp->oh.handle = NULL;
+      char *msg = LOAD_ERR();
+      LOG(resp->oh.verbose, "dlerr: %s\n", msg);
+      UNLOAD_LIBRARY(resp->oh.handle);
+      snprintf(buf, buflen, "symbol lookup for %s failed: %s", l[i].s, msg);
+      free(msg);
+      resp->err = strdup(buf);
+      return;
+    }
+  }
+
+  ret = (*resp->oh.zesInit)(0);
+  if (ret != ZE_RESULT_SUCCESS)
+  {
+    LOG(resp->oh.verbose, "zesInit err: %d\n", ret);
+    UNLOAD_LIBRARY(resp->oh.handle);
+    resp->oh.handle = NULL;
+    snprintf(buf, buflen, "oneapi vram init failure: %d", ret);
+    resp->err = strdup(buf);
+  }
+
+  (*resp->oh.zesDriverGet)(&resp->num_devices, NULL);
+
+  return;
+}
+
+void oneapi_check_vram(oneapi_handle_t h, mem_info_t *resp)
+{
+  ze_result_t ret;
+  resp->err = NULL;
+  uint64_t totalMem = 0;
+  uint64_t usedMem = 0;
+  const int buflen = 256;
+  char buf[buflen + 1];
+  int i, d, m;
+
+  if (h.handle == NULL)
+  {
+    resp->err = strdup("Level-Zero handle not initialized");
+    return;
+  }
+
+  uint32_t driversCount = 0;
+  ret = (*h.zesDriverGet)(&driversCount, NULL);
+  if (ret != ZE_RESULT_SUCCESS)
+  {
+    snprintf(buf, buflen, "unable to get driver count: %d", ret);
+    resp->err = strdup(buf);
+    return;
+  }
+  LOG(h.verbose, "discovered %d Level-Zero drivers\n", driversCount);
+
+  zes_driver_handle_t *allDrivers =
+      malloc(driversCount * sizeof(zes_driver_handle_t));
+  (*h.zesDriverGet)(&driversCount, allDrivers);
+
+  resp->total = 0;
+  resp->free = 0;
+
+  for (d = 0; d < driversCount; d++)
+  {
+    uint32_t deviceCount = 0;
+    ret = (*h.zesDeviceGet)(allDrivers[d], &deviceCount, NULL);
+    if (ret != ZE_RESULT_SUCCESS)
+    {
+      snprintf(buf, buflen, "unable to get device count: %d", ret);
+      resp->err = strdup(buf);
+      free(allDrivers);
+      return;
+    }
+
+    LOG(h.verbose, "discovered %d Level-Zero devices\n", deviceCount);
+
+    zes_device_handle_t *devices =
+        malloc(deviceCount * sizeof(zes_device_handle_t));
+    (*h.zesDeviceGet)(allDrivers[d], &deviceCount, devices);
+
+    for (i = 0; i < deviceCount; i++)
+    {
+      zes_device_ext_properties_t ext_props;
+      ext_props.stype = ZES_STRUCTURE_TYPE_DEVICE_EXT_PROPERTIES;
+      ext_props.pNext = NULL;
+
+      zes_device_properties_t props;
+      props.stype = ZES_STRUCTURE_TYPE_DEVICE_PROPERTIES;
+      props.pNext = &ext_props;
+
+      ret = (*h.zesDeviceGetProperties)(devices[i], &props);
+      if (ret != ZE_RESULT_SUCCESS)
+      {
+        snprintf(buf, buflen, "unable to get device properties: %d", ret);
+        resp->err = strdup(buf);
+        free(allDrivers);
+        free(devices);
+        return;
+      }
+
+      if (h.verbose)
+      {
+        // When in verbose mode, report more information about
+        // the card we discover.
+        LOG(h.verbose, "[%d] oneAPI device name: %s\n", i,
+            props.modelName);
+        LOG(h.verbose, "[%d] oneAPI brand: %s\n", i,
+            props.brandName);
+        LOG(h.verbose, "[%d] oneAPI vendor: %s\n", i,
+            props.vendorName);
+        LOG(h.verbose, "[%d] oneAPI S/N: %s\n", i,
+            props.serialNumber);
+        LOG(h.verbose, "[%d] oneAPI board number: %s\n", i,
+            props.boardNumber);
+      }
+
+      uint32_t memCount = 0;
+      ret = (*h.zesDeviceEnumMemoryModules)(devices[i], &memCount, NULL);
+      if (ret != ZE_RESULT_SUCCESS)
+      {
+        snprintf(buf, buflen,
+                 "unable to enumerate Level-Zero memory modules: %d", ret);
+        resp->err = strdup(buf);
+        free(allDrivers);
+        free(devices);
+        return;
+      }
+
+      LOG(h.verbose, "discovered %d Level-Zero memory modules\n", memCount);
+
+      zes_mem_handle_t *mems = malloc(memCount * sizeof(zes_mem_handle_t));
+      (*h.zesDeviceEnumMemoryModules)(devices[i], &memCount, mems);
+
+      for (m = 0; m < memCount; m++)
+      {
+        zes_mem_state_t state;
+        state.stype = ZES_STRUCTURE_TYPE_MEM_STATE;
+        state.pNext = NULL;
+        ret = (*h.zesMemoryGetState)(mems[m], &state);
+        if (ret != ZE_RESULT_SUCCESS)
+        {
+          snprintf(buf, buflen, "unable to get memory state: %d", ret);
+          resp->err = strdup(buf);
+          free(allDrivers);
+          free(devices);
+          free(mems);
+          return;
+        }
+
+        resp->total += state.size;
+        resp->free += state.free;
+      }
+
+      free(mems);
+    }
+
+    free(devices);
+  }
+
+  free(allDrivers);
+}
+
+#endif // __APPLE__

gpu/gpu_info_oneapi.h

@@ -0,0 +1,211 @@
+#ifndef __APPLE__
+#ifndef __GPU_INFO_ONEAPI_H__
+#define __GPU_INFO_ONEAPI_H__
+#include "gpu_info.h"
+
+#define ZE_MAX_DEVICE_NAME 256
+#define ZE_MAX_DEVICE_UUID_SIZE 16
+#define ZES_STRING_PROPERTY_SIZE 64
+#define ZE_BIT(_i) (1 << _i)
+
+// Just enough typedef's to dlopen/dlsym for memory information
+typedef enum ze_result_t
+{
+  ZE_RESULT_SUCCESS = 0,
+  // Other values omitted for now...
+} ze_result_t;
+
+typedef uint8_t ze_bool_t;
+typedef struct _zes_driver_handle_t *zes_driver_handle_t;
+typedef struct _zes_device_handle_t *zes_device_handle_t;
+typedef struct _zes_mem_handle_t *zes_mem_handle_t;
+
+typedef enum _ze_structure_type_t
+{
+  ZE_STRUCTURE_TYPE_FORCE_UINT32 = 0x7fffffff
+} ze_structure_type_t;
+
+typedef enum _zes_structure_type_t
+{
+  ZES_STRUCTURE_TYPE_DEVICE_PROPERTIES = 0x1,
+  ZES_STRUCTURE_TYPE_MEM_PROPERTIES = 0xb,
+  ZES_STRUCTURE_TYPE_MEM_STATE = 0x1e,
+  ZES_STRUCTURE_TYPE_DEVICE_EXT_PROPERTIES = 0x2d,
+  ZES_STRUCTURE_TYPE_FORCE_UINT32 = 0x7fffffff
+} zes_structure_type_t;
+
+typedef enum _zes_mem_type_t
+{
+  ZES_MEM_TYPE_FORCE_UINT32 = 0x7fffffff
+} zes_mem_type_t;
+
+typedef enum _zes_mem_loc_t
+{
+  ZES_MEM_LOC_SYSTEM = 0,
+  ZES_MEM_LOC_DEVICE = 1,
+  ZES_MEM_LOC_FORCE_UINT32 = 0x7fffffff
+} zes_mem_loc_t;
+
+typedef enum _zes_mem_health_t
+{
+  ZES_MEM_HEALTH_FORCE_UINT32 = 0x7fffffff
+} zes_mem_health_t;
+
+typedef struct _ze_device_uuid_t
+{
+  uint8_t id[ZE_MAX_DEVICE_UUID_SIZE];
+} ze_device_uuid_t;
+
+typedef struct _zes_uuid_t
+{
+  uint8_t id[ZE_MAX_DEVICE_UUID_SIZE];
+} zes_uuid_t;
+
+typedef enum _ze_device_type_t
+{
+  ZE_DEVICE_TYPE_GPU = 1,
+  ZE_DEVICE_TYPE_CPU = 2,
+  ZE_DEVICE_TYPE_FPGA = 3,
+  ZE_DEVICE_TYPE_MCA = 4,
+  ZE_DEVICE_TYPE_VPU = 5,
+  ZE_DEVICE_TYPE_FORCE_UINT32 = 0x7fffffff
+} ze_device_type_t;
+
+typedef enum _zes_device_type_t
+{
+  ZES_DEVICE_TYPE_GPU = 1,
+  ZES_DEVICE_TYPE_CPU = 2,
+  ZES_DEVICE_TYPE_FPGA = 3,
+  ZES_DEVICE_TYPE_MCA = 4,
+  ZES_DEVICE_TYPE_VPU = 5,
+  ZES_DEVICE_TYPE_FORCE_UINT32 = 0x7fffffff
+} zes_device_type_t;
+
+typedef uint32_t ze_device_property_flags_t;
+typedef enum _ze_device_property_flag_t
+{
+  ZE_DEVICE_PROPERTY_FLAG_INTEGRATED = ZE_BIT(0),
+  ZE_DEVICE_PROPERTY_FLAG_SUBDEVICE = ZE_BIT(1),
+  ZE_DEVICE_PROPERTY_FLAG_ECC = ZE_BIT(2),
+  ZE_DEVICE_PROPERTY_FLAG_ONDEMANDPAGING = ZE_BIT(3),
+  ZE_DEVICE_PROPERTY_FLAG_FORCE_UINT32 = 0x7fffffff
+} ze_device_property_flag_t;
+
+typedef uint32_t zes_device_property_flags_t;
+typedef enum _zes_device_property_flag_t
+{
+  ZES_DEVICE_PROPERTY_FLAG_INTEGRATED = ZE_BIT(0),
+  ZES_DEVICE_PROPERTY_FLAG_SUBDEVICE = ZE_BIT(1),
+  ZES_DEVICE_PROPERTY_FLAG_ECC = ZE_BIT(2),
+  ZES_DEVICE_PROPERTY_FLAG_ONDEMANDPAGING = ZE_BIT(3),
+  ZES_DEVICE_PROPERTY_FLAG_FORCE_UINT32 = 0x7fffffff
+} zes_device_property_flag_t;
+
+typedef struct _ze_device_properties_t
+{
+  ze_structure_type_t stype;
+  void *pNext;
+  ze_device_type_t type;
+  uint32_t vendorId;
+  uint32_t deviceId;
+  ze_device_property_flags_t flags;
+  uint32_t subdeviceId;
+  uint32_t coreClockRate;
+  uint64_t maxMemAllocSize;
+  uint32_t maxHardwareContexts;
+  uint32_t maxCommandQueuePriority;
+  uint32_t numThreadsPerEU;
+  uint32_t physicalEUSimdWidth;
+  uint32_t numEUsPerSubslice;
+  uint32_t numSubslicesPerSlice;
+  uint32_t numSlices;
+  uint64_t timerResolution;
+  uint32_t timestampValidBits;
+  uint32_t kernelTimestampValidBits;
+  ze_device_uuid_t uuid;
+  char name[ZE_MAX_DEVICE_NAME];
+} ze_device_properties_t;
+
+typedef struct _zes_device_properties_t
+{
+  zes_structure_type_t stype;
+  void *pNext;
+  ze_device_properties_t core;
+  uint32_t numSubdevices;
+  char serialNumber[ZES_STRING_PROPERTY_SIZE];
+  char boardNumber[ZES_STRING_PROPERTY_SIZE];
+  char brandName[ZES_STRING_PROPERTY_SIZE];
+  char modelName[ZES_STRING_PROPERTY_SIZE];
+  char vendorName[ZES_STRING_PROPERTY_SIZE];
+  char driverVersion[ZES_STRING_PROPERTY_SIZE];
+} zes_device_properties_t;
+
+typedef struct _zes_device_ext_properties_t
+{
+  zes_structure_type_t stype;
+  void *pNext;
+  zes_uuid_t uuid;
+  zes_device_type_t type;
+  zes_device_property_flags_t flags;
+} zes_device_ext_properties_t;
+
+typedef struct _zes_mem_properties_t
+{
+  zes_structure_type_t stype;
+  void *pNext;
+  zes_mem_type_t type;
+  ze_bool_t onSubdevice;
+  uint32_t subdeviceId;
+  zes_mem_loc_t location;
+  uint64_t physicalSize;
+  int32_t busWidth;
+  int32_t numChannels;
+} zes_mem_properties_t;
+
+typedef struct _zes_mem_state_t
+{
+  zes_structure_type_t stype;
+  const void *pNext;
+  zes_mem_health_t health;
+  uint64_t free;
+  uint64_t size;
+} zes_mem_state_t;
+
+typedef struct oneapi_handle
+{
+  void *handle;
+  uint16_t verbose;
+  ze_result_t (*zesInit)(int);
+  ze_result_t (*zesDriverGet)(uint32_t *pCount, zes_driver_handle_t *phDrivers);
+  ze_result_t (*zesDeviceGet)(zes_driver_handle_t hDriver, uint32_t *pCount,
+                              zes_device_handle_t *phDevices);
+  ze_result_t (*zesDeviceGetProperties)(zes_device_handle_t hDevice,
+                                        zes_device_properties_t *pProperties);
+  ze_result_t (*zesDeviceEnumMemoryModules)(zes_device_handle_t hDevice,
+                                            uint32_t *pCount,
+                                            zes_mem_handle_t *phMemory);
+  ze_result_t (*zesMemoryGetProperties)(zes_mem_handle_t hMemory,
+                                        zes_mem_properties_t *pProperties);
+  ze_result_t (*zesMemoryGetState)(zes_mem_handle_t hMemory,
+                                   zes_mem_state_t *pState);
+
+} oneapi_handle_t;
+
+typedef struct oneapi_init_resp
+{
+  char *err; // If err is non-null handle is invalid
+  int num_devices;
+  oneapi_handle_t oh;
+} oneapi_init_resp_t;
+
+typedef struct oneapi_version_resp
+{
+  ze_result_t status;
+  char *str; // Contains version or error string if status != 0
+} oneapi_version_resp_t;
+
+void oneapi_init(char *oneapi_lib_path, oneapi_init_resp_t *resp);
+void oneapi_check_vram(oneapi_handle_t rh, mem_info_t *resp);
+
+#endif // __GPU_INFO_INTEL_H__
+#endif // __APPLE__

gpu/gpu_oneapi.go

@@ -0,0 +1,21 @@
+//go:build linux || windows
+
+package gpu
+
+import (
+	"log/slog"
+	"strings"
+)
+
+func oneapiGetVisibleDevicesEnv(gpuInfo []GpuInfo) (string, string) {
+	ids := []string{}
+	for _, info := range gpuInfo {
+		if info.Library != "oneapi" {
+			// TODO shouldn't happen if things are wired correctly...
+			slog.Debug("oneapiGetVisibleDevicesEnv skipping over non-sycl device", "library", info.Library)
+			continue
+		}
+		ids = append(ids, info.ID)
+	}
+	return "ONEAPI_DEVICE_SELECTOR", "level_zero:" + strings.Join(ids, ",")
+}

gpu/types.go

@@ -1,5 +1,12 @@
 package gpu
 
+import (
+	"fmt"
+	"log/slog"
+
+	"github.com/ollama/ollama/format"
+)
+
 type memInfo struct {
 	TotalMemory uint64 `json:"total_memory,omitempty"`
 	FreeMemory  uint64 `json:"free_memory,omitempty"`
@@ -20,11 +27,13 @@ type GpuInfo struct {
 	DependencyPath string `json:"lib_path,omitempty"`
 
 	// GPU information
-	ID    string `json:"gpu_id"`          // string to use for selection of this specific GPU
-	Name  string `json:"name"`            // user friendly name if available
-	Major int    `json:"major,omitempty"` // Major compatibility version (CC or gfx)
-	Minor int    `json:"minor,omitempty"` // Minor compatibility version (CC or gfx)
-	Patch int    `json:"patch,omitempty"` // Patch compatibility only matters on AMD
+	ID      string `json:"gpu_id"`  // string to use for selection of this specific GPU
+	Name    string `json:"name"`    // user friendly name if available
+	Compute string `json:"compute"` // Compute Capability or gfx
+
+	// Driver Information - TODO no need to put this on each GPU
+	DriverMajor int `json:"driver_major,omitempty"`
+	DriverMinor int `json:"driver_minor,omitempty"`
 
 	// TODO other performance capability info to help in scheduling decisions
 }
@@ -56,6 +65,21 @@ func (l GpuInfoList) ByLibrary() []GpuInfoList {
 	return resp
 }
 
+// Report the GPU information into the log an Info level
+func (l GpuInfoList) LogDetails() {
+	for _, g := range l {
+		slog.Info("inference compute",
+			"id", g.ID,
+			"library", g.Library,
+			"compute", g.Compute,
+			"driver", fmt.Sprintf("%d.%d", g.DriverMajor, g.DriverMinor),
+			"name", g.Name,
+			"total", format.HumanBytes2(g.TotalMemory),
+			"available", format.HumanBytes2(g.FreeMemory),
+		)
+	}
+}
+
 // Sort by Free Space
 type ByFreeMemory []GpuInfo
 

integration/concurrency_test.go

@@ -217,7 +217,7 @@ func TestMultiModelStress(t *testing.T) {
 			defer wg.Done()
 			for j := 0; j < 3; j++ {
 				slog.Info("Starting", "req", i, "iter", j, "model", req[i].Model)
-				DoGenerate(ctx, t, client, req[i], resp[i], 90*time.Second, 5*time.Second)
+				DoGenerate(ctx, t, client, req[i], resp[i], 120*time.Second, 5*time.Second)
 			}
 		}(i)
 	}

integration/max_queue_test.go

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

integration/utils_test.go

@@ -85,7 +85,7 @@ func GetTestEndpoint() (*api.Client, string) {
 var serverMutex sync.Mutex
 var serverReady bool
 
-func startServer(ctx context.Context, ollamaHost string) error {
+func startServer(t *testing.T, ctx context.Context, ollamaHost string) error {
 	// Make sure the server has been built
 	CLIName, err := filepath.Abs("../ollama")
 	if err != nil {
@@ -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)
@@ -200,7 +200,7 @@ func InitServerConnection(ctx context.Context, t *testing.T) (*api.Client, strin
 		}
 		lifecycle.ServerLogFile = fp.Name()
 		fp.Close()
-		require.NoError(t, startServer(ctx, testEndpoint))
+		require.NoError(t, startServer(t, ctx, testEndpoint))
 	}
 
 	return client, testEndpoint, func() {

llm/ext_server/server.cpp

@@ -66,7 +66,7 @@ struct server_params {
 };
 
 bool server_verbose = false;
-bool server_log_json = true;
+bool server_log_json = false;
 
 enum stop_type {
     STOP_FULL,
@@ -266,7 +266,7 @@ struct server_slot {
         sprintf(buffer, "prompt eval time     = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)",
                 t_prompt_processing, n_prompt_tokens_processed,
                 t_token, n_tokens_second);
-        LOG_INFO(buffer, {
+        LOG_DEBUG(buffer, {
             {"slot_id",                   id},
             {"task_id",                   task_id},
             {"t_prompt_processing",       t_prompt_processing},
@@ -280,7 +280,7 @@ struct server_slot {
         sprintf(buffer, "generation eval time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)",
                 t_token_generation, n_decoded,
                 t_token, n_tokens_second);
-        LOG_INFO(buffer, {
+        LOG_DEBUG(buffer, {
             {"slot_id",            id},
             {"task_id",            task_id},
             {"t_token_generation", t_token_generation},
@@ -290,7 +290,7 @@ struct server_slot {
         });
 
         sprintf(buffer, "          total time = %10.2f ms", t_prompt_processing + t_token_generation);
-        LOG_INFO(buffer, {
+        LOG_DEBUG(buffer, {
             {"slot_id",             id},
             {"task_id",             task_id},
             {"t_prompt_processing", t_prompt_processing},
@@ -334,6 +334,7 @@ struct server_metrics {
 struct llama_server_context
 {
     llama_model *model = nullptr;
+    float modelProgress = 0.0;
     llama_context *ctx = nullptr;
 
     clip_ctx *clp_ctx = nullptr;
@@ -371,7 +372,7 @@ struct llama_server_context
     {
         if (clp_ctx)
         {
-            LOG_INFO("freeing clip model", {});
+            LOG_DEBUG("freeing clip model", {});
             clip_free(clp_ctx);
             clp_ctx = nullptr;
         }
@@ -392,7 +393,7 @@ struct llama_server_context
         params = params_;
         if (!params.mmproj.empty()) {
             multimodal = true;
-            LOG_INFO("Multi Modal Mode Enabled", {});
+            LOG_DEBUG("Multi Modal Mode Enabled", {});
             clp_ctx = clip_model_load(params.mmproj.c_str(), /*verbosity=*/ 1);
             if(clp_ctx == nullptr) {
                 LOG_ERROR("unable to load clip model", {{"model", params.mmproj}});
@@ -445,7 +446,7 @@ struct llama_server_context
 
         const int32_t n_ctx_slot = n_ctx / params.n_parallel;
 
-        LOG_INFO("initializing slots", {{"n_slots", params.n_parallel}});
+        LOG_DEBUG("initializing slots", {{"n_slots", params.n_parallel}});
         for (int i = 0; i < params.n_parallel; i++)
         {
             server_slot slot;
@@ -454,7 +455,7 @@ struct llama_server_context
             slot.n_ctx = n_ctx_slot;
             slot.n_predict = params.n_predict;
 
-            LOG_INFO("new slot", {
+            LOG_DEBUG("new slot", {
                 {"slot_id",    slot.id},
                 {"n_ctx_slot", slot.n_ctx}
             });
@@ -468,7 +469,7 @@ struct llama_server_context
                 //GGML_ASSERT(n_ctx_train % ga_w == 0     && "n_ctx_train must be a multiple of ga_w");    // NOLINT
                 //GGML_ASSERT(n_ctx >= n_ctx_train * ga_n && "n_ctx must be at least n_ctx_train * ga_n"); // NOLINT
 
-                LOG_INFO("slot self-extend", {
+                LOG_DEBUG("slot self-extend", {
                     {"slot_id",   slot.id},
                     {"ga_n",      ga_n},
                     {"ga_w",      ga_w}
@@ -737,7 +738,7 @@ struct llama_server_context
                     sampler_names.emplace_back(sampler_name);
                 }
             }
-            slot->sparams.samplers_sequence = sampler_types_from_names(sampler_names, false);
+            slot->sparams.samplers_sequence = llama_sampling_types_from_names(sampler_names, false);
         }
         else
         {
@@ -827,7 +828,7 @@ struct llama_server_context
 
         all_slots_are_idle = false;
 
-        LOG_INFO("slot is processing task", {
+        LOG_DEBUG("slot is processing task", {
             {"slot_id", slot->id},
             {"task_id", slot->task_id},
         });
@@ -1095,7 +1096,7 @@ struct llama_server_context
         std::vector<std::string> samplers_sequence;
         for (const auto &sampler_type : slot.sparams.samplers_sequence)
         {
-            samplers_sequence.emplace_back(sampler_type_to_name_string(sampler_type));
+            samplers_sequence.emplace_back(llama_sampling_type_to_str(sampler_type));
         }
 
         return json {
@@ -1504,7 +1505,7 @@ struct llama_server_context
                     }
                     slots_data.push_back(slot_data);
                 }
-                LOG_INFO("slot data", {
+                LOG_DEBUG("slot data", {
                     {"task_id",            task.id},
                     {"n_idle_slots",       n_idle_slots},
                     {"n_processing_slots", n_processing_slots}
@@ -1566,7 +1567,7 @@ struct llama_server_context
     bool update_slots() {
         if (system_need_update)
         {
-            LOG_INFO("updating system prompt", {});
+            LOG_DEBUG("updating system prompt", {});
             system_prompt_update();
         }
 
@@ -1576,7 +1577,7 @@ struct llama_server_context
         {
             if (system_prompt.empty() && clean_kv_cache)
             {
-                LOG_INFO("all slots are idle and system prompt is empty, clear the KV cache", {});
+                LOG_DEBUG("all slots are idle and system prompt is empty, clear the KV cache", {});
                 kv_cache_clear();
             }
             return true;
@@ -1599,7 +1600,7 @@ struct llama_server_context
                     const int n_left    = (int) system_tokens.size() + slot.n_past - n_keep;
                     const int n_discard = n_left / 2;
 
-                    LOG_INFO("slot context shift", {
+                    LOG_DEBUG("slot context shift", {
                         {"slot_id",         slot.id},
                         {"task_id",         slot.task_id},
                         {"n_keep",          n_keep},
@@ -1638,7 +1639,7 @@ struct llama_server_context
                 slot.command = NONE;
                 slot.t_last_used = ggml_time_us();
 
-                LOG_INFO("slot released", {
+                LOG_DEBUG("slot released", {
                     {"slot_id",         slot.id},
                     {"task_id",         slot.task_id},
                     {"n_ctx",           n_ctx},
@@ -1807,7 +1808,7 @@ struct llama_server_context
                             slot.ga_i = ga_i;
                         }
 
-                        LOG_INFO("slot progression", {
+                        LOG_DEBUG("slot progression", {
                             { "slot_id",    slot.id },
                             { "task_id",    slot.task_id },
                             { "n_past",     slot.n_past },
@@ -1822,7 +1823,7 @@ struct llama_server_context
                     if (slot.n_past == slot.n_prompt_tokens && slot.n_past > 0)
                     {
                         // we have to evaluate at least 1 token to generate logits.
-                        LOG_INFO("we have to evaluate at least 1 token to generate logits", {
+                        LOG_DEBUG("we have to evaluate at least 1 token to generate logits", {
                             { "slot_id", slot.id },
                             { "task_id", slot.task_id }
                         });
@@ -1834,7 +1835,7 @@ struct llama_server_context
                     }
 
                     int p0 = (int) system_tokens.size() + slot.n_past;
-                    LOG_INFO("kv cache rm [p0, end)", {
+                    LOG_DEBUG("kv cache rm [p0, end)", {
                         { "slot_id", slot.id },
                         { "task_id", slot.task_id },
                         { "p0",      p0 }
@@ -2104,6 +2105,7 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     printf("  --embedding               enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled");
     printf("  -np N, --parallel N       number of slots for process requests (default: %d)\n", params.n_parallel);
     printf("  -cb, --cont-batching      enable continuous batching (a.k.a dynamic batching) (default: disabled)\n");
+    printf("  -fa, --flash-attn         enable Flash Attention (default: %s)\n", params.flash_attn ? "enabled" : "disabled");
     printf("  -spf FNAME, --system-prompt-file FNAME\n");
     printf("                            set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications.\n");
     printf("  -ctk TYPE, --cache-type-k TYPE\n");
@@ -2491,11 +2493,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
         }
         else if (arg == "-v" || arg == "--verbose")
         {
-#if SERVER_VERBOSE != 1
-            LOG_WARNING("server.cpp is not built with verbose logging.", {});
-#else
             server_verbose = true;
-#endif
         }
         else if (arg == "--mlock")
         {
@@ -2505,7 +2503,8 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
         {
             params.use_mmap = false;
         }
-        else if (arg == "--numa") {
+        else if (arg == "--numa")
+        {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
@@ -2525,6 +2524,10 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
         {
             params.cont_batching = true;
         }
+        else if (arg == "-fa" || arg == "--flash-attn")
+        {
+            params.flash_attn = true;
+        }
         else if (arg == "-np" || arg == "--parallel")
         {
             if (++i >= argc)
@@ -2533,7 +2536,8 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
                 break;
             }
             params.n_parallel = std::stoi(argv[i]);
-        } else if (arg == "-n" || arg == "--n-predict")
+        }
+        else if (arg == "-n" || arg == "--n-predict")
         {
             if (++i >= argc)
             {
@@ -2541,7 +2545,8 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
                 break;
             }
             params.n_predict = std::stoi(argv[i]);
-        } else if (arg == "-spf" || arg == "--system-prompt-file")
+        }
+        else if (arg == "-spf" || arg == "--system-prompt-file")
         {
             if (++i >= argc)
             {
@@ -2601,7 +2606,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
         else if (arg == "--log-disable")
         {
             log_set_target(stdout);
-            LOG_INFO("logging to file is disabled.", {});
+            LOG_DEBUG("logging to file is disabled.", {});
         }
         else if (arg == "--slots-endpoint-disable")
         {
@@ -2727,12 +2732,12 @@ static json format_detokenized_response(std::string content)
 static void log_server_request(const httplib::Request &req, const httplib::Response &res)
 {
     // skip GH copilot requests when using default port
-    if (req.path == "/v1/health" || req.path == "/v1/completions")
+    if (req.path == "/health" || req.path == "/v1/health" || req.path == "/v1/completions")
     {
         return;
     }
 
-    LOG_INFO("request", {
+    LOG_DEBUG("request", {
         {"remote_addr", req.remote_addr},
         {"remote_port", req.remote_port},
         {"status",      res.status},
@@ -2775,6 +2780,12 @@ inline void signal_handler(int signal) {
     shutdown_handler(signal);
 }
 
+static bool update_load_progress(float progress, void *data)
+{
+    ((llama_server_context*)data)->modelProgress = progress;
+    return true;
+}
+
 #if defined(_WIN32)
 char* wchar_to_char(const wchar_t* wstr) {
     if (wstr == nullptr) return nullptr;
@@ -2880,7 +2891,9 @@ int main(int argc, char **argv) {
                 break;
             }
             case SERVER_STATE_LOADING_MODEL:
-                res.set_content(R"({"status": "loading model"})", "application/json");
+                char buf[128];
+                snprintf(&buf[0], 128, R"({"status": "loading model", "progress": %0.2f})", llama.modelProgress);
+                res.set_content(buf, "application/json");
                 res.status = 503; // HTTP Service Unavailable
                 break;
             case SERVER_STATE_ERROR:
@@ -3054,7 +3067,30 @@ int main(int argc, char **argv) {
         log_data["api_key"] = "api_key: " + std::to_string(sparams.api_keys.size()) + " keys loaded";
     }
 
+    if (sparams.n_threads_http < 1) {
+        // +2 threads for monitoring endpoints
+        sparams.n_threads_http = std::max(params.n_parallel + 2, (int32_t) std::thread::hardware_concurrency() - 1);
+    }
+    log_data["n_threads_http"] =  std::to_string(sparams.n_threads_http);
+    svr.new_task_queue = [&sparams] { return new httplib::ThreadPool(sparams.n_threads_http); };
+
+    LOG_INFO("HTTP server listening", log_data);
+    // run the HTTP server in a thread - see comment below
+    std::thread t([&]()
+            {
+                if (!svr.listen_after_bind())
+                {
+                    state.store(SERVER_STATE_ERROR);
+                    return 1;
+                }
+
+                return 0;
+            });
+
     // load the model
+    params.progress_callback = update_load_progress;
+    params.progress_callback_user_data = (void*)&llama;
+
     if (!llama.load_model(params))
     {
         state.store(SERVER_STATE_ERROR);
@@ -3258,26 +3294,6 @@ int main(int argc, char **argv) {
     }*/
     //);
 
-    if (sparams.n_threads_http < 1) {
-        // +2 threads for monitoring endpoints
-        sparams.n_threads_http = std::max(params.n_parallel + 2, (int32_t) std::thread::hardware_concurrency() - 1);
-    }
-    log_data["n_threads_http"] =  std::to_string(sparams.n_threads_http);
-    svr.new_task_queue = [&sparams] { return new httplib::ThreadPool(sparams.n_threads_http); };
-
-    LOG_INFO("HTTP server listening", log_data);
-    // run the HTTP server in a thread - see comment below
-    std::thread t([&]()
-            {
-                if (!svr.listen_after_bind())
-                {
-                    state.store(SERVER_STATE_ERROR);
-                    return 1;
-                }
-
-                return 0;
-            });
-
     llama.queue_tasks.on_new_task(std::bind(
         &llama_server_context::process_single_task, &llama, std::placeholders::_1));
     llama.queue_tasks.on_finish_multitask(std::bind(

llm/ext_server/utils.hpp

@@ -55,9 +55,10 @@ extern bool server_log_json;
     } while (0)
 #endif
 
-#define LOG_ERROR(  MSG, ...) server_log("ERR",  __func__, __LINE__, MSG, __VA_ARGS__)
+#define LOG_ERROR(  MSG, ...) server_log("ERROR",  __func__, __LINE__, MSG, __VA_ARGS__)
 #define LOG_WARNING(MSG, ...) server_log("WARN", __func__, __LINE__, MSG, __VA_ARGS__)
 #define LOG_INFO(   MSG, ...) server_log("INFO", __func__, __LINE__, MSG, __VA_ARGS__)
+#define LOG_DEBUG(  MSG, ...) server_log("DEBUG", __func__, __LINE__, MSG, __VA_ARGS__)
 
 enum server_state {
     SERVER_STATE_LOADING_MODEL,  // Server is starting up, model not fully loaded yet
@@ -123,6 +124,10 @@ static inline void server_log(const char *level, const char *function, int line,
         {"timestamp", time(nullptr)},
     };
 
+    if (strncmp("DEBUG", level, strlen(level)) == 0 && !server_verbose) {
+        return;
+    }
+
     if (server_log_json) {
         log.merge_patch(
                 {
@@ -137,14 +142,12 @@ static inline void server_log(const char *level, const char *function, int line,
 
         std::cout << log.dump(-1, ' ', false, json::error_handler_t::replace) << "\n" << std::flush;
     } else {
-        char buf[1024];
-        snprintf(buf, 1024, "%4s [%24s] %s", level, function, message);
-
         if (!extra.empty()) {
             log.merge_patch(extra);
         }
+
         std::stringstream ss;
-        ss << buf << " |";
+        ss << level << " [" << function << "] " << message << " |";
         for (const auto& el : log.items())
         {
             const std::string value = el.value().dump(-1, ' ', false, json::error_handler_t::replace);

llm/filetype.go

@@ -0,0 +1,180 @@
+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
+	fileTypeIQ3_XS
+	fileTypeIQ3_XXS
+	fileTypeIQ1_S
+	fileTypeIQ4_NL
+	fileTypeIQ3_S
+	fileTypeIQ2_S
+	fileTypeIQ4_XS
+	fileTypeIQ2_M
+	fileTypeIQ1_M
+	fileTypeBF16
+
+	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 "IQ3_XS":
+		return fileTypeIQ3_XS, nil
+	case "IQ3_XXS":
+		return fileTypeIQ3_XXS, nil
+	case "IQ1_S":
+		return fileTypeIQ1_S, nil
+	case "IQ4_NL":
+		return fileTypeIQ4_NL, nil
+	case "IQ3_S":
+		return fileTypeIQ3_S, nil
+	case "IQ2_S":
+		return fileTypeIQ2_S, nil
+	case "IQ4_XS":
+		return fileTypeIQ4_XS, nil
+	case "IQ2_M":
+		return fileTypeIQ2_M, nil
+	case "IQ1_M":
+		return fileTypeIQ1_M, nil
+	case "BF16":
+		return fileTypeBF16, nil
+	default:
+		return fileTypeUnknown, fmt.Errorf("unknown fileType: %s", s)
+	}
+}
+
+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 fileTypeIQ3_XS:
+		return "IQ3_XS"
+	case fileTypeIQ3_XXS:
+		return "IQ3_XXS"
+	case fileTypeIQ1_S:
+		return "IQ1_S"
+	case fileTypeIQ4_NL:
+		return "IQ4_NL"
+	case fileTypeIQ3_S:
+		return "IQ3_S"
+	case fileTypeIQ2_S:
+		return "IQ2_S"
+	case fileTypeIQ4_XS:
+		return "IQ4_XS"
+	case fileTypeIQ2_M:
+		return "IQ2_M"
+	case fileTypeIQ1_M:
+		return "IQ1_M"
+	case fileTypeBF16:
+		return "BF16"
+	default:
+		return "unknown"
+	}
+}
+
+func (t fileType) Value() uint32 {
+	return uint32(t)
+}

llm/generate/gen_linux.sh

@@ -156,7 +156,7 @@ if [ -z "${CUDART_LIB_DIR}" ]; then
     CUDART_LIB_DIR="${CUDA_LIB_DIR}"
 fi
 
-if [ -d "${CUDA_LIB_DIR}" ]; then
+if [ -z "${OLLAMA_SKIP_CUDA_GENERATE}" -a -d "${CUDA_LIB_DIR}" ]; then
     echo "CUDA libraries detected - building dynamic CUDA library"
     init_vars
     CUDA_MAJOR=$(ls "${CUDA_LIB_DIR}"/libcudart.so.* | head -1 | cut -f3 -d. || true)
@@ -206,6 +206,36 @@ if [ -d "${CUDA_LIB_DIR}" ]; then
 
 fi
 
+if [ -z "${ONEAPI_ROOT}" ]; then
+    # Try the default location in case it exists
+    ONEAPI_ROOT=/opt/intel/oneapi
+fi
+
+if [ -d "${ONEAPI_ROOT}" ]; then
+    echo "OneAPI libraries detected - building dynamic OneAPI library"
+    init_vars
+    source ${ONEAPI_ROOT}/setvars.sh --force # set up environment variables for oneAPI
+    CC=icx
+    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL=ON -DLLAMA_SYCL_F16=OFF"
+    BUILD_DIR="../build/linux/${ARCH}/oneapi"
+    EXTRA_LIBS="-fsycl -Wl,-rpath,${ONEAPI_ROOT}/compiler/latest/lib,-rpath,${ONEAPI_ROOT}/mkl/latest/lib,-rpath,${ONEAPI_ROOT}/tbb/latest/lib,-rpath,${ONEAPI_ROOT}/compiler/latest/opt/oclfpga/linux64/lib -lOpenCL -lmkl_core -lmkl_sycl_blas -lmkl_intel_ilp64 -lmkl_tbb_thread -ltbb"
+    DEBUG_FLAGS="" # icx compiles with -O0 if we pass -g, so we must remove it
+    build
+
+    # copy oneAPI dependencies
+    for dep in $(ldd "${BUILD_DIR}/bin/ollama_llama_server" | grep "=>" | cut -f2 -d= | cut -f2 -d' ' | grep -e sycl -e mkl -e tbb); do
+        cp "${dep}" "${BUILD_DIR}/bin/"
+    done
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libOpenCL.so" "${BUILD_DIR}/bin/"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libimf.so" "${BUILD_DIR}/bin/"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libintlc.so.5" "${BUILD_DIR}/bin/"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libirng.so" "${BUILD_DIR}/bin/"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libpi_level_zero.so" "${BUILD_DIR}/bin/"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libsvml.so" "${BUILD_DIR}/bin/"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libur_loader.so.0" "${BUILD_DIR}/bin/"
+    compress
+fi
+
 if [ -z "${ROCM_PATH}" ]; then
     # Try the default location in case it exists
     ROCM_PATH=/opt/rocm
@@ -218,7 +248,7 @@ if [ -z "${CLBlast_DIR}" ]; then
     fi
 fi
 
-if [ -d "${ROCM_PATH}" ]; then
+if [ -z "${OLLAMA_SKIP_ROCM_GENERATE}" -a -d "${ROCM_PATH}" ]; then
     echo "ROCm libraries detected - building dynamic ROCm library"
     if [ -f ${ROCM_PATH}/lib/librocblas.so.*.*.????? ]; then
         ROCM_VARIANT=_v$(ls ${ROCM_PATH}/lib/librocblas.so.*.*.????? | cut -f5 -d. || true)

llm/generate/gen_windows.ps1

@@ -289,6 +289,49 @@ function build_cuda() {
     }
 }
 
+function build_oneapi() {
+  if ((-not "${env:OLLAMA_SKIP_CUDA_GENERATE}") -and ("${env:ONEAPI_ROOT}"))  {
+    # Get oneAPI version
+    $script:ONEAPI_VERSION = icpx --version
+    $script:ONEAPI_VERSION = [regex]::Match($script:ONEAPI_VERSION, '(?<=oneAPI DPC\+\+/C\+\+ Compiler )(?<version>\d+\.\d+\.\d+)').Value
+    if ($null -ne $script:ONEAPI_VERSION) {
+      $script:ONEAPI_VARIANT = "_v" + $script:ONEAPI_VERSION
+    }
+    init_vars
+    $script:buildDir = "../build/windows/${script:ARCH}/oneapi$script:ONEAPI_VARIANT"
+    $script:distDir ="$script:DIST_BASE\oneapi$script:ONEAPI_VARIANT"
+    $script:cmakeDefs += @(
+      "-G", "MinGW Makefiles",
+      "-DLLAMA_SYCL=ON",
+      "-DCMAKE_C_COMPILER=icx",
+      "-DCMAKE_CXX_COMPILER=icx",
+      "-DCMAKE_BUILD_TYPE=Release"
+    )
+
+    Write-Host "Building oneAPI"
+    build
+    # Ninja doesn't prefix with config name
+    if ($null -ne $script:DUMPBIN) {
+      & "$script:DUMPBIN" /dependents "${script:buildDir}/bin/ollama_llama_server.exe" | Select-String ".dll"
+    }
+    sign
+    install
+
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libirngmd.dll" "${script:distDir}"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libmmd.dll" "${script:distDir}"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_level_zero.dll" "${script:distDir}"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_unified_runtime.dll" "${script:distDir}"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_win_proxy_loader.dll" "${script:distDir}"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\svml_dispmd.dll" "${script:distDir}"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\sycl7.dll" "${script:distDir}"
+    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_core.2.dll" "${script:distDir}"
+    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_sycl_blas.4.dll" "${script:distDir}"
+    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_tbb_thread.2.dll" "${script:distDir}"
+  } else {
+    Write-Host "Skipping oneAPI generation step"
+  }
+}
+
 function build_rocm() {
     if ((-not "${env:OLLAMA_SKIP_ROCM_GENERATE}") -and ("${env:HIP_PATH}")) {
         $script:ROCM_VERSION=(get-item $env:HIP_PATH).Basename
@@ -356,6 +399,7 @@ if ($($args.count) -eq 0) {
         build_cpu_avx
         build_cpu_avx2
         build_cuda
+        build_oneapi
         build_rocm
     }
 

llm/ggla.go

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

llm/ggml.go

@@ -13,82 +13,6 @@ 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
@@ -121,12 +45,12 @@ func (kv KV) ParameterCount() uint64 {
 	return kv.u64("general.parameter_count")
 }
 
-func (kv KV) FileType() string {
+func (kv KV) FileType() fileType {
 	if u64 := kv.u64("general.file_type"); u64 > 0 {
 		return fileType(uint32(u64))
 	}
 
-	return "unknown"
+	return fileTypeUnknown
 }
 
 func (kv KV) BlockCount() uint64 {
@@ -182,7 +106,7 @@ type Layer map[string]*Tensor
 
 func (l Layer) size() (size uint64) {
 	for _, t := range l {
-		size += t.size()
+		size += t.Size()
 	}
 
 	return size
@@ -200,12 +124,12 @@ type Tensor struct {
 }
 
 func (t Tensor) blockSize() uint64 {
-	switch {
-	case t.Kind < 2:
+	switch t.Kind {
+	case 0, 1, 24, 25, 26, 27, 28, 30: // F32, F16, I8, I16, I32, I64, F64, BF16
 		return 1
-	case t.Kind < 10:
+	case 2, 3, 4, 5, 6, 7, 8, 9, 20: // Q4_0, Q4_1, Q5_0, Q5_1, Q8_0, Q8_1, IQ4_NL
 		return 32
-	default:
+	default: // All others
 		return 256
 	}
 }
@@ -247,7 +171,29 @@ func (t Tensor) typeSize() uint64 {
 	case 17: // IQ2_XS
 		return 2 + 2*blockSize/8 + blockSize/32
 	case 18: // IQ3_XXS
-		return 2 + 3*blockSize/8
+		return 2 + blockSize/4 + blockSize/8
+	case 19: // IQ1_S
+		return 2 + blockSize/8 + blockSize/16
+	case 20: // IQ4_NL
+		return 2 + blockSize/2
+	case 21: // IQ3_S
+		return 2 + blockSize/4 + blockSize/8 + blockSize/32 + 4
+	case 22: // IQ2_S
+		return 2 + blockSize/4 + blockSize/16
+	case 23: // IQ4_XS
+		return 2 + 2 + blockSize/2 + blockSize/64
+	case 24: // I8
+		return 1
+	case 25: // I16
+		return 2
+	case 26: // I32
+		return 4
+	case 27: // I64
+		return 8
+	case 28: // F64
+		return 8
+	case 29: // IQ1_M
+		return blockSize/8 + blockSize/16 + blockSize/32
 	default:
 		return 0
 	}
@@ -261,7 +207,7 @@ func (t Tensor) parameters() uint64 {
 	return count
 }
 
-func (t Tensor) size() uint64 {
+func (t Tensor) Size() uint64 {
 	return t.parameters() * t.typeSize() / t.blockSize()
 }
 
@@ -286,6 +232,23 @@ 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 {
@@ -347,7 +310,7 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 			// mixtral 8x22b
 			ff := uint64(llm.KV()["llama.feed_forward_length"].(uint32))
 			partialOffload = max(
-				3*ffnGateExpsWeight.size()+4*batch*(2*ff+headsKV+embedding+context+embedding/heads*headsKV),
+				3*ffnGateExpsWeight.Size()+4*batch*(2*ff+headsKV+embedding+context+embedding/heads*headsKV),
 				4*(context*batch*heads+context*embedding/heads*headsKV+batch*1024+embedding/heads*headsKV*batch),
 			)
 		} else if ffnGateWeight, ok := layers["blk.0"]["ffn_gate.0.weight"]; ok {
@@ -388,7 +351,10 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 			4*batch*(1+4*embedding+context+context*heads),
 		)
 
-		partialOffload = 4*batch*(2*embedding+vocab) + embedding*vocab*105/128
+		partialOffload = max(
+			4*batch*(2*embedding+vocab)+embedding*vocab*105/128,
+			4*batch*(2+3*embedding+context+context*heads),
+		)
 	case "stablelm":
 		fullOffload = 4 * batch * (context*(1+heads) + 3*embedding + 2)
 		partialOffload = max(

llm/gguf.go

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

llm/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/memory.go

@@ -7,25 +7,16 @@ import (
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/gpu"
-	"github.com/ollama/ollama/server/envconfig"
+	"github.com/ollama/ollama/envconfig"
 )
 
 // This algorithm looks for a complete fit to determine if we need to unload other models
 func PredictServerFit(allGpus gpu.GpuInfoList, ggml *GGML, adapters, projectors []string, opts api.Options) (bool, uint64) {
-	var estimatedVRAM uint64
-	if opts.NumCtx > int(ggml.KV().ContextLength()) {
-		slog.Warn("requested context length is greater than model max context length", "requested", opts.NumCtx, "model", ggml.KV().ContextLength())
-		opts.NumCtx = int(ggml.KV().ContextLength())
-	}
-
-	if opts.NumCtx < 4 {
-		opts.NumCtx = 4
-	}
-
 	// Split up the GPUs by type and try them
+	var estimatedVRAM uint64
 	for _, gpus := range allGpus.ByLibrary() {
 		var layerCount int
-		layerCount, estimatedVRAM = EstimateGPULayers(gpus, ggml, projectors, opts)
+		layerCount, estimatedVRAM, _ = EstimateGPULayers(gpus, ggml, projectors, opts)
 		if opts.NumGPU < 0 {
 			if layerCount > 0 && layerCount >= int(ggml.KV().BlockCount()+1) {
 				return true, estimatedVRAM
@@ -39,12 +30,9 @@ func PredictServerFit(allGpus gpu.GpuInfoList, ggml *GGML, adapters, projectors
 	return false, estimatedVRAM
 }
 
-// Given a model and one or more GPU targets, predict how many layers and bytes we can load
+// Given a model and one or more GPU targets, predict how many layers and bytes we can load, and the total size
 // The GPUs provided must all be the same Library
-func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts api.Options) (int, uint64) {
-	if gpus[0].Library == "cpu" {
-		return 0, 0
-	}
+func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts api.Options) (int, uint64, uint64) {
 	var memoryAvailable uint64
 	for _, info := range gpus {
 		memoryAvailable += info.FreeMemory
@@ -65,6 +53,12 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 		opts.NumCtx = max(opts.NumCtx, 2048)
 	}
 
+	layers := ggml.Tensors().Layers()
+	// add one layer worth of memory as a buffer
+	if blk0, ok := layers["blk.0"]; ok {
+		memoryMinimum += blk0.size()
+	}
+
 	// fp16 k,v = (1 (k) + 1 (v)) * sizeof(float16) * n_ctx * n_layer * n_embd / n_head * n_head_kv
 	var kv uint64 = 2 * 2 * uint64(opts.NumCtx) * ggml.KV().BlockCount() * ggml.KV().EmbeddingLength() / ggml.KV().HeadCount() * ggml.KV().HeadCountKV()
 
@@ -85,18 +79,11 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 		graphPartialOffload = graphFullOffload
 	}
 
-	layers := ggml.Tensors().Layers()
-
 	// memoryRequiredTotal represents the memory required for full GPU offloading (all layers)
-	memoryRequiredTotal := memoryMinimum + graphFullOffload + layers["blk.0"].size()
+	memoryRequiredTotal := memoryMinimum + graphFullOffload
 
 	// memoryRequiredPartial represents the memory required for partial GPU offloading (n > 0, n < layers)
-	memoryRequiredPartial := memoryMinimum + graphPartialOffload + layers["blk.0"].size()
-
-	if memoryRequiredPartial > memoryAvailable {
-		slog.Debug("insufficient VRAM to load any model layers")
-		return 0, 0
-	}
+	memoryRequiredPartial := memoryMinimum + graphPartialOffload
 
 	var memoryLayerOutput uint64
 	if layer, ok := layers["output_norm"]; ok {
@@ -117,15 +104,17 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 
 	var layerCount int
 	for i := 0; i < int(ggml.KV().BlockCount()); i++ {
-		memoryLayer := layers[fmt.Sprintf("blk.%d", i)].size()
+		if blk, ok := layers[fmt.Sprintf("blk.%d", i)]; ok {
+			memoryLayer := blk.size()
 
-		// KV is proportional to the number of layers
-		memoryLayer += kv / ggml.KV().BlockCount()
+			// KV is proportional to the number of layers
+			memoryLayer += kv / ggml.KV().BlockCount()
 
-		memoryRequiredTotal += memoryLayer
-		if memoryAvailable > memoryRequiredPartial+memoryLayer {
-			memoryRequiredPartial += memoryLayer
-			layerCount++
+			memoryRequiredTotal += memoryLayer
+			if (opts.NumGPU >= 0 && layerCount+1 <= opts.NumGPU) || (opts.NumGPU < 0 && memoryAvailable > memoryRequiredPartial+memoryLayer) {
+				memoryRequiredPartial += memoryLayer
+				layerCount++
+			}
 		}
 	}
 
@@ -134,7 +123,7 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 		memoryRequiredTotal += memoryLayerOutput
 	}
 
-	if memoryAvailable > memoryRequiredTotal {
+	if (opts.NumGPU >= 0 && layerCount+1 <= opts.NumGPU) || (opts.NumGPU < 0 && memoryAvailable > memoryRequiredTotal) {
 		layerCount = int(ggml.KV().BlockCount()) + 1
 		memoryRequiredPartial = memoryRequiredTotal
 	}
@@ -145,10 +134,10 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 		"offload to gpu",
 		slog.Group(
 			"layers",
-			// actual number of layers offloaded
-			"real", opts.NumGPU,
+			// requested number of layers to offload
+			"requested", opts.NumGPU,
 			// estimated number of layers that can be offloaded
-			"estimate", layerCount,
+			"real", layerCount,
 		),
 		slog.Group(
 			"memory",
@@ -181,5 +170,13 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 			),
 		),
 	)
-	return layerCount, uint64(memoryRequiredPartial)
+	if gpus[0].Library == "cpu" {
+		return 0, 0, memoryRequiredTotal
+	}
+	if memoryRequiredPartial > memoryAvailable {
+		slog.Debug("insufficient VRAM to load any model layers")
+		return 0, 0, memoryRequiredTotal
+	}
+
+	return layerCount, memoryRequiredPartial, memoryRequiredTotal
 }

llm/patches/01-load-progress.diff

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

llm/patches/03-load_exception.diff

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

llm/patches/05-clip-fix.diff

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

llm/patches/05-default-pretokenizer.diff

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

llm/server.go

@@ -24,9 +24,9 @@ import (
 	"golang.org/x/sync/semaphore"
 
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/gpu"
-	"github.com/ollama/ollama/server/envconfig"
 )
 
 type LlamaServer interface {
@@ -38,6 +38,7 @@ type LlamaServer interface {
 	Detokenize(ctx context.Context, tokens []int) (string, error)
 	Close() error
 	EstimatedVRAM() uint64
+	EstimatedTotal() uint64
 }
 
 // llmServer is an instance of the llama.cpp server
@@ -49,7 +50,12 @@ type llmServer struct {
 	options api.Options
 
 	// TODO - this should be broken down by GPU
-	estimatedVRAM uint64 // Estimated usage of VRAM by the loaded model
+	estimatedVRAM  uint64 // Estimated usage of VRAM by the loaded model
+	estimatedTotal uint64 // Total size of model
+	totalLayers    uint64
+	gpuCount       int
+	loadDuration   time.Duration // Record how long it took the model to load
+	loadProgress   float32
 
 	sem *semaphore.Weighted
 }
@@ -73,22 +79,18 @@ func LoadModel(model string) (*GGML, error) {
 // The gpu list must be a single family.
 func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, projectors []string, opts api.Options) (LlamaServer, error) {
 	var err error
-	if opts.NumCtx > int(ggml.KV().ContextLength()) {
-		slog.Warn("requested context length is greater than the model's training context window size", "requested", opts.NumCtx, "training size", ggml.KV().ContextLength())
-	}
-
-	if opts.NumCtx < 4 {
-		opts.NumCtx = 4
-	}
-
-	cpuRunner := ""
+	var cpuRunner string
 	var estimatedVRAM uint64
+	var estimatedTotal uint64
 	var systemMemory uint64
+	gpuCount := len(gpus)
 	if (len(gpus) == 1 && gpus[0].Library == "cpu") || opts.NumGPU == 0 {
 
 		// TODO evaluate system memory to see if we should block the load, or force an unload of another CPU runner
 
 		cpuRunner = serverForCpu()
+		gpuCount = 0
+		_, _, estimatedTotal = EstimateGPULayers(gpus, ggml, projectors, opts)
 	} else {
 		if gpus[0].Library == "metal" {
 			memInfo, err := gpu.GetCPUMem()
@@ -100,12 +102,16 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 			}
 		}
 		var layers int
-		layers, estimatedVRAM = EstimateGPULayers(gpus, ggml, projectors, opts)
+		layers, estimatedVRAM, estimatedTotal = EstimateGPULayers(gpus, ggml, projectors, opts)
 
 		if gpus[0].Library == "metal" && estimatedVRAM > systemMemory {
 			// disable partial offloading when model is greater than total system memory as this
 			// can lead to locking up the system
 			opts.NumGPU = 0
+		} else if gpus[0].Library != "metal" && layers == 0 {
+			// Don't bother loading into the GPU if no layers can fit
+			cpuRunner = serverForCpu()
+			gpuCount = 0
 		} else if opts.NumGPU < 0 && layers > 0 && gpus[0].Library != "cpu" {
 			opts.NumGPU = layers
 		}
@@ -133,6 +139,10 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		} else {
 			slog.Info("user override", "OLLAMA_LLM_LIBRARY", demandLib, "path", serverPath)
 			servers = []string{demandLib}
+			if strings.HasPrefix(demandLib, "cpu") {
+				// Omit the GPU flag to silence the warning
+				opts.NumGPU = -1
+			}
 		}
 	}
 
@@ -146,11 +156,8 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		"--batch-size", fmt.Sprintf("%d", opts.NumBatch),
 		"--embedding",
 	}
-	if envconfig.Debug {
-		params = append(params, "--log-format", "json")
-	} else {
-		params = append(params, "--log-disable")
-	}
+
+	params = append(params, "--log-disable")
 
 	if opts.NumGPU >= 0 {
 		params = append(params, "--n-gpu-layers", fmt.Sprintf("%d", opts.NumGPU))
@@ -194,6 +201,23 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		params = append(params, "--numa")
 	}
 
+	flashAttnEnabled := envconfig.FlashAttention
+
+	// partial offloading does not support flash attention
+	if uint64(opts.NumGPU) < ggml.KV().BlockCount()+1 {
+		flashAttnEnabled = false
+	}
+
+	// only cuda (compute capability 7+) and metal support flash attention
+	for _, g := range gpus {
+		if g.Library != "metal" && (g.Library != "cuda" || g.DriverMajor < 7) {
+			flashAttnEnabled = false
+		}
+	}
+	if flashAttnEnabled {
+		params = append(params, "--flash-attn")
+	}
+
 	numParallel := envconfig.NumParallel
 
 	// TODO (jmorganca): multimodal models don't support parallel yet
@@ -210,11 +234,16 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		if dir == "" {
 			// Shouldn't happen
 			finalErr = fmt.Errorf("[%d] server %s not listed in available servers %v", i, servers[i], availableServers)
-			slog.Error("sever list inconsistent", "error", finalErr)
+			slog.Error("server list inconsistent", "error", finalErr)
 			continue
 		}
 
-		// Find an availableServers  port, retry on each iterration in case the failure was a port conflict race
+		if strings.HasPrefix(servers[i], "cpu") {
+			// TODO if we tried a gpu runner first, and it failed, record the error and bubble that back up
+			gpuCount = 0
+		}
+
+		// Find an availableServers  port, retry on each iteration in case the failure was a port conflict race
 		port := 0
 		if a, err := net.ResolveTCPAddr("tcp", "localhost:0"); err == nil {
 			var l *net.TCPListener
@@ -267,12 +296,16 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		}
 
 		s := &llmServer{
-			port:          port,
-			cmd:           exec.Command(server, finalParams...),
-			status:        NewStatusWriter(os.Stderr),
-			options:       opts,
-			estimatedVRAM: estimatedVRAM,
-			sem:           semaphore.NewWeighted(int64(numParallel)),
+			port:           port,
+			cmd:            exec.Command(server, finalParams...),
+			status:         NewStatusWriter(os.Stderr),
+			options:        opts,
+			estimatedVRAM:  estimatedVRAM,
+			estimatedTotal: estimatedTotal,
+			sem:            semaphore.NewWeighted(int64(numParallel)),
+			totalLayers:    ggml.KV().BlockCount() + 1,
+			gpuCount:       gpuCount,
+			done:           make(chan error, 1),
 		}
 
 		s.cmd.Env = os.Environ()
@@ -303,10 +336,29 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		}
 
 		slog.Info("starting llama server", "cmd", s.cmd.String())
-		// Log at debug as the environment is inherited and might contain sensitive information
-		slog.Debug("subprocess", "environment", s.cmd.Env)
+		if envconfig.Debug {
+			filteredEnv := []string{}
+			for _, ev := range s.cmd.Env {
+				if strings.HasPrefix(ev, "CUDA_") ||
+					strings.HasPrefix(ev, "ROCM_") ||
+					strings.HasPrefix(ev, "HIP_") ||
+					strings.HasPrefix(ev, "HSA_") ||
+					strings.HasPrefix(ev, "GGML_") ||
+					strings.HasPrefix(ev, "PATH=") ||
+					strings.HasPrefix(ev, "LD_LIBRARY_PATH=") {
+					filteredEnv = append(filteredEnv, ev)
+				}
+			}
+			// Log at debug as the environment is inherited and might contain sensitive information
+			slog.Debug("subprocess", "environment", filteredEnv)
+		}
 
 		if err = s.cmd.Start(); err != nil {
+			// Detect permission denied and augment them essage about noexec
+			if errors.Is(err, os.ErrPermission) {
+				finalErr = fmt.Errorf("unable to start server %w.  %s may have noexec set.  Set OLLAMA_TMPDIR for server to a writable executable directory", err, dir)
+				continue
+			}
 			msg := ""
 			if s.status != nil && s.status.LastErrMsg != "" {
 				msg = s.status.LastErrMsg
@@ -316,6 +368,11 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 			continue
 		}
 
+		// reap subprocess when it exits
+		go func() {
+			s.done <- s.cmd.Wait()
+		}()
+
 		return s, nil
 	}
 
@@ -369,10 +426,11 @@ func (s ServerStatus) ToString() string {
 }
 
 type ServerStatusResp struct {
-	Status          string `json:"status"`
-	SlotsIdle       int    `json:"slots_idle"`
-	SlotsProcessing int    `json:"slots_processing"`
-	Error           string `json:"error"`
+	Status          string  `json:"status"`
+	SlotsIdle       int     `json:"slots_idle"`
+	SlotsProcessing int     `json:"slots_processing"`
+	Error           string  `json:"error"`
+	Progress        float32 `json:"progress"`
 }
 
 func (s *llmServer) getServerStatus(ctx context.Context) (ServerStatus, error) {
@@ -382,6 +440,10 @@ func (s *llmServer) getServerStatus(ctx context.Context) (ServerStatus, error) {
 		if s.status != nil && s.status.LastErrMsg != "" {
 			msg = s.status.LastErrMsg
 		}
+		if s.cmd.ProcessState.ExitCode() == -1 {
+			// Most likely a signal killed it, log some more details to try to help troubleshoot
+			slog.Warn("llama runner process no longer running", "sys", s.cmd.ProcessState.Sys(), "string", s.cmd.ProcessState.String())
+		}
 		return ServerStatusError, fmt.Errorf("llama runner process no longer running: %d %s", s.cmd.ProcessState.ExitCode(), msg)
 	}
 
@@ -416,6 +478,7 @@ func (s *llmServer) getServerStatus(ctx context.Context) (ServerStatus, error) {
 	case "no slot available":
 		return ServerStatusNoSlotsAvailable, nil
 	case "loading model":
+		s.loadProgress = status.Progress
 		return ServerStatusLoadingModel, nil
 	default:
 		return ServerStatusError, fmt.Errorf("server error: %+v", status)
@@ -456,17 +519,18 @@ func (s *llmServer) Ping(ctx context.Context) error {
 
 func (s *llmServer) WaitUntilRunning(ctx context.Context) error {
 	start := time.Now()
-	// TODO we need to wire up a better way to detect hangs during model load and startup of the server
-	expiresAt := time.Now().Add(10 * time.Minute) // be generous with timeout, large models can take a while to load
-	ticker := time.NewTicker(50 * time.Millisecond)
-	defer ticker.Stop()
+	stallDuration := 5 * time.Minute            // If no progress happens
+	finalLoadDuration := 5 * time.Minute        // After we hit 100%, give the runner more time to come online
+	stallTimer := time.Now().Add(stallDuration) // give up if we stall
 
 	slog.Info("waiting for llama runner to start responding")
 	var lastStatus ServerStatus = -1
+	fullyLoaded := false
+
 	for {
 		select {
 		case <-ctx.Done():
-			slog.Info("context expired before server started")
+			slog.Warn("client connection closed before server finished loading, aborting load")
 			return fmt.Errorf("timed out waiting for llama runner to start: %w", ctx.Err())
 		case err := <-s.done:
 			msg := ""
@@ -474,41 +538,49 @@ func (s *llmServer) WaitUntilRunning(ctx context.Context) error {
 				msg = s.status.LastErrMsg
 			}
 			return fmt.Errorf("llama runner process has terminated: %v %s", err, msg)
-		case <-ticker.C:
-			if time.Now().After(expiresAt) {
-				// timeout
-				msg := ""
-				if s.status != nil && s.status.LastErrMsg != "" {
-					msg = s.status.LastErrMsg
-				}
-				return fmt.Errorf("timed out waiting for llama runner to start: %s", msg)
+		default:
+		}
+		if time.Now().After(stallTimer) {
+			// timeout
+			msg := ""
+			if s.status != nil && s.status.LastErrMsg != "" {
+				msg = s.status.LastErrMsg
 			}
-			if s.cmd.ProcessState != nil {
-				msg := ""
-				if s.status != nil && s.status.LastErrMsg != "" {
-					msg = s.status.LastErrMsg
-				}
-				return fmt.Errorf("llama runner process no longer running: %d %s", s.cmd.ProcessState.ExitCode(), msg)
+			return fmt.Errorf("timed out waiting for llama runner to start - progress %0.2f - %s", s.loadProgress, msg)
+		}
+		if s.cmd.ProcessState != nil {
+			msg := ""
+			if s.status != nil && s.status.LastErrMsg != "" {
+				msg = s.status.LastErrMsg
 			}
-
-			c, cancel := context.WithTimeout(ctx, 200*time.Millisecond)
-			defer cancel()
-			status, err := s.getServerStatus(c)
-			if err != nil && lastStatus != status {
-				slog.Debug("server not yet available", "error", err)
-				lastStatus = status
-				continue
-			}
-
-			switch status {
-			case ServerStatusLoadingModel:
-				// TODO - this state never seems to happen with the current server.cpp code (bug?)
-				// it doesn't respond to the health endpoint until after the model is loaded
-				slog.Debug("loading model")
-			case ServerStatusReady:
-				slog.Debug(fmt.Sprintf("llama runner started in %f seconds", time.Since(start).Seconds()))
-				return nil
+			return fmt.Errorf("llama runner process no longer running: %d %s", s.cmd.ProcessState.ExitCode(), msg)
+		}
+		ctx, cancel := context.WithTimeout(ctx, 200*time.Millisecond)
+		defer cancel()
+		priorProgress := s.loadProgress
+		status, _ := s.getServerStatus(ctx)
+		if lastStatus != status && status != ServerStatusReady {
+			// Only log on status changes
+			slog.Info("waiting for server to become available", "status", status.ToString())
+		}
+		switch status {
+		case ServerStatusReady:
+			s.loadDuration = time.Since(start)
+			slog.Info(fmt.Sprintf("llama runner started in %0.2f seconds", s.loadDuration.Seconds()))
+			return nil
+		default:
+			lastStatus = status
+			// Reset the timer as long as we're making forward progress on the load
+			if priorProgress != s.loadProgress {
+				slog.Debug(fmt.Sprintf("model load progress %0.2f", s.loadProgress))
+				stallTimer = time.Now().Add(stallDuration)
+			} else if !fullyLoaded && int(s.loadProgress*100.0) >= 100 {
+				slog.Debug("model load completed, waiting for server to become available", "status", status.ToString())
+				stallTimer = time.Now().Add(finalLoadDuration)
+				fullyLoaded = true
 			}
+			time.Sleep(time.Millisecond * 250)
+			continue
 		}
 	}
 }
@@ -549,10 +621,11 @@ type ImageData struct {
 }
 
 type completion struct {
-	Content string `json:"content"`
-	Model   string `json:"model"`
-	Prompt  string `json:"prompt"`
-	Stop    bool   `json:"stop"`
+	Content      string `json:"content"`
+	Model        string `json:"model"`
+	Prompt       string `json:"prompt"`
+	Stop         bool   `json:"stop"`
+	StoppedLimit bool   `json:"stopped_limit"`
 
 	Timings struct {
 		PredictedN  int     `json:"predicted_n"`
@@ -571,6 +644,7 @@ type CompletionRequest struct {
 
 type CompletionResponse struct {
 	Content            string
+	DoneReason         string
 	Done               bool
 	PromptEvalCount    int
 	PromptEvalDuration time.Duration
@@ -688,7 +762,7 @@ func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn fu
 
 			var c completion
 			if err := json.Unmarshal(evt, &c); err != nil {
-				return fmt.Errorf("error unmarshaling llm prediction response: %v", err)
+				return fmt.Errorf("error unmarshalling llm prediction response: %v", err)
 			}
 
 			switch {
@@ -712,8 +786,14 @@ func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn fu
 			}
 
 			if c.Stop {
+				doneReason := "stop"
+				if c.StoppedLimit {
+					doneReason = "length"
+				}
+
 				fn(CompletionResponse{
 					Done:               true,
+					DoneReason:         doneReason,
 					PromptEvalCount:    c.Timings.PromptN,
 					PromptEvalDuration: parseDurationMs(c.Timings.PromptMS),
 					EvalCount:          c.Timings.PredictedN,
@@ -908,8 +988,11 @@ func (s *llmServer) Close() error {
 		if err := s.cmd.Process.Kill(); err != nil {
 			return err
 		}
-
-		_ = s.cmd.Wait()
+		// if ProcessState is already populated, Wait already completed, no need to wait again
+		if s.cmd.ProcessState == nil {
+			slog.Debug("waiting for llama server to exit")
+			<-s.done
+		}
 
 		slog.Debug("llama server stopped")
 	}
@@ -921,6 +1004,10 @@ func (s *llmServer) EstimatedVRAM() uint64 {
 	return s.estimatedVRAM
 }
 
+func (s *llmServer) EstimatedTotal() uint64 {
+	return s.estimatedTotal
+}
+
 func parseDurationMs(ms float64) time.Duration {
 	dur, err := time.ParseDuration(fmt.Sprintf("%fms", ms))
 	if err != nil {

macapp/src/index.ts

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

openai/openai.go

@@ -109,13 +109,12 @@ func toChatCompletion(id string, r api.ChatResponse) ChatCompletion {
 		Choices: []Choice{{
 			Index:   0,
 			Message: Message{Role: r.Message.Role, Content: r.Message.Content},
-			FinishReason: func(done bool) *string {
-				if done {
-					reason := "stop"
+			FinishReason: func(reason string) *string {
+				if len(reason) > 0 {
 					return &reason
 				}
 				return nil
-			}(r.Done),
+			}(r.DoneReason),
 		}},
 		Usage: Usage{
 			// TODO: ollama returns 0 for prompt eval if the prompt was cached, but openai returns the actual count
@@ -133,19 +132,16 @@ func toChunk(id string, r api.ChatResponse) ChatCompletionChunk {
 		Created:           time.Now().Unix(),
 		Model:             r.Model,
 		SystemFingerprint: "fp_ollama",
-		Choices: []ChunkChoice{
-			{
-				Index: 0,
-				Delta: Message{Role: "assistant", Content: r.Message.Content},
-				FinishReason: func(done bool) *string {
-					if done {
-						reason := "stop"
-						return &reason
-					}
-					return nil
-				}(r.Done),
-			},
-		},
+		Choices: []ChunkChoice{{
+			Index: 0,
+			Delta: Message{Role: "assistant", Content: r.Message.Content},
+			FinishReason: func(reason string) *string {
+				if len(reason) > 0 {
+					return &reason
+				}
+				return nil
+			}(r.DoneReason),
+		}},
 	}
 }
 

parser/parser.go

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

parser/parser_test.go

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

readline/buffer.go

@@ -5,16 +5,20 @@ import (
 	"os"
 
 	"github.com/emirpasic/gods/lists/arraylist"
+	"github.com/mattn/go-runewidth"
 	"golang.org/x/term"
 )
 
 type Buffer struct {
-	Pos       int
-	Buf       *arraylist.List
-	Prompt    *Prompt
-	LineWidth int
-	Width     int
-	Height    int
+	DisplayPos int
+	Pos        int
+	Buf        *arraylist.List
+	//LineHasSpace is an arraylist of bools to keep track of whether a line has a space at the end
+	LineHasSpace *arraylist.List
+	Prompt       *Prompt
+	LineWidth    int
+	Width        int
+	Height       int
 }
 
 func NewBuffer(prompt *Prompt) (*Buffer, error) {
@@ -27,25 +31,57 @@ func NewBuffer(prompt *Prompt) (*Buffer, error) {
 	lwidth := width - len(prompt.prompt())
 
 	b := &Buffer{
-		Pos:       0,
-		Buf:       arraylist.New(),
-		Prompt:    prompt,
-		Width:     width,
-		Height:    height,
-		LineWidth: lwidth,
+		DisplayPos:   0,
+		Pos:          0,
+		Buf:          arraylist.New(),
+		LineHasSpace: arraylist.New(),
+		Prompt:       prompt,
+		Width:        width,
+		Height:       height,
+		LineWidth:    lwidth,
 	}
 
 	return b, nil
 }
 
+func (b *Buffer) GetLineSpacing(line int) bool {
+	hasSpace, _ := b.LineHasSpace.Get(line)
+
+	if hasSpace == nil {
+		return false
+	}
+
+	return hasSpace.(bool)
+
+}
+
 func (b *Buffer) MoveLeft() {
 	if b.Pos > 0 {
-		if b.Pos%b.LineWidth == 0 {
-			fmt.Printf(CursorUp + CursorBOL + cursorRightN(b.Width))
-		} else {
-			fmt.Print(CursorLeft)
+		//asserts that we retrieve a rune
+		if e, ok := b.Buf.Get(b.Pos - 1); ok {
+			if r, ok := e.(rune); ok {
+				rLength := runewidth.RuneWidth(r)
+
+				if b.DisplayPos%b.LineWidth == 0 {
+					fmt.Printf(CursorUp + CursorBOL + cursorRightN(b.Width))
+					if rLength == 2 {
+						fmt.Print(CursorLeft)
+					}
+
+					line := b.DisplayPos/b.LineWidth - 1
+					hasSpace := b.GetLineSpacing(line)
+					if hasSpace {
+						b.DisplayPos -= 1
+						fmt.Print(CursorLeft)
+					}
+				} else {
+					fmt.Print(cursorLeftN(rLength))
+				}
+
+				b.Pos -= 1
+				b.DisplayPos -= rLength
+			}
 		}
-		b.Pos -= 1
 	}
 }
 
@@ -71,18 +107,35 @@ func (b *Buffer) MoveLeftWord() {
 }
 
 func (b *Buffer) MoveRight() {
-	if b.Pos < b.Size() {
-		b.Pos += 1
-		if b.Pos%b.LineWidth == 0 {
-			fmt.Printf(CursorDown + CursorBOL + cursorRightN(len(b.Prompt.prompt())))
-		} else {
-			fmt.Print(CursorRight)
+	if b.Pos < b.Buf.Size() {
+		if e, ok := b.Buf.Get(b.Pos); ok {
+			if r, ok := e.(rune); ok {
+				rLength := runewidth.RuneWidth(r)
+				b.Pos += 1
+				hasSpace := b.GetLineSpacing(b.DisplayPos / b.LineWidth)
+				b.DisplayPos += rLength
+
+				if b.DisplayPos%b.LineWidth == 0 {
+					fmt.Printf(CursorDown + CursorBOL + cursorRightN(len(b.Prompt.prompt())))
+
+				} else if (b.DisplayPos-rLength)%b.LineWidth == b.LineWidth-1 && hasSpace {
+					fmt.Printf(CursorDown + CursorBOL + cursorRightN(len(b.Prompt.prompt())+rLength))
+					b.DisplayPos += 1
+
+				} else if b.LineHasSpace.Size() > 0 && b.DisplayPos%b.LineWidth == b.LineWidth-1 && hasSpace {
+					fmt.Printf(CursorDown + CursorBOL + cursorRightN(len(b.Prompt.prompt())))
+					b.DisplayPos += 1
+
+				} else {
+					fmt.Print(cursorRightN(rLength))
+				}
+			}
 		}
 	}
 }
 
 func (b *Buffer) MoveRightWord() {
-	if b.Pos < b.Size() {
+	if b.Pos < b.Buf.Size() {
 		for {
 			b.MoveRight()
 			v, _ := b.Buf.Get(b.Pos)
@@ -90,7 +143,7 @@ func (b *Buffer) MoveRightWord() {
 				break
 			}
 
-			if b.Pos == b.Size() {
+			if b.Pos == b.Buf.Size() {
 				break
 			}
 		}
@@ -99,7 +152,7 @@ func (b *Buffer) MoveRightWord() {
 
 func (b *Buffer) MoveToStart() {
 	if b.Pos > 0 {
-		currLine := b.Pos / b.LineWidth
+		currLine := b.DisplayPos / b.LineWidth
 		if currLine > 0 {
 			for cnt := 0; cnt < currLine; cnt++ {
 				fmt.Print(CursorUp)
@@ -107,81 +160,195 @@ func (b *Buffer) MoveToStart() {
 		}
 		fmt.Printf(CursorBOL + cursorRightN(len(b.Prompt.prompt())))
 		b.Pos = 0
+		b.DisplayPos = 0
 	}
 }
 
 func (b *Buffer) MoveToEnd() {
-	if b.Pos < b.Size() {
-		currLine := b.Pos / b.LineWidth
-		totalLines := b.Size() / b.LineWidth
+	if b.Pos < b.Buf.Size() {
+		currLine := b.DisplayPos / b.LineWidth
+		totalLines := b.DisplaySize() / b.LineWidth
 		if currLine < totalLines {
 			for cnt := 0; cnt < totalLines-currLine; cnt++ {
 				fmt.Print(CursorDown)
 			}
-			remainder := b.Size() % b.LineWidth
+			remainder := b.DisplaySize() % b.LineWidth
 			fmt.Printf(CursorBOL + cursorRightN(len(b.Prompt.prompt())+remainder))
 		} else {
-			fmt.Print(cursorRightN(b.Size() - b.Pos))
+			fmt.Print(cursorRightN(b.DisplaySize() - b.DisplayPos))
 		}
 
-		b.Pos = b.Size()
+		b.Pos = b.Buf.Size()
+		b.DisplayPos = b.DisplaySize()
 	}
 }
 
-func (b *Buffer) Size() int {
-	return b.Buf.Size()
+func (b *Buffer) DisplaySize() int {
+	sum := 0
+	for i := 0; i < b.Buf.Size(); i++ {
+		if e, ok := b.Buf.Get(i); ok {
+			if r, ok := e.(rune); ok {
+				sum += runewidth.RuneWidth(r)
+			}
+		}
+	}
+
+	return sum
 }
 
 func (b *Buffer) Add(r rune) {
+
 	if b.Pos == b.Buf.Size() {
-		fmt.Printf("%c", r)
-		b.Buf.Add(r)
-		b.Pos += 1
-		if b.Pos > 0 && b.Pos%b.LineWidth == 0 {
+		b.AddChar(r, false)
+	} else {
+		b.AddChar(r, true)
+	}
+}
+
+func (b *Buffer) AddChar(r rune, insert bool) {
+	rLength := runewidth.RuneWidth(r)
+	b.DisplayPos += rLength
+
+	if b.Pos > 0 {
+
+		if b.DisplayPos%b.LineWidth == 0 {
+			fmt.Printf("%c", r)
 			fmt.Printf("\n%s", b.Prompt.AltPrompt)
+
+			if insert {
+				b.LineHasSpace.Set(b.DisplayPos/b.LineWidth-1, false)
+			} else {
+				b.LineHasSpace.Add(false)
+			}
+
+			// this case occurs when a double-width rune crosses the line boundary
+		} else if b.DisplayPos%b.LineWidth < (b.DisplayPos-rLength)%b.LineWidth {
+			if insert {
+				fmt.Print(ClearToEOL)
+			}
+			fmt.Printf("\n%s", b.Prompt.AltPrompt)
+			b.DisplayPos += 1
+			fmt.Printf("%c", r)
+
+			if insert {
+				b.LineHasSpace.Set(b.DisplayPos/b.LineWidth-1, true)
+			} else {
+				b.LineHasSpace.Add(true)
+			}
+
+		} else {
+			fmt.Printf("%c", r)
 		}
 	} else {
 		fmt.Printf("%c", r)
+	}
+
+	if insert {
 		b.Buf.Insert(b.Pos, r)
-		b.Pos += 1
-		if b.Pos > 0 && b.Pos%b.LineWidth == 0 {
-			fmt.Printf("\n%s", b.Prompt.AltPrompt)
-		}
+	} else {
+		b.Buf.Add(r)
+	}
+
+	b.Pos += 1
+
+	if insert {
 		b.drawRemaining()
 	}
 }
 
+func (b *Buffer) countRemainingLineWidth(place int) int {
+	var sum int
+	counter := -1
+	var prevLen int
+
+	for place <= b.LineWidth {
+		counter += 1
+		sum += prevLen
+		if e, ok := b.Buf.Get(b.Pos + counter); ok {
+			if r, ok := e.(rune); ok {
+				place += runewidth.RuneWidth(r)
+				prevLen = len(string(r))
+			}
+		} else {
+			break
+		}
+	}
+
+	return sum
+}
+
 func (b *Buffer) drawRemaining() {
 	var place int
 	remainingText := b.StringN(b.Pos)
 	if b.Pos > 0 {
-		place = b.Pos % b.LineWidth
+		place = b.DisplayPos % b.LineWidth
 	}
 	fmt.Print(CursorHide)
 
 	// render the rest of the current line
-	currLine := remainingText[:min(b.LineWidth-place, len(remainingText))]
+	currLineLength := b.countRemainingLineWidth(place)
+
+	currLine := remainingText[:min(currLineLength, len(remainingText))]
+	currLineSpace := runewidth.StringWidth(currLine)
+	remLength := runewidth.StringWidth(remainingText)
+
 	if len(currLine) > 0 {
 		fmt.Printf(ClearToEOL + currLine)
-		fmt.Print(cursorLeftN(len(currLine)))
+		fmt.Print(cursorLeftN(currLineSpace))
 	} else {
 		fmt.Print(ClearToEOL)
 	}
 
+	if currLineSpace != b.LineWidth-place && currLineSpace != remLength {
+		b.LineHasSpace.Set(b.DisplayPos/b.LineWidth, true)
+	} else if currLineSpace != b.LineWidth-place {
+		b.LineHasSpace.Remove(b.DisplayPos / b.LineWidth)
+	} else {
+		b.LineHasSpace.Set(b.DisplayPos/b.LineWidth, false)
+	}
+
+	if (b.DisplayPos+currLineSpace)%b.LineWidth == 0 && currLine == remainingText {
+		fmt.Print(cursorRightN(currLineSpace))
+		fmt.Printf("\n%s", b.Prompt.AltPrompt)
+		fmt.Printf(CursorUp + CursorBOL + cursorRightN(b.Width-currLineSpace))
+	}
+
 	// render the other lines
-	if len(remainingText) > len(currLine) {
-		remaining := []rune(remainingText[len(currLine):])
+	if remLength > currLineSpace {
+		remaining := (remainingText[len(currLine):])
 		var totalLines int
-		for i, c := range remaining {
-			if i%b.LineWidth == 0 {
+		var displayLength int
+		var lineLength int = currLineSpace
+
+		for _, c := range remaining {
+			if displayLength == 0 || (displayLength+runewidth.RuneWidth(c))%b.LineWidth < displayLength%b.LineWidth {
 				fmt.Printf("\n%s", b.Prompt.AltPrompt)
 				totalLines += 1
+
+				if displayLength != 0 {
+					if lineLength == b.LineWidth {
+						b.LineHasSpace.Set(b.DisplayPos/b.LineWidth+totalLines-1, false)
+					} else {
+						b.LineHasSpace.Set(b.DisplayPos/b.LineWidth+totalLines-1, true)
+					}
+				}
+
+				lineLength = 0
 			}
+
+			displayLength += runewidth.RuneWidth(c)
+			lineLength += runewidth.RuneWidth(c)
 			fmt.Printf("%c", c)
 		}
 		fmt.Print(ClearToEOL)
 		fmt.Print(cursorUpN(totalLines))
-		fmt.Printf(CursorBOL + cursorRightN(b.Width-len(currLine)))
+		fmt.Printf(CursorBOL + cursorRightN(b.Width-currLineSpace))
+
+		hasSpace := b.GetLineSpacing(b.DisplayPos / b.LineWidth)
+
+		if hasSpace && b.DisplayPos%b.LineWidth != b.LineWidth-1 {
+			fmt.Print(CursorLeft)
+		}
 	}
 
 	fmt.Print(CursorShow)
@@ -189,46 +356,84 @@ func (b *Buffer) drawRemaining() {
 
 func (b *Buffer) Remove() {
 	if b.Buf.Size() > 0 && b.Pos > 0 {
-		if b.Pos%b.LineWidth == 0 {
-			// if the user backspaces over the word boundary, do this magic to clear the line
-			// and move to the end of the previous line
-			fmt.Printf(CursorBOL + ClearToEOL)
-			fmt.Printf(CursorUp + CursorBOL + cursorRightN(b.Width) + " " + CursorLeft)
-		} else {
-			fmt.Printf(CursorLeft + " " + CursorLeft)
-		}
 
-		var eraseExtraLine bool
-		if (b.Size()-1)%b.LineWidth == 0 {
-			eraseExtraLine = true
-		}
+		if e, ok := b.Buf.Get(b.Pos - 1); ok {
+			if r, ok := e.(rune); ok {
+				rLength := runewidth.RuneWidth(r)
+				hasSpace := b.GetLineSpacing(b.DisplayPos/b.LineWidth - 1)
 
-		b.Pos -= 1
-		b.Buf.Remove(b.Pos)
+				if b.DisplayPos%b.LineWidth == 0 {
+					// if the user backspaces over the word boundary, do this magic to clear the line
+					// and move to the end of the previous line
+					fmt.Printf(CursorBOL + ClearToEOL)
+					fmt.Printf(CursorUp + CursorBOL + cursorRightN(b.Width))
 
-		if b.Pos < b.Size() {
-			b.drawRemaining()
-			// this erases a line which is left over when backspacing in the middle of a line and there
-			// are trailing characters which go over the line width boundary
-			if eraseExtraLine {
-				remainingLines := (b.Size() - b.Pos) / b.LineWidth
-				fmt.Printf(cursorDownN(remainingLines+1) + CursorBOL + ClearToEOL)
-				place := b.Pos % b.LineWidth
-				fmt.Printf(cursorUpN(remainingLines+1) + cursorRightN(place+len(b.Prompt.prompt())))
+					if b.DisplaySize()%b.LineWidth < (b.DisplaySize()-rLength)%b.LineWidth {
+						b.LineHasSpace.Remove(b.DisplayPos/b.LineWidth - 1)
+					}
+
+					if hasSpace {
+						b.DisplayPos -= 1
+						fmt.Print(CursorLeft)
+					}
+
+					if rLength == 2 {
+						fmt.Print(CursorLeft + "  " + cursorLeftN(2))
+					} else {
+						fmt.Print(" " + CursorLeft)
+					}
+
+				} else if (b.DisplayPos-rLength)%b.LineWidth == 0 && hasSpace {
+					fmt.Printf(CursorBOL + ClearToEOL)
+					fmt.Printf(CursorUp + CursorBOL + cursorRightN(b.Width))
+
+					if b.Pos == b.Buf.Size() {
+						b.LineHasSpace.Remove(b.DisplayPos/b.LineWidth - 1)
+					}
+					b.DisplayPos -= 1
+
+				} else {
+					fmt.Print(cursorLeftN(rLength))
+					for i := 0; i < rLength; i++ {
+						fmt.Print(" ")
+					}
+					fmt.Print(cursorLeftN(rLength))
+				}
+
+				var eraseExtraLine bool
+				if (b.DisplaySize()-1)%b.LineWidth == 0 || (rLength == 2 && ((b.DisplaySize()-2)%b.LineWidth == 0)) || b.DisplaySize()%b.LineWidth == 0 {
+					eraseExtraLine = true
+				}
+
+				b.Pos -= 1
+				b.DisplayPos -= rLength
+				b.Buf.Remove(b.Pos)
+
+				if b.Pos < b.Buf.Size() {
+					b.drawRemaining()
+					// this erases a line which is left over when backspacing in the middle of a line and there
+					// are trailing characters which go over the line width boundary
+					if eraseExtraLine {
+						remainingLines := (b.DisplaySize() - b.DisplayPos) / b.LineWidth
+						fmt.Printf(cursorDownN(remainingLines+1) + CursorBOL + ClearToEOL)
+						place := b.DisplayPos % b.LineWidth
+						fmt.Printf(cursorUpN(remainingLines+1) + cursorRightN(place+len(b.Prompt.prompt())))
+					}
+				}
 			}
 		}
 	}
 }
 
 func (b *Buffer) Delete() {
-	if b.Size() > 0 && b.Pos < b.Size() {
+	if b.Buf.Size() > 0 && b.Pos < b.Buf.Size() {
 		b.Buf.Remove(b.Pos)
 		b.drawRemaining()
-		if b.Size()%b.LineWidth == 0 {
-			if b.Pos != b.Size() {
-				remainingLines := (b.Size() - b.Pos) / b.LineWidth
+		if b.DisplaySize()%b.LineWidth == 0 {
+			if b.DisplayPos != b.DisplaySize() {
+				remainingLines := (b.DisplaySize() - b.DisplayPos) / b.LineWidth
 				fmt.Printf(cursorDownN(remainingLines) + CursorBOL + ClearToEOL)
-				place := b.Pos % b.LineWidth
+				place := b.DisplayPos % b.LineWidth
 				fmt.Printf(cursorUpN(remainingLines) + cursorRightN(place+len(b.Prompt.prompt())))
 			}
 		}
@@ -244,8 +449,8 @@ func (b *Buffer) DeleteBefore() {
 }
 
 func (b *Buffer) DeleteRemaining() {
-	if b.Size() > 0 && b.Pos < b.Size() {
-		charsToDel := b.Size() - b.Pos
+	if b.DisplaySize() > 0 && b.Pos < b.DisplaySize() {
+		charsToDel := b.Buf.Size() - b.Pos
 		for cnt := 0; cnt < charsToDel; cnt++ {
 			b.Delete()
 		}
@@ -281,8 +486,10 @@ func (b *Buffer) ClearScreen() {
 		ph := b.Prompt.placeholder()
 		fmt.Printf(ColorGrey + ph + cursorLeftN(len(ph)) + ColorDefault)
 	} else {
-		currPos := b.Pos
+		currPos := b.DisplayPos
+		currIndex := b.Pos
 		b.Pos = 0
+		b.DisplayPos = 0
 		b.drawRemaining()
 		fmt.Printf(CursorReset + cursorRightN(len(b.Prompt.prompt())))
 		if currPos > 0 {
@@ -300,7 +507,8 @@ func (b *Buffer) ClearScreen() {
 				fmt.Printf(CursorBOL + b.Prompt.AltPrompt)
 			}
 		}
-		b.Pos = currPos
+		b.Pos = currIndex
+		b.DisplayPos = currPos
 	}
 }
 
@@ -309,9 +517,20 @@ func (b *Buffer) IsEmpty() bool {
 }
 
 func (b *Buffer) Replace(r []rune) {
+	b.DisplayPos = 0
 	b.Pos = 0
+	lineNums := b.DisplaySize() / b.LineWidth
+
 	b.Buf.Clear()
-	fmt.Printf(ClearLine + CursorBOL + b.Prompt.prompt())
+
+	fmt.Printf(CursorBOL + ClearToEOL)
+
+	for i := 0; i < lineNums; i++ {
+		fmt.Print(CursorUp + CursorBOL + ClearToEOL)
+	}
+
+	fmt.Printf(CursorBOL + b.Prompt.prompt())
+
 	for _, c := range r {
 		b.Add(c)
 	}
@@ -328,7 +547,7 @@ func (b *Buffer) StringN(n int) string {
 func (b *Buffer) StringNM(n, m int) string {
 	var s string
 	if m == 0 {
-		m = b.Size()
+		m = b.Buf.Size()
 	}
 	for cnt := n; cnt < m; cnt++ {
 		c, _ := b.Buf.Get(cnt)

readline/readline.go

@@ -150,7 +150,7 @@ func (i *Instance) Readline() (string, error) {
 					i.Pasting = false
 				}
 			case KeyDel:
-				if buf.Size() > 0 {
+				if buf.DisplaySize() > 0 {
 					buf.Delete()
 				}
 				metaDel = true
@@ -202,7 +202,7 @@ func (i *Instance) Readline() (string, error) {
 				buf.Add(' ')
 			}
 		case CharDelete:
-			if buf.Size() > 0 {
+			if buf.DisplaySize() > 0 {
 				buf.Delete()
 			} else {
 				return "", io.EOF
@@ -218,7 +218,7 @@ func (i *Instance) Readline() (string, error) {
 		case CharCtrlZ:
 			fd := int(syscall.Stdin)
 			return handleCharCtrlZ(fd, i.Terminal.termios)
-		case CharEnter:
+		case CharEnter, CharCtrlJ:
 			output := buf.String()
 			if output != "" {
 				i.History.Add([]rune(output))
@@ -232,7 +232,7 @@ func (i *Instance) Readline() (string, error) {
 				metaDel = false
 				continue
 			}
-			if r >= CharSpace || r == CharEnter {
+			if r >= CharSpace || r == CharEnter || r == CharCtrlJ {
 				buf.Add(r)
 			}
 		}

scripts/install.sh

@@ -33,9 +33,11 @@ case "$ARCH" in
     *) error "Unsupported architecture: $ARCH" ;;
 esac
 
+IS_WSL2=false
+
 KERN=$(uname -r)
 case "$KERN" in
-    *icrosoft*WSL2 | *icrosoft*wsl2) ;;
+    *icrosoft*WSL2 | *icrosoft*wsl2) IS_WSL2=true;;
     *icrosoft) error "Microsoft WSL1 is not currently supported. Please upgrade to WSL2 with 'wsl --set-version <distro> 2'" ;;
     *) ;;
 esac
@@ -72,7 +74,7 @@ status "Installing ollama to $BINDIR..."
 $SUDO install -o0 -g0 -m755 -d $BINDIR
 $SUDO install -o0 -g0 -m755 $TEMP_DIR/ollama $BINDIR/ollama
 
-install_success() { 
+install_success() {
     status 'The Ollama API is now available at 127.0.0.1:11434.'
     status 'Install complete. Run "ollama" from the command line.'
 }
@@ -131,6 +133,17 @@ if available systemctl; then
     configure_systemd
 fi
 
+# WSL2 only supports GPUs via nvidia passthrough
+# so check for nvidia-smi to determine if GPU is available
+if [ "$IS_WSL2" = true ]; then
+    if available nvidia-smi && [ -n "$(nvidia-smi | grep -o "CUDA Version: [0-9]*\.[0-9]*")" ]; then
+        status "Nvidia GPU detected."
+    fi
+    install_success
+    exit 0
+fi
+
+# Install GPU dependencies on Linux
 if ! available lspci && ! available lshw; then
     warning "Unable to detect NVIDIA/AMD GPU. Install lspci or lshw to automatically detect and install GPU dependencies."
     exit 0
@@ -139,12 +152,12 @@ fi
 check_gpu() {
     # Look for devices based on vendor ID for NVIDIA and AMD
     case $1 in
-        lspci) 
+        lspci)
             case $2 in
                 nvidia) available lspci && lspci -d '10de:' | grep -q 'NVIDIA' || return 1 ;;
                 amdgpu) available lspci && lspci -d '1002:' | grep -q 'AMD' || return 1 ;;
             esac ;;
-        lshw) 
+        lshw)
             case $2 in
                 nvidia) available lshw && $SUDO lshw -c display -numeric | grep -q 'vendor: .* \[10DE\]' || return 1 ;;
                 amdgpu) available lshw && $SUDO lshw -c display -numeric | grep -q 'vendor: .* \[1002\]' || return 1 ;;
@@ -181,7 +194,7 @@ if check_gpu lspci amdgpu || check_gpu lshw amdgpu; then
     curl --fail --show-error --location --progress-bar "https://ollama.com/download/ollama-linux-amd64-rocm.tgz${VER_PARAM}" \
         | $SUDO tar zx --owner ollama --group ollama -C /usr/share/ollama/lib/rocm .
     install_success
-    status "AMD GPU dependencies installed."
+    status "AMD GPU ready."
     exit 0
 fi
 
@@ -274,7 +287,7 @@ if ! check_gpu nvidia-smi || [ -z "$(nvidia-smi | grep -o "CUDA Version: [0-9]*\
     esac
 fi
 
-if ! lsmod | grep -q nvidia; then
+if ! lsmod | grep -q nvidia || ! lsmod | grep -q nvidia_uvm; then
     KERNEL_RELEASE="$(uname -r)"
     case $OS_NAME in
         rocky) $SUDO $PACKAGE_MANAGER -y install kernel-devel kernel-headers ;;
@@ -295,7 +308,19 @@ if ! lsmod | grep -q nvidia; then
     fi
 
     $SUDO modprobe nvidia
+    $SUDO modprobe nvidia_uvm
 fi
 
+# make sure the NVIDIA modules are loaded on boot with nvidia-persistenced
+if command -v nvidia-persistenced > /dev/null 2>&1; then
+    $SUDO touch /etc/modules-load.d/nvidia.conf
+    MODULES="nvidia nvidia-uvm"
+    for MODULE in $MODULES; do
+        if ! grep -qxF "$MODULE" /etc/modules-load.d/nvidia.conf; then
+            echo "$MODULE" | sudo tee -a /etc/modules-load.d/nvidia.conf > /dev/null
+        fi
+    done
+fi
 
-status "NVIDIA CUDA drivers installed."
+status "NVIDIA GPU ready."
+install_success

server/download.go

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

server/envconfig/config_test.go

@@ -1,20 +0,0 @@
-package envconfig
-
-import (
-	"os"
-	"testing"
-
-	"github.com/stretchr/testify/require"
-)
-
-func TestConfig(t *testing.T) {
-	os.Setenv("OLLAMA_DEBUG", "")
-	LoadConfig()
-	require.False(t, Debug)
-	os.Setenv("OLLAMA_DEBUG", "false")
-	LoadConfig()
-	require.False(t, Debug)
-	os.Setenv("OLLAMA_DEBUG", "1")
-	LoadConfig()
-	require.True(t, Debug)
-}

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"
@@ -26,10 +25,10 @@ import (
 
 	"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/server/envconfig"
+	"github.com/ollama/ollama/parser"
+	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/types/errtypes"
 	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/version"
@@ -54,7 +53,6 @@ type Model struct {
 	System         string
 	License        []string
 	Digest         string
-	Size           int64
 	Options        map[string]interface{}
 	Messages       []Message
 }
@@ -64,52 +62,52 @@ func (m *Model) IsEmbedding() bool {
 }
 
 func (m *Model) String() string {
-	var modelfile model.File
+	var modelfile parser.File
 
-	modelfile.Commands = append(modelfile.Commands, model.Command{
+	modelfile.Commands = append(modelfile.Commands, parser.Command{
 		Name: "model",
 		Args: m.ModelPath,
 	})
 
-	if m.Template != "" {
-		modelfile.Commands = append(modelfile.Commands, model.Command{
-			Name: "template",
-			Args: m.Template,
-		})
-	}
-
-	if m.System != "" {
-		modelfile.Commands = append(modelfile.Commands, model.Command{
-			Name: "system",
-			Args: m.System,
-		})
-	}
-
 	for _, adapter := range m.AdapterPaths {
-		modelfile.Commands = append(modelfile.Commands, model.Command{
+		modelfile.Commands = append(modelfile.Commands, parser.Command{
 			Name: "adapter",
 			Args: adapter,
 		})
 	}
 
 	for _, projector := range m.ProjectorPaths {
-		modelfile.Commands = append(modelfile.Commands, model.Command{
-			Name: "projector",
+		modelfile.Commands = append(modelfile.Commands, parser.Command{
+			Name: "model",
 			Args: projector,
 		})
 	}
 
+	if m.Template != "" {
+		modelfile.Commands = append(modelfile.Commands, parser.Command{
+			Name: "template",
+			Args: m.Template,
+		})
+	}
+
+	if m.System != "" {
+		modelfile.Commands = append(modelfile.Commands, parser.Command{
+			Name: "system",
+			Args: m.System,
+		})
+	}
+
 	for k, v := range m.Options {
 		switch v := v.(type) {
 		case []any:
 			for _, s := range v {
-				modelfile.Commands = append(modelfile.Commands, model.Command{
+				modelfile.Commands = append(modelfile.Commands, parser.Command{
 					Name: k,
 					Args: fmt.Sprintf("%v", s),
 				})
 			}
 		default:
-			modelfile.Commands = append(modelfile.Commands, model.Command{
+			modelfile.Commands = append(modelfile.Commands, parser.Command{
 				Name: k,
 				Args: fmt.Sprintf("%v", v),
 			})
@@ -117,14 +115,14 @@ func (m *Model) String() string {
 	}
 
 	for _, license := range m.License {
-		modelfile.Commands = append(modelfile.Commands, model.Command{
+		modelfile.Commands = append(modelfile.Commands, parser.Command{
 			Name: "license",
 			Args: license,
 		})
 	}
 
 	for _, msg := range m.Messages {
-		modelfile.Commands = append(modelfile.Commands, model.Command{
+		modelfile.Commands = append(modelfile.Commands, parser.Command{
 			Name: "message",
 			Args: fmt.Sprintf("%s %s", msg.Role, msg.Content),
 		})
@@ -158,50 +156,11 @@ 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"`
 }
 
-func (m *ManifestV2) GetTotalSize() (total int64) {
-	for _, layer := range m.Layers {
-		total += layer.Size
-	}
-
-	total += m.Config.Size
-	return total
-}
-
 func GetManifest(mp ModelPath) (*ManifestV2, string, error) {
 	fp, err := mp.GetManifestPath()
 	if err != nil {
@@ -242,7 +201,6 @@ func GetModel(name string) (*Model, error) {
 		Digest:    digest,
 		Template:  "{{ .Prompt }}",
 		License:   []string{},
-		Size:      manifest.GetTotalSize(),
 	}
 
 	filename, err := GetBlobsPath(manifest.Config.Digest)
@@ -332,7 +290,7 @@ 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
@@ -349,22 +307,15 @@ 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, modelfile *model.File, 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, modelfile *parser.File, fn func(resp api.ProgressResponse)) (err error) {
 	config := ConfigV2{
 		OS:           "linux",
 		Architecture: "amd64",
@@ -373,250 +324,205 @@ func CreateModel(ctx context.Context, name, modelFileDir, quantization string, m
 		},
 	}
 
-	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 modelfile.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
 				}
-
-				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)
+			} else if strings.HasPrefix(c.Args, "@") {
+				digest := strings.TrimPrefix(c.Args, "@")
+				if ib, ok := intermediateBlobs[digest]; ok {
+					p, err := GetBlobsPath(ib)
 					if err != nil {
 						return err
 					}
-					defer os.RemoveAll(tempfile.Name())
 
-					if err := llm.Quantize(ggufName, tempfile.Name(), quantization); err != nil {
+					if _, err := os.Stat(p); errors.Is(err, os.ErrNotExist) {
+						// pass
+					} else if err != nil {
 						return err
+					} else {
+						fn(api.ProgressResponse{Status: fmt.Sprintf("using cached layer %s", ib)})
+						digest = ib
 					}
-
-					if err := tempfile.Close(); err != nil {
-						return err
-					}
-
-					pathName = tempfile.Name()
 				}
+
+				blobpath, err := GetBlobsPath(digest)
+				if err != nil {
+					return err
+				}
+
+				blob, err := os.Open(blobpath)
+				if err != nil {
+					return err
+				}
+				defer blob.Close()
+
+				baseLayers, err = parseFromFile(ctx, blob, digest, fn)
+				if err != nil {
+					return err
+				}
+			} else if file, err := os.Open(realpath(modelFileDir, c.Args)); err == nil {
+				defer file.Close()
+
+				baseLayers, err = parseFromFile(ctx, file, "", fn)
+				if err != nil {
+					return err
+				}
+			} else {
+				return fmt.Errorf("invalid model reference: %s", c.Args)
 			}
 
-			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)
+			for _, baseLayer := range baseLayers {
+				if quantization != "" &&
+					baseLayer.MediaType == "application/vnd.ollama.image.model" &&
+					baseLayer.GGML != nil &&
+					baseLayer.GGML.Name() == "gguf" {
+					want, err := llm.ParseFileType(quantization)
 					if err != nil {
 						return err
 					}
-				case err != nil:
-					return err
-				}
 
-				fn(api.ProgressResponse{Status: "reading model metadata"})
-				fromConfigPath, err := GetBlobsPath(manifest.Config.Digest)
-				if err != nil {
-					return err
-				}
+					ft := baseLayer.GGML.KV().FileType()
+					if !slices.Contains([]string{"F16", "F32"}, ft.String()) {
+						return errors.New("quantization is only supported for F16 and F32 models")
+					} else if want != ft {
+						fn(api.ProgressResponse{Status: fmt.Sprintf("quantizing %s model to %s", ft, quantization)})
 
-				fromConfigFile, err := os.Open(fromConfigPath)
-				if err != nil {
-					return err
-				}
-				defer fromConfigFile.Close()
-
-				var fromConfig ConfigV2
-				if err := json.NewDecoder(fromConfigFile).Decode(&fromConfig); err != nil {
-					return err
-				}
-
-				// 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)
+						blob, err := GetBlobsPath(baseLayer.Digest)
 						if err != nil {
 							return err
 						}
 
-						fromParamsFile, err := os.Open(fromParamsPath)
+						temp, err := os.CreateTemp(filepath.Dir(blob), quantization)
 						if err != nil {
 							return err
 						}
-						defer fromParamsFile.Close()
+						defer temp.Close()
+						defer os.Remove(temp.Name())
 
-						if err := json.NewDecoder(fromParamsFile).Decode(&fromParams); err != nil {
+						if err := llm.Quantize(blob, temp.Name(), want); err != nil {
 							return err
 						}
+
+						layers, err := parseFromFile(ctx, temp, "", fn)
+						if err != nil {
+							return err
+						}
+
+						if len(layers) != 1 {
+							return errors.New("quantization failed")
+						}
+
+						baseLayer.Layer = layers[0].Layer
+						baseLayer.GGML = layers[0].GGML
 					}
-
-					layer, err := NewLayerFromLayer(layer.Digest, layer.MediaType, modelpath.GetShortTagname())
-					if err != nil {
-						return err
-					}
-
-					layers.Add(layer)
 				}
 
-				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().String())
+					config.ModelFamilies = append(config.ModelFamilies, baseLayer.GGML.KV().Architecture())
+				}
+
+				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
 		}
 
@@ -625,39 +531,25 @@ func CreateModel(ctx context.Context, name, modelFileDir, quantization string, m
 			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
 	}
 
@@ -668,36 +560,37 @@ func CreateModel(ctx context.Context, name, modelFileDir, quantization string, m
 		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.status != "" {
+			fn(api.ProgressResponse{Status: layer.status})
+		}
+	}
 
-	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 !envconfig.NoPrune {
-		if err := deleteUnusedLayers(nil, deleteMap, false); err != nil {
+		if err := deleteUnusedLayers(nil, unref); err != nil {
 			return err
 		}
 	}
@@ -706,74 +599,6 @@ func CreateModel(ctx context.Context, name, modelFileDir, quantization string, m
 	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)
@@ -813,7 +638,7 @@ func CopyModel(src, dst model.Name) error {
 	return err
 }
 
-func deleteUnusedLayers(skipModelPath *ModelPath, deleteMap map[string]struct{}, dryRun bool) error {
+func deleteUnusedLayers(skipModelPath *ModelPath, deleteMap map[string]struct{}) error {
 	fp, err := GetManifestPath()
 	if err != nil {
 		return err
@@ -860,13 +685,9 @@ func deleteUnusedLayers(skipModelPath *ModelPath, deleteMap map[string]struct{},
 			slog.Info(fmt.Sprintf("couldn't get file path for '%s': %v", k, err))
 			continue
 		}
-		if !dryRun {
-			if err := os.Remove(fp); err != nil {
-				slog.Info(fmt.Sprintf("couldn't remove file '%s': %v", fp, err))
-				continue
-			}
-		} else {
-			slog.Info(fmt.Sprintf("wanted to remove: %s", fp))
+		if err := os.Remove(fp); err != nil {
+			slog.Info(fmt.Sprintf("couldn't remove file '%s': %v", fp, err))
+			continue
 		}
 	}
 
@@ -889,14 +710,25 @@ func PruneLayers() error {
 	for _, blob := range blobs {
 		name := blob.Name()
 		name = strings.ReplaceAll(name, "-", ":")
-		if strings.HasPrefix(name, "sha256:") {
-			deleteMap[name] = struct{}{}
+
+		_, err := GetBlobsPath(name)
+		if err != nil {
+			if errors.Is(err, ErrInvalidDigestFormat) {
+				// remove invalid blobs (e.g. partial downloads)
+				if err := os.Remove(filepath.Join(p, blob.Name())); err != nil {
+					slog.Error("couldn't remove blob", "blob", blob.Name(), "error", err)
+				}
+			}
+
+			continue
 		}
+
+		deleteMap[name] = struct{}{}
 	}
 
 	slog.Info(fmt.Sprintf("total blobs: %d", len(deleteMap)))
 
-	err = deleteUnusedLayers(nil, deleteMap, false)
+	err = deleteUnusedLayers(nil, deleteMap)
 	if err != nil {
 		return err
 	}
@@ -952,7 +784,7 @@ func DeleteModel(name string) error {
 	}
 	deleteMap[manifest.Config.Digest] = struct{}{}
 
-	err = deleteUnusedLayers(&mp, deleteMap, false)
+	err = deleteUnusedLayers(&mp, deleteMap)
 	if err != nil {
 		return err
 	}
@@ -1112,7 +944,7 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 
 	if noprune == "" {
 		fn(api.ProgressResponse{Status: "removing any unused layers"})
-		err = deleteUnusedLayers(nil, deleteMap, false)
+		err = deleteUnusedLayers(nil, deleteMap)
 		if err != nil {
 			return err
 		}

server/layer.go

@@ -5,39 +5,14 @@ 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
+	status    string
 }
 
 func NewLayer(r io.Reader, mediatype string) (*Layer, error) {
@@ -46,14 +21,12 @@ 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 +34,29 @@ 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,
+		status:    fmt.Sprintf("%s %s", status, digest),
 	}, nil
 }
 
@@ -85,21 +76,15 @@ func NewLayerFromLayer(digest, mediatype, from string) (*Layer, error) {
 		Digest:    digest,
 		Size:      fi.Size(),
 		From:      from,
+		status:    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() (io.ReadSeekCloser, 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/manifest.go

@@ -0,0 +1,79 @@
+package server
+
+import (
+	"bytes"
+	"crypto/sha256"
+	"encoding/json"
+	"fmt"
+	"io"
+	"os"
+	"path/filepath"
+
+	"github.com/ollama/ollama/types/model"
+)
+
+type Manifest struct {
+	ManifestV2
+	Digest string `json:"-"`
+}
+
+func (m *Manifest) Size() (size int64) {
+	for _, layer := range append(m.Layers, m.Config) {
+		size += layer.Size
+	}
+
+	return
+}
+
+func ParseNamedManifest(name model.Name) (*Manifest, error) {
+	if !name.IsFullyQualified() {
+		return nil, model.Unqualified(name)
+	}
+
+	manifests, err := GetManifestPath()
+	if err != nil {
+		return nil, err
+	}
+
+	var manifest ManifestV2
+	manifestfile, err := os.Open(filepath.Join(manifests, name.Filepath()))
+	if err != nil {
+		return nil, err
+	}
+
+	sha256sum := sha256.New()
+	if err := json.NewDecoder(io.TeeReader(manifestfile, sha256sum)).Decode(&manifest); err != nil {
+		return nil, err
+	}
+
+	return &Manifest{
+		ManifestV2: manifest,
+		Digest:     fmt.Sprintf("%x", sha256sum.Sum(nil)),
+	}, nil
+}
+
+func WriteManifest(name string, config *Layer, layers []*Layer) error {
+	manifest := ManifestV2{
+		SchemaVersion: 2,
+		MediaType:     "application/vnd.docker.distribution.manifest.v2+json",
+		Config:        config,
+		Layers:        layers,
+	}
+
+	var b bytes.Buffer
+	if err := json.NewEncoder(&b).Encode(manifest); err != nil {
+		return err
+	}
+
+	modelpath := ParseModelPath(name)
+	manifestPath, err := modelpath.GetManifestPath()
+	if err != nil {
+		return err
+	}
+
+	if err := os.MkdirAll(filepath.Dir(manifestPath), 0o755); err != nil {
+		return err
+	}
+
+	return os.WriteFile(manifestPath, b.Bytes(), 0o644)
+}

server/manifests.go

@@ -1,34 +0,0 @@
-package server
-
-import (
-	"bytes"
-	"encoding/json"
-	"os"
-	"path/filepath"
-)
-
-func WriteManifest(name string, config *Layer, layers []*Layer) error {
-	manifest := ManifestV2{
-		SchemaVersion: 2,
-		MediaType:     "application/vnd.docker.distribution.manifest.v2+json",
-		Config:        config,
-		Layers:        layers,
-	}
-
-	var b bytes.Buffer
-	if err := json.NewEncoder(&b).Encode(manifest); err != nil {
-		return err
-	}
-
-	modelpath := ParseModelPath(name)
-	manifestPath, err := modelpath.GetManifestPath()
-	if err != nil {
-		return err
-	}
-
-	if err := os.MkdirAll(filepath.Dir(manifestPath), 0o755); err != nil {
-		return err
-	}
-
-	return os.WriteFile(manifestPath, b.Bytes(), 0o644)
-}

server/model.go

@@ -0,0 +1,255 @@
+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"
+)
+
+var intermediateBlobs map[string]string = make(map[string]string)
+
+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, digest string, 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
+		}
+		defer outfile.Close()
+
+		infile, err := f.Open()
+		if err != nil {
+			return nil, err
+		}
+		defer infile.Close()
+
+		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, err
+	}
+
+	bin, err := layer.Open()
+	if err != nil {
+		return nil, err
+	}
+	defer bin.Close()
+
+	ggml, _, err := llm.DecodeGGML(bin)
+	if err != nil {
+		return nil, err
+	}
+
+	layers = append(layers, &layerWithGGML{layer, ggml})
+
+	intermediateBlobs[digest] = layer.Digest
+	return layers, nil
+}
+
+func parseFromFile(ctx context.Context, file *os.File, digest string, 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, digest, 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/modelpath.go

@@ -154,9 +154,6 @@ func GetBlobsPath(digest string) (string, error) {
 	// only accept actual sha256 digests
 	pattern := "^sha256[:-][0-9a-fA-F]{64}$"
 	re := regexp.MustCompile(pattern)
-	if err != nil {
-		return "", err
-	}
 
 	if digest != "" && !re.MatchString(digest) {
 		return "", ErrInvalidDigestFormat

server/routes.go

@@ -26,10 +26,12 @@ import (
 	"golang.org/x/exp/slices"
 
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/gpu"
 	"github.com/ollama/ollama/llm"
 	"github.com/ollama/ollama/openai"
-	"github.com/ollama/ollama/server/envconfig"
+	"github.com/ollama/ollama/parser"
+	"github.com/ollama/ollama/types/errtypes"
 	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/version"
 )
@@ -127,10 +129,6 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 
 	opts, err := modelOptions(model, req.Options)
 	if err != nil {
-		if errors.Is(err, api.ErrInvalidOpts) {
-			c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
-			return
-		}
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 		return
 	}
@@ -156,9 +154,10 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 	// of `raw` mode so we need to check for it too
 	if req.Prompt == "" && req.Template == "" && req.System == "" {
 		c.JSON(http.StatusOK, api.GenerateResponse{
-			CreatedAt: time.Now().UTC(),
-			Model:     req.Model,
-			Done:      true,
+			CreatedAt:  time.Now().UTC(),
+			Model:      req.Model,
+			Done:       true,
+			DoneReason: "load",
 		})
 		return
 	}
@@ -226,10 +225,11 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 			}
 
 			resp := api.GenerateResponse{
-				Model:     req.Model,
-				CreatedAt: time.Now().UTC(),
-				Done:      r.Done,
-				Response:  r.Content,
+				Model:      req.Model,
+				CreatedAt:  time.Now().UTC(),
+				Done:       r.Done,
+				Response:   r.Content,
+				DoneReason: r.DoneReason,
 				Metrics: api.Metrics{
 					PromptEvalCount:    r.PromptEvalCount,
 					PromptEvalDuration: r.PromptEvalDuration,
@@ -315,10 +315,10 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 }
 
 func getDefaultSessionDuration() time.Duration {
-	if t, exists := os.LookupEnv("OLLAMA_KEEP_ALIVE"); exists {
-		v, err := strconv.Atoi(t)
+	if envconfig.KeepAlive != "" {
+		v, err := strconv.Atoi(envconfig.KeepAlive)
 		if err != nil {
-			d, err := time.ParseDuration(t)
+			d, err := time.ParseDuration(envconfig.KeepAlive)
 			if err != nil {
 				return defaultSessionDuration
 			}
@@ -370,10 +370,6 @@ func (s *Server) EmbeddingsHandler(c *gin.Context) {
 
 	opts, err := modelOptions(model, req.Options)
 	if err != nil {
-		if errors.Is(err, api.ErrInvalidOpts) {
-			c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
-			return
-		}
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 		return
 	}
@@ -523,7 +519,7 @@ func (s *Server) CreateModelHandler(c *gin.Context) {
 
 	name := model.ParseName(cmp.Or(req.Model, req.Name))
 	if !name.IsValid() {
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "invalid model name"})
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": errtypes.InvalidModelNameErrMsg})
 		return
 	}
 
@@ -544,7 +540,7 @@ func (s *Server) CreateModelHandler(c *gin.Context) {
 		r = f
 	}
 
-	modelfile, err := model.ParseFile(r)
+	modelfile, err := parser.ParseFile(r)
 	if err != nil {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
 		return
@@ -560,7 +556,12 @@ func (s *Server) CreateModelHandler(c *gin.Context) {
 		ctx, cancel := context.WithCancel(c.Request.Context())
 		defer cancel()
 
-		if err := CreateModel(ctx, name.String(), filepath.Dir(req.Path), req.Quantization, modelfile, fn); err != nil {
+		quantization := req.Quantization
+		if req.Quantize != "" {
+			quantization = req.Quantize
+		}
+
+		if err := CreateModel(ctx, name.String(), filepath.Dir(req.Path), strings.ToUpper(quantization), modelfile, fn); err != nil {
 			ch <- gin.H{"error": err.Error()}
 		}
 	}()
@@ -709,7 +710,7 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 	}
 
 	var sb strings.Builder
-	fmt.Fprintln(&sb, "# Modelfile generate by \"ollama show\"")
+	fmt.Fprintln(&sb, "# Modelfile generated by \"ollama show\"")
 	fmt.Fprintln(&sb, "# To build a new Modelfile based on this, replace FROM with:")
 	fmt.Fprintf(&sb, "# FROM %s\n\n", model.ShortName)
 	fmt.Fprint(&sb, model.String())
@@ -719,62 +720,79 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 }
 
 func (s *Server) ListModelsHandler(c *gin.Context) {
-	models := make([]api.ModelResponse, 0)
-	manifestsPath, err := GetManifestPath()
+	manifests, err := GetManifestPath()
 	if err != nil {
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 		return
 	}
 
-	modelResponse := func(modelName string) (api.ModelResponse, error) {
-		model, err := GetModel(modelName)
-		if err != nil {
-			return api.ModelResponse{}, err
-		}
-
-		modelDetails := api.ModelDetails{
-			Format:            model.Config.ModelFormat,
-			Family:            model.Config.ModelFamily,
-			Families:          model.Config.ModelFamilies,
-			ParameterSize:     model.Config.ModelType,
-			QuantizationLevel: model.Config.FileType,
-		}
-
-		return api.ModelResponse{
-			Model:   model.ShortName,
-			Name:    model.ShortName,
-			Size:    model.Size,
-			Digest:  model.Digest,
-			Details: modelDetails,
-		}, nil
-	}
-
-	walkFunc := func(path string, info os.FileInfo, _ error) error {
+	models := []api.ModelResponse{}
+	if err := filepath.Walk(manifests, func(path string, info os.FileInfo, _ error) error {
 		if !info.IsDir() {
-			path, tag := filepath.Split(path)
-			model := strings.Trim(strings.TrimPrefix(path, manifestsPath), string(os.PathSeparator))
-			modelPath := strings.Join([]string{model, tag}, ":")
-			canonicalModelPath := strings.ReplaceAll(modelPath, string(os.PathSeparator), "/")
-
-			resp, err := modelResponse(canonicalModelPath)
+			rel, err := filepath.Rel(manifests, path)
 			if err != nil {
-				slog.Info(fmt.Sprintf("skipping file: %s", canonicalModelPath))
-				// nolint: nilerr
+				return err
+			}
+
+			if hidden, err := filepath.Match(".*", filepath.Base(rel)); err != nil {
+				return err
+			} else if hidden {
 				return nil
 			}
 
-			resp.ModifiedAt = info.ModTime()
-			models = append(models, resp)
+			n := model.ParseNameFromFilepath(rel)
+			if !n.IsValid() {
+				slog.Warn("bad manifest filepath", "path", rel)
+				return nil
+			}
+
+			m, err := ParseNamedManifest(n)
+			if err != nil {
+				slog.Warn("bad manifest", "name", n, "error", err)
+				return nil
+			}
+
+			f, err := m.Config.Open()
+			if err != nil {
+				slog.Warn("bad manifest config filepath", "name", n, "error", err)
+				return nil
+			}
+			defer f.Close()
+
+			var c ConfigV2
+			if err := json.NewDecoder(f).Decode(&c); err != nil {
+				slog.Warn("bad manifest config", "name", n, "error", err)
+				return nil
+			}
+
+			// tag should never be masked
+			models = append(models, api.ModelResponse{
+				Model:      n.DisplayShortest(),
+				Name:       n.DisplayShortest(),
+				Size:       m.Size(),
+				Digest:     m.Digest,
+				ModifiedAt: info.ModTime(),
+				Details: api.ModelDetails{
+					Format:            c.ModelFormat,
+					Family:            c.ModelFamily,
+					Families:          c.ModelFamilies,
+					ParameterSize:     c.ModelType,
+					QuantizationLevel: c.FileType,
+				},
+			})
 		}
 
 		return nil
-	}
-
-	if err := filepath.Walk(manifestsPath, walkFunc); err != nil {
+	}); err != nil {
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 		return
 	}
 
+	slices.SortStableFunc(models, func(i, j api.ModelResponse) int {
+		// most recently modified first
+		return cmp.Compare(j.ModifiedAt.Unix(), i.ModifiedAt.Unix())
+	})
+
 	c.JSON(http.StatusOK, api.ListResponse{Models: models})
 }
 
@@ -796,7 +814,7 @@ func (s *Server) CopyModelHandler(c *gin.Context) {
 
 	dst := model.ParseName(r.Destination)
 	if !dst.IsValid() {
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("destination %q is invalid", r.Source)})
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("destination %q is invalid", r.Destination)})
 		return
 	}
 
@@ -823,6 +841,25 @@ func (s *Server) HeadBlobHandler(c *gin.Context) {
 }
 
 func (s *Server) CreateBlobHandler(c *gin.Context) {
+	if ib, ok := intermediateBlobs[c.Param("digest")]; ok {
+		p, err := GetBlobsPath(ib)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+			return
+		}
+
+		if _, err := os.Stat(p); errors.Is(err, os.ErrNotExist) {
+			slog.Info("evicting intermediate blob which no longer exists", "digest", ib)
+			delete(intermediateBlobs, c.Param("digest"))
+		} else if err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+			return
+		} else {
+			c.Status(http.StatusOK)
+			return
+		}
+	}
+
 	path, err := GetBlobsPath(c.Param("digest"))
 	if err != nil {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
@@ -852,11 +889,6 @@ 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)
 }
 
@@ -931,6 +963,11 @@ func allowedHostsMiddleware(addr net.Addr) gin.HandlerFunc {
 		}
 
 		if allowedHost(host) {
+			if c.Request.Method == "OPTIONS" {
+				c.AbortWithStatus(http.StatusNoContent)
+				return
+			}
+
 			c.Next()
 			return
 		}
@@ -943,6 +980,7 @@ func (s *Server) GenerateRoutes() http.Handler {
 	config := cors.DefaultConfig()
 	config.AllowWildcard = true
 	config.AllowBrowserExtensions = true
+	config.AllowHeaders = []string{"Authorization", "Content-Type", "User-Agent", "Accept", "X-Requested-With"}
 	config.AllowOrigins = envconfig.AllowOrigins
 
 	r := gin.Default()
@@ -962,6 +1000,7 @@ func (s *Server) GenerateRoutes() http.Handler {
 	r.POST("/api/show", s.ShowModelHandler)
 	r.POST("/api/blobs/:digest", s.CreateBlobHandler)
 	r.HEAD("/api/blobs/:digest", s.HeadBlobHandler)
+	r.GET("/api/ps", s.ProcessHandler)
 
 	// Compatibility endpoints
 	r.POST("/v1/chat/completions", openai.Middleware(), s.ChatHandler)
@@ -986,7 +1025,7 @@ func Serve(ln net.Listener) error {
 		level = slog.LevelDebug
 	}
 
-	slog.Info("server config", "env", envconfig.AsMap())
+	slog.Info("server config", "env", envconfig.Values())
 	handler := slog.NewTextHandler(os.Stderr, &slog.HandlerOptions{
 		Level:     level,
 		AddSource: true,
@@ -1027,7 +1066,8 @@ func Serve(ln net.Listener) error {
 	}
 
 	ctx, done := context.WithCancel(context.Background())
-	sched := InitScheduler(ctx)
+	schedCtx, schedDone := context.WithCancel(ctx)
+	sched := InitScheduler(schedCtx)
 	s := &Server{addr: ln.Addr(), sched: sched}
 	r := s.GenerateRoutes()
 
@@ -1042,23 +1082,31 @@ func Serve(ln net.Listener) error {
 	go func() {
 		<-signals
 		srvr.Close()
-		done()
+		schedDone()
 		sched.unloadAllRunners()
 		gpu.Cleanup()
-		os.Exit(0)
+		done()
 	}()
 
 	if err := llm.Init(); err != nil {
 		return fmt.Errorf("unable to initialize llm library %w", err)
 	}
 
-	s.sched.Run(ctx)
+	s.sched.Run(schedCtx)
 
 	// At startup we retrieve GPU information so we can get log messages before loading a model
 	// This will log warnings to the log in case we have problems with detected GPUs
-	_ = gpu.GetGPUInfo()
+	gpus := gpu.GetGPUInfo()
+	gpus.LogDetails()
 
-	return srvr.Serve(ln)
+	err = srvr.Serve(ln)
+	// If server is closed from the signal handler, wait for the ctx to be done
+	// otherwise error out quickly
+	if !errors.Is(err, http.ErrServerClosed) {
+		return err
+	}
+	<-ctx.Done()
+	return nil
 }
 
 func waitForStream(c *gin.Context, ch chan interface{}) {
@@ -1111,6 +1159,42 @@ func streamResponse(c *gin.Context, ch chan any) {
 	})
 }
 
+func (s *Server) ProcessHandler(c *gin.Context) {
+	models := []api.ModelResponse{}
+
+	for _, v := range s.sched.loaded {
+		model := v.model
+		modelDetails := api.ModelDetails{
+			Format:            model.Config.ModelFormat,
+			Family:            model.Config.ModelFamily,
+			Families:          model.Config.ModelFamilies,
+			ParameterSize:     model.Config.ModelType,
+			QuantizationLevel: model.Config.FileType,
+		}
+
+		mr := api.ModelResponse{
+			Model:     model.ShortName,
+			Name:      model.ShortName,
+			Size:      int64(v.estimatedTotal),
+			SizeVRAM:  int64(v.estimatedVRAM),
+			Digest:    model.Digest,
+			Details:   modelDetails,
+			ExpiresAt: v.expiresAt,
+		}
+		// The scheduler waits to set expiresAt, so if a model is loading it's
+		// possible that it will be set to the unix epoch. For those cases, just
+		// calculate the time w/ the sessionDuration instead.
+		var epoch time.Time
+		if v.expiresAt == epoch {
+			mr.ExpiresAt = time.Now().Add(v.sessionDuration)
+		}
+
+		models = append(models, mr)
+	}
+
+	c.JSON(http.StatusOK, api.ListResponse{Models: models})
+}
+
 // ChatPrompt builds up a prompt from a series of messages for the currently `loaded` model
 func chatPrompt(ctx context.Context, runner *runnerRef, template string, messages []api.Message, numCtx int) (string, error) {
 	encode := func(s string) ([]int, error) {
@@ -1167,10 +1251,6 @@ func (s *Server) ChatHandler(c *gin.Context) {
 
 	opts, err := modelOptions(model, req.Options)
 	if err != nil {
-		if errors.Is(err, api.ErrInvalidOpts) {
-			c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
-			return
-		}
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 		return
 	}
@@ -1212,10 +1292,11 @@ func (s *Server) ChatHandler(c *gin.Context) {
 	// an empty request loads the model
 	if len(req.Messages) == 0 || prompt == "" {
 		resp := api.ChatResponse{
-			CreatedAt: time.Now().UTC(),
-			Model:     req.Model,
-			Done:      true,
-			Message:   api.Message{Role: "assistant"},
+			CreatedAt:  time.Now().UTC(),
+			Model:      req.Model,
+			Done:       true,
+			DoneReason: "load",
+			Message:    api.Message{Role: "assistant"},
 		}
 		c.JSON(http.StatusOK, resp)
 		return
@@ -1248,10 +1329,11 @@ func (s *Server) ChatHandler(c *gin.Context) {
 		fn := func(r llm.CompletionResponse) {
 
 			resp := api.ChatResponse{
-				Model:     req.Model,
-				CreatedAt: time.Now().UTC(),
-				Message:   api.Message{Role: "assistant", Content: r.Content},
-				Done:      r.Done,
+				Model:      req.Model,
+				CreatedAt:  time.Now().UTC(),
+				Message:    api.Message{Role: "assistant", Content: r.Content},
+				Done:       r.Done,
+				DoneReason: r.DoneReason,
 				Metrics: api.Metrics{
 					PromptEvalCount:    r.PromptEvalCount,
 					PromptEvalDuration: r.PromptEvalDuration,

server/routes_test.go

@@ -17,7 +17,7 @@ import (
 	"github.com/stretchr/testify/assert"
 
 	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/types/model"
+	"github.com/ollama/ollama/parser"
 	"github.com/ollama/ollama/version"
 )
 
@@ -56,7 +56,7 @@ func Test_Routes(t *testing.T) {
 		fname := createTestFile(t, "ollama-model")
 
 		r := strings.NewReader(fmt.Sprintf("FROM %s\nPARAMETER seed 42\nPARAMETER top_p 0.9\nPARAMETER stop foo\nPARAMETER stop bar", fname))
-		modelfile, err := model.ParseFile(r)
+		modelfile, err := parser.ParseFile(r)
 		assert.Nil(t, err)
 		fn := func(resp api.ProgressResponse) {
 			t.Logf("Status: %s", resp.Status)
@@ -95,6 +95,7 @@ func Test_Routes(t *testing.T) {
 				err = json.Unmarshal(body, &modelList)
 				assert.Nil(t, err)
 
+				assert.NotNil(t, modelList.Models)
 				assert.Equal(t, 0, len(modelList.Models))
 			},
 		},
@@ -124,14 +125,12 @@ 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)
@@ -210,33 +209,31 @@ func Test_Routes(t *testing.T) {
 		},
 	}
 
+	t.Setenv("OLLAMA_MODELS", t.TempDir())
+
 	s := &Server{}
 	router := s.GenerateRoutes()
 
 	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)
-
 	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()
+			resp, err := httpSrv.Client().Do(req)
+			assert.Nil(t, err)
+			defer resp.Body.Close()
 
-		if tc.Expected != nil {
-			tc.Expected(t, resp)
-		}
+			if tc.Expected != nil {
+				tc.Expected(t, resp)
+			}
+		})
 	}
 }

server/sched.go

@@ -6,6 +6,7 @@ import (
 	"fmt"
 	"log/slog"
 	"reflect"
+	"runtime"
 	"sort"
 	"strings"
 	"sync"
@@ -15,7 +16,7 @@ import (
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/gpu"
 	"github.com/ollama/ollama/llm"
-	"github.com/ollama/ollama/server/envconfig"
+	"github.com/ollama/ollama/envconfig"
 	"golang.org/x/exp/slices"
 )
 
@@ -61,6 +62,10 @@ func InitScheduler(ctx context.Context) *Scheduler {
 // context must be canceled to decrement ref count and release the runner
 func (s *Scheduler) GetRunner(c context.Context, model *Model, opts api.Options, sessionDuration time.Duration) (chan *runnerRef, chan error) {
 	// allocate a large enough kv cache for all parallel requests
+	if opts.NumCtx < 4 {
+		opts.NumCtx = 4
+	}
+
 	opts.NumCtx = opts.NumCtx * envconfig.NumParallel
 
 	req := &LlmRequest{
@@ -173,7 +178,7 @@ func (s *Scheduler) processPending(ctx context.Context) {
 				}
 				// Trigger an expiration to unload once it's done
 				runnerToExpire.refMu.Lock()
-				slog.Debug("resetting model to expire immediately to make room", "model", runnerToExpire.model, "refCount", runnerToExpire.refCount)
+				slog.Debug("resetting model to expire immediately to make room", "modelPath", runnerToExpire.modelPath, "refCount", runnerToExpire.refCount)
 				if runnerToExpire.expireTimer != nil {
 					runnerToExpire.expireTimer.Stop()
 					runnerToExpire.expireTimer = nil
@@ -186,13 +191,13 @@ func (s *Scheduler) processPending(ctx context.Context) {
 				// Wait for the unload to happen
 				// Note: at this point we're queueing up all incoming requests, even if they were for
 				// a different model that's loaded and not scheduled to be removed.
-				slog.Debug("waiting for pending requests to complete and unload to occur", "model", runnerToExpire.model)
+				slog.Debug("waiting for pending requests to complete and unload to occur", "modelPath", runnerToExpire.modelPath)
 				select {
 				case <-ctx.Done():
 					slog.Debug("shutting down scheduler pending loop")
 					return
 				case <-s.unloadedCh:
-					slog.Debug("unload completed", "model", runnerToExpire.model)
+					slog.Debug("unload completed", "modelPath", runnerToExpire.modelPath)
 					continue
 				}
 			}
@@ -215,23 +220,23 @@ func (s *Scheduler) processCompleted(ctx context.Context) {
 			runner := s.loaded[finished.model.ModelPath]
 			s.loadedMu.Unlock()
 			if runner == nil {
-				slog.Error("finished requeset signal received after model unloaded", "model", finished.model.ModelPath)
+				slog.Error("finished request signal received after model unloaded", "modelPath", finished.model.ModelPath)
 				continue
 			}
 			runner.refMu.Lock()
 			runner.refCount--
 			if runner.refCount <= 0 {
 				if runner.sessionDuration <= 0 {
-					slog.Debug("runner with zero duration has gone idle, expiring to unload", "model", runner.model)
+					slog.Debug("runner with zero duration has gone idle, expiring to unload", "modelPath", runner.modelPath)
 					if runner.expireTimer != nil {
 						runner.expireTimer.Stop()
 						runner.expireTimer = nil
 					}
 					s.expiredCh <- runner
 				} else if runner.expireTimer == nil {
-					slog.Debug("runner with non-zero duration has gone idle, adding timer", "model", runner.model, "duration", runner.sessionDuration)
+					slog.Debug("runner with non-zero duration has gone idle, adding timer", "modelPath", runner.modelPath, "duration", runner.sessionDuration)
 					runner.expireTimer = time.AfterFunc(runner.sessionDuration, func() {
-						slog.Debug("timer expired, expiring to unload", "model", runner.model)
+						slog.Debug("timer expired, expiring to unload", "modelPath", runner.modelPath)
 						runner.refMu.Lock()
 						defer runner.refMu.Unlock()
 						if runner.expireTimer != nil {
@@ -240,19 +245,21 @@ func (s *Scheduler) processCompleted(ctx context.Context) {
 						}
 						s.expiredCh <- runner
 					})
+					runner.expiresAt = time.Now().Add(runner.sessionDuration)
 				} else {
-					slog.Debug("runner with non-zero duration has gone idle, resetting timer", "model", runner.model, "duration", runner.sessionDuration)
+					slog.Debug("runner with non-zero duration has gone idle, resetting timer", "modelPath", runner.modelPath, "duration", runner.sessionDuration)
 					runner.expireTimer.Reset(runner.sessionDuration)
+					runner.expiresAt = time.Now().Add(runner.sessionDuration)
 				}
 			}
-			slog.Debug("after processing request finished event", "model", runner.model, "refCount", runner.refCount)
+			slog.Debug("after processing request finished event", "modelPath", runner.modelPath, "refCount", runner.refCount)
 			runner.refMu.Unlock()
 		case runner := <-s.expiredCh:
-			slog.Debug("runner expired event received", "model", runner.model)
+			slog.Debug("runner expired event received", "modelPath", runner.modelPath)
 			runner.refMu.Lock()
 			if runner.refCount > 0 {
 				// Shouldn't happen, but safeguard to ensure no leaked runners
-				slog.Debug("expired event with positive ref count, retrying", "model", runner.model, "refCount", runner.refCount)
+				slog.Debug("expired event with positive ref count, retrying", "modelPath", runner.modelPath, "refCount", runner.refCount)
 				go func(runner *runnerRef) {
 					// We can't unload yet, but want to as soon as the current request completes
 					// So queue up another expired event
@@ -264,13 +271,16 @@ func (s *Scheduler) processCompleted(ctx context.Context) {
 			}
 
 			s.loadedMu.Lock()
-			slog.Debug("got lock to unload", "model", runner.model)
+			slog.Debug("got lock to unload", "modelPath", runner.modelPath)
+			finished := runner.waitForVRAMRecovery()
 			runner.unload()
-			delete(s.loaded, runner.model)
+			delete(s.loaded, runner.modelPath)
 			s.loadedMu.Unlock()
-			slog.Debug("runner released", "model", runner.model)
+			slog.Debug("runner released", "modelPath", runner.modelPath)
 			runner.refMu.Unlock()
-			slog.Debug("sending an unloaded event", "model", runner.model)
+
+			<-finished
+			slog.Debug("sending an unloaded event", "modelPath", runner.modelPath)
 			s.unloadedCh <- struct{}{}
 		}
 	}
@@ -309,18 +319,20 @@ func (s *Scheduler) load(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList)
 		req.errCh <- err
 		return
 	}
-	runner := &runnerRef{}
-	runner.model = req.model.ModelPath
-	runner.adapters = req.model.AdapterPaths
-	runner.projectors = req.model.ProjectorPaths
-	runner.llama = llama
-	runner.Options = &req.opts
-	runner.sessionDuration = req.sessionDuration
-	runner.gpus = gpus
-	runner.estimatedVRAM = llama.EstimatedVRAM()
-	runner.loading = true
-	runner.refCount = 1
+	runner := &runnerRef{
+		model:           req.model,
+		modelPath:       req.model.ModelPath,
+		llama:           llama,
+		Options:         &req.opts,
+		sessionDuration: req.sessionDuration,
+		gpus:            gpus,
+		estimatedVRAM:   llama.EstimatedVRAM(),
+		estimatedTotal:  llama.EstimatedTotal(),
+		loading:         true,
+		refCount:        1,
+	}
 	runner.refMu.Lock()
+
 	s.loadedMu.Lock()
 	s.loaded[req.model.ModelPath] = runner
 	slog.Info("loaded runners", "count", len(s.loaded))
@@ -332,7 +344,7 @@ func (s *Scheduler) load(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList)
 			slog.Error("error loading llama server", "error", err)
 			runner.refCount--
 			req.errCh <- err
-			slog.Debug("triggering expiration for failed load", "model", runner.model)
+			slog.Debug("triggering expiration for failed load", "model", runner.modelPath)
 			s.expiredCh <- runner
 			return
 		}
@@ -401,17 +413,18 @@ type runnerRef struct {
 	refCount uint // prevent unloading if > 0
 	// unloading bool      // set to true when we are trying to unload the runner
 
-	llama         llm.LlamaServer
-	loading       bool            // True only during initial load, then false forever
-	gpus          gpu.GpuInfoList // Recorded at time of provisioning
-	estimatedVRAM uint64
+	llama          llm.LlamaServer
+	loading        bool            // True only during initial load, then false forever
+	gpus           gpu.GpuInfoList // Recorded at time of provisioning
+	estimatedVRAM  uint64
+	estimatedTotal uint64
 
 	sessionDuration time.Duration
 	expireTimer     *time.Timer
+	expiresAt       time.Time
 
-	model      string
-	adapters   []string
-	projectors []string
+	model     *Model
+	modelPath string
 	*api.Options
 }
 
@@ -424,9 +437,8 @@ func (runner *runnerRef) unload() {
 	if runner.llama != nil {
 		runner.llama.Close()
 	}
+	runner.model = nil
 	runner.llama = nil
-	runner.adapters = nil
-	runner.projectors = nil
 	runner.Options = nil
 	runner.gpus = nil
 }
@@ -455,8 +467,8 @@ func (runner *runnerRef) needsReload(ctx context.Context, req *LlmRequest) bool
 
 	ctx, cancel := context.WithTimeout(ctx, timeout)
 	defer cancel()
-	if !reflect.DeepEqual(runner.adapters, req.model.AdapterPaths) || // have the adapters changed?
-		!reflect.DeepEqual(runner.projectors, req.model.ProjectorPaths) || // have the projectors changed?
+	if !reflect.DeepEqual(runner.model.AdapterPaths, req.model.AdapterPaths) || // have the adapters changed?
+		!reflect.DeepEqual(runner.model.ProjectorPaths, req.model.ProjectorPaths) || // have the projectors changed?
 		!reflect.DeepEqual(optsExisting, optsNew) || // have the runner options changed?
 		runner.llama.Ping(ctx) != nil {
 		return true
@@ -465,6 +477,61 @@ func (runner *runnerRef) needsReload(ctx context.Context, req *LlmRequest) bool
 	return false
 }
 
+// Free memory reporting on GPUs can lag for a while even after the runner
+// exits, so we have to keep checking until we see the available memory recover,
+// otherwise subsequent model loads will get far less layers loaded or worse
+// case, may completely fall back to CPU mode.
+// This routine must be called before the runner unloads so it can establish
+// a before and after GPU memory allocation.  The returned channel
+// will be notified when we're done waiting, or have timed out and should
+// proceed anyway
+func (runner *runnerRef) waitForVRAMRecovery() chan interface{} {
+	finished := make(chan interface{}, 1)
+
+	// CPU or Metal don't need checking, so no waiting required, windows can page VRAM, and the APIs we query tend to be optimistic on free space
+	if (len(runner.gpus) == 1 && (runner.gpus[0].Library == "cpu" || runner.gpus[0].Library == "metal")) || runtime.GOOS == "windows" {
+		finished <- struct{}{}
+		return finished
+	}
+	start := time.Now()
+
+	// Establish a baseline before we unload
+	gpusBefore := gpu.GetGPUInfo()
+	var totalMemoryBefore, freeMemoryBefore uint64
+	for _, gpu := range gpusBefore {
+		totalMemoryBefore += gpu.TotalMemory
+		freeMemoryBefore += gpu.FreeMemory
+	}
+	go func() {
+		expiresAt := start.Add(5 * time.Second) // typical convergence is 0.5-1.5s
+		ticker := time.NewTicker(250 * time.Millisecond)
+		defer ticker.Stop()
+		for {
+			<-ticker.C
+			if time.Now().After(expiresAt) {
+				slog.Warn("gpu VRAM usage didn't recover within timeout", "seconds", time.Since(start).Seconds())
+				finished <- struct{}{}
+			}
+
+			// Query GPUs, look for free to go back up
+			gpusNow := gpu.GetGPUInfo()
+			var totalMemoryNow, freeMemoryNow uint64
+			for _, gpu := range gpusNow {
+				totalMemoryNow += gpu.TotalMemory
+				freeMemoryNow += gpu.FreeMemory
+			}
+			// If we're within ~80% of the estimated memory usage recovered, bail out
+			if float32(freeMemoryNow-freeMemoryBefore) > float32(runner.estimatedVRAM)*0.8 {
+				slog.Debug(fmt.Sprintf("gpu VRAM free memory converged after %0.2f seconds", time.Since(start).Seconds()))
+				finished <- struct{}{}
+				return
+			}
+		}
+	}()
+	return finished
+
+}
+
 type ByDuration []*runnerRef
 
 func (a ByDuration) Len() int      { return len(a) }
@@ -505,9 +572,9 @@ func pickBestFitGPUs(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList) gpu.
 		// - try subsets of GPUs instead of just falling back to 1 or all in a family
 
 		// Now try all the GPUs
-		if ok, estimatedVRAM = llm.PredictServerFit(gl, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
-			slog.Debug("new model will fit in available VRAM, loading", "model", req.model.ModelPath, "library", gl[0].Library, "required", format.HumanBytes2(estimatedVRAM))
-			return gl
+		if ok, estimatedVRAM = llm.PredictServerFit(sgl, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
+			slog.Debug("new model will fit in available VRAM, loading", "model", req.model.ModelPath, "library", sgl[0].Library, "required", format.HumanBytes2(estimatedVRAM))
+			return sgl
 		}
 	}
 	return nil

server/sched_test.go

@@ -15,7 +15,7 @@ import (
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/gpu"
 	"github.com/ollama/ollama/llm"
-	"github.com/ollama/ollama/server/envconfig"
+	"github.com/ollama/ollama/envconfig"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 )
@@ -151,7 +151,7 @@ func newScenario(t *testing.T, ctx context.Context, modelName string, estimatedV
 }
 
 func TestRequests(t *testing.T) {
-	ctx, done := context.WithTimeout(context.Background(), 500*time.Millisecond)
+	ctx, done := context.WithTimeout(context.Background(), time.Second)
 	defer done()
 
 	// Same model, same request
@@ -164,7 +164,8 @@ func TestRequests(t *testing.T) {
 
 	// simple reload of same model
 	scenario2a := newScenario(t, ctx, "ollama-model-1", 20)
-	scenario2a.req.model = scenario1a.req.model
+	tmpModel := *scenario1a.req.model
+	scenario2a.req.model = &tmpModel
 	scenario2a.ggml = scenario1a.ggml
 
 	// Multiple loaded models
@@ -496,10 +497,9 @@ func TestNeedsReload(t *testing.T) {
 	llm := &mockLlm{}
 	do := api.DefaultOptions()
 	runner := &runnerRef{
-		adapters:   []string{"adapter1"},
-		projectors: []string{"projector1"},
-		Options:    &do,
-		llama:      llm,
+		model:   &Model{AdapterPaths: []string{"adapter1"}, ProjectorPaths: []string{"projector1"}},
+		Options: &do,
+		llama:   llm,
 	}
 	req := &LlmRequest{
 		model: &Model{
@@ -510,10 +510,10 @@ func TestNeedsReload(t *testing.T) {
 	}
 	resp := runner.needsReload(ctx, req)
 	require.True(t, resp)
-	req.model.AdapterPaths = runner.adapters
+	req.model.AdapterPaths = runner.model.AdapterPaths
 	resp = runner.needsReload(ctx, req)
 	require.True(t, resp)
-	req.model.ProjectorPaths = runner.projectors
+	req.model.ProjectorPaths = runner.model.ProjectorPaths
 	runner.loading = true
 	req.opts.NumBatch = 1234
 	resp = runner.needsReload(ctx, req)
@@ -558,11 +558,11 @@ func TestUnloadAllRunners(t *testing.T) {
 func TestUnload(t *testing.T) {
 	llm1 := &mockLlm{}
 	r1 := &runnerRef{llama: llm1}
-	r2 := &runnerRef{adapters: []string{"A"}}
+	r2 := &runnerRef{model: &Model{AdapterPaths: []string{"A"}}}
 	r1.unload()
 	require.True(t, llm1.closeCalled)
 	r2.unload()
-	require.Nil(t, r2.adapters)
+	require.Nil(t, r2.model)
 }
 
 type mockLlm struct {
@@ -578,6 +578,7 @@ type mockLlm struct {
 	closeResp         error
 	closeCalled       bool
 	estimatedVRAM     uint64
+	estimatedTotal    uint64
 }
 
 func (s *mockLlm) Ping(ctx context.Context) error             { return s.pingResp }
@@ -598,4 +599,5 @@ func (s *mockLlm) Close() error {
 	s.closeCalled = true
 	return s.closeResp
 }
-func (s *mockLlm) EstimatedVRAM() uint64 { return s.estimatedVRAM }
+func (s *mockLlm) EstimatedVRAM() uint64  { return s.estimatedVRAM }
+func (s *mockLlm) EstimatedTotal() uint64 { return s.estimatedTotal }

types/errtypes/errtypes.go

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

types/model/name.go

@@ -10,6 +10,7 @@ import (
 	"log/slog"
 	"path/filepath"
 	"strings"
+	"unicode"
 )
 
 // Errors
@@ -35,6 +36,12 @@ func Unqualified(n Name) error {
 // spot in logs.
 const MissingPart = "!MISSING!"
 
+const (
+	defaultHost      = "registry.ollama.ai"
+	defaultNamespace = "library"
+	defaultTag       = "latest"
+)
+
 // DefaultName returns a name with the default values for the host, namespace,
 // and tag parts. The model and digest parts are empty.
 //
@@ -43,9 +50,9 @@ const MissingPart = "!MISSING!"
 //   - The default tag is ("latest")
 func DefaultName() Name {
 	return Name{
-		Host:      "registry.ollama.ai",
-		Namespace: "library",
-		Tag:       "latest",
+		Host:      defaultHost,
+		Namespace: defaultNamespace,
+		Tag:       defaultTag,
 	}
 }
 
@@ -169,6 +176,27 @@ func ParseNameBare(s string) Name {
 	return n
 }
 
+// ParseNameFromFilepath parses a 4-part filepath as a Name. The parts are
+// expected to be in the form:
+//
+// { host } "/" { namespace } "/" { model } "/" { tag }
+func ParseNameFromFilepath(s string) (n Name) {
+	parts := strings.Split(s, string(filepath.Separator))
+	if len(parts) != 4 {
+		return Name{}
+	}
+
+	n.Host = parts[0]
+	n.Namespace = parts[1]
+	n.Model = parts[2]
+	n.Tag = parts[3]
+	if !n.IsFullyQualified() {
+		return Name{}
+	}
+
+	return n
+}
+
 // Merge merges the host, namespace, and tag parts of the two names,
 // preferring the non-empty parts of a.
 func Merge(a, b Name) Name {
@@ -203,6 +231,27 @@ func (n Name) String() string {
 	return b.String()
 }
 
+// DisplayShort returns a short string version of the name.
+func (n Name) DisplayShortest() string {
+	var sb strings.Builder
+
+	if n.Host != defaultHost {
+		sb.WriteString(n.Host)
+		sb.WriteByte('/')
+		sb.WriteString(n.Namespace)
+		sb.WriteByte('/')
+	} else if n.Namespace != defaultNamespace {
+		sb.WriteString(n.Namespace)
+		sb.WriteByte('/')
+	}
+
+	// always include model and tag
+	sb.WriteString(n.Model)
+	sb.WriteString(":")
+	sb.WriteString(n.Tag)
+	return sb.String()
+}
+
 // IsValid reports whether all parts of the name are present and valid. The
 // digest is a special case, and is checked for validity only if present.
 func (n Name) IsValid() bool {
@@ -242,12 +291,12 @@ func (n Name) Filepath() string {
 	if !n.IsFullyQualified() {
 		panic("illegal attempt to get filepath of invalid name")
 	}
-	return strings.ToLower(filepath.Join(
+	return filepath.Join(
 		n.Host,
 		n.Namespace,
 		n.Model,
 		n.Tag,
-	))
+	)
 }
 
 // LogValue returns a slog.Value that represents the name as a string.
@@ -270,14 +319,14 @@ func isValidPart(kind partKind, s string) bool {
 	if !isValidLen(kind, s) {
 		return false
 	}
-	for i := range s {
+	for i, c := range s {
 		if i == 0 {
-			if !isAlphanumericOrUnderscore(s[i]) {
+			if !isAlphanumericOrUnderscore(c) {
 				return false
 			}
 			continue
 		}
-		switch s[i] {
+		switch c {
 		case '_', '-':
 		case '.':
 			if kind == kindNamespace {
@@ -288,7 +337,7 @@ func isValidPart(kind partKind, s string) bool {
 				return false
 			}
 		default:
-			if !isAlphanumericOrUnderscore(s[i]) {
+			if !isAlphanumericOrUnderscore(c) {
 				return false
 			}
 		}
@@ -296,8 +345,8 @@ func isValidPart(kind partKind, s string) bool {
 	return true
 }
 
-func isAlphanumericOrUnderscore(c byte) bool {
-	return c >= 'A' && c <= 'Z' || c >= 'a' && c <= 'z' || c >= '0' && c <= '9' || c == '_'
+func isAlphanumericOrUnderscore(c rune) bool {
+	return unicode.IsLetter(c) || unicode.IsDigit(c) || c == '_'
 }
 
 func cutLast(s, sep string) (before, after string, ok bool) {

types/model/name_test.go

@@ -19,6 +19,16 @@ func TestParseNameParts(t *testing.T) {
 		wantFilepath    string
 		wantValidDigest bool
 	}{
+		{
+			in: "registry.ollama.ai/library/dolphin-mistral:7b-v2.6-dpo-laser-q6_K",
+			want: Name{
+				Host:      "registry.ollama.ai",
+				Namespace: "library",
+				Model:     "dolphin-mistral",
+				Tag:       "7b-v2.6-dpo-laser-q6_K",
+			},
+			wantFilepath: filepath.Join("registry.ollama.ai", "library", "dolphin-mistral", "7b-v2.6-dpo-laser-q6_K"),
+		},
 		{
 			in: "scheme://host:port/namespace/model:tag",
 			want: Name{
@@ -266,9 +276,9 @@ func TestFilepathAllocs(t *testing.T) {
 	allocs := testing.AllocsPerRun(1000, func() {
 		n.Filepath()
 	})
-	allowedAllocs := 2.0
+	var allowedAllocs float64 = 1
 	if runtime.GOOS == "windows" {
-		allowedAllocs = 4
+		allowedAllocs = 3
 	}
 	if allocs > allowedAllocs {
 		t.Errorf("allocs = %v; allowed %v", allocs, allowedAllocs)
@@ -309,6 +319,49 @@ func TestParseDigest(t *testing.T) {
 	}
 }
 
+func TestParseNameFromFilepath(t *testing.T) {
+	cases := map[string]Name{
+		filepath.Join("host", "namespace", "model", "tag"):      {Host: "host", Namespace: "namespace", Model: "model", Tag: "tag"},
+		filepath.Join("host:port", "namespace", "model", "tag"): {Host: "host:port", Namespace: "namespace", Model: "model", Tag: "tag"},
+		filepath.Join("namespace", "model", "tag"):              {},
+		filepath.Join("model", "tag"):                           {},
+		filepath.Join("model"):                                  {},
+		filepath.Join("..", "..", "model", "tag"):               {},
+		filepath.Join("", "namespace", ".", "tag"):              {},
+		filepath.Join(".", ".", ".", "."):                       {},
+		filepath.Join("/", "path", "to", "random", "file"):      {},
+	}
+
+	for in, want := range cases {
+		t.Run(in, func(t *testing.T) {
+			got := ParseNameFromFilepath(in)
+
+			if !reflect.DeepEqual(got, want) {
+				t.Errorf("parseNameFromFilepath(%q) = %v; want %v", in, got, want)
+			}
+		})
+	}
+}
+
+func TestDisplayShortest(t *testing.T) {
+	cases := map[string]string{
+		"registry.ollama.ai/library/model:latest": "model:latest",
+		"registry.ollama.ai/library/model:tag":    "model:tag",
+		"registry.ollama.ai/namespace/model:tag":  "namespace/model:tag",
+		"host/namespace/model:tag":                "host/namespace/model:tag",
+		"host/library/model:tag":                  "host/library/model:tag",
+	}
+
+	for in, want := range cases {
+		t.Run(in, func(t *testing.T) {
+			got := ParseNameBare(in).DisplayShortest()
+			if got != want {
+				t.Errorf("parseName(%q).DisplayShortest() = %q; want %q", in, got, want)
+			}
+		})
+	}
+}
+
 func FuzzName(f *testing.F) {
 	for s := range testCases {
 		f.Add(s)