num parallel embed

README.md

@@ -35,10 +35,10 @@ The official [Ollama Docker image](https://hub.docker.com/r/ollama/ollama) `olla
 
 ## Quickstart
 
-To run and chat with [Llama 3.1](https://ollama.com/library/llama3.1):
+To run and chat with [Llama 3](https://ollama.com/library/llama3):
 
 ```
-ollama run llama3.1
+ollama run llama3
 ```
 
 ## Model library
@@ -49,9 +49,8 @@ Here are some example models that can be downloaded:
 
 | Model              | Parameters | Size  | Download                       |
 | ------------------ | ---------- | ----- | ------------------------------ |
-| Llama 3.1          | 8B         | 4.7GB | `ollama run llama3.1`          |
-| Llama 3.1          | 70B        | 40GB  | `ollama run llama3.1:70b`      |
-| Llama 3.1          | 405B       | 231GB | `ollama run llama3.1:405b`     |
+| Llama 3            | 8B         | 4.7GB | `ollama run llama3`            |
+| Llama 3            | 70B        | 40GB  | `ollama run llama3:70b`        |
 | Phi 3 Mini         | 3.8B       | 2.3GB | `ollama run phi3`              |
 | Phi 3 Medium       | 14B        | 7.9GB | `ollama run phi3:medium`       |
 | Gemma 2            | 9B         | 5.5GB | `ollama run gemma2`            |
@@ -98,16 +97,16 @@ See the [guide](docs/import.md) on importing models for more information.
 
 ### Customize a prompt
 
-Models from the Ollama library can be customized with a prompt. For example, to customize the `llama3.1` model:
+Models from the Ollama library can be customized with a prompt. For example, to customize the `llama3` model:
 
 ```
-ollama pull llama3.1
+ollama pull llama3
 ```
 
 Create a `Modelfile`:
 
 ```
-FROM llama3.1
+FROM llama3
 
 # set the temperature to 1 [higher is more creative, lower is more coherent]
 PARAMETER temperature 1
@@ -142,7 +141,7 @@ ollama create mymodel -f ./Modelfile
 ### Pull a model
 
 ```
-ollama pull llama3.1
+ollama pull llama3
 ```
 
 > This command can also be used to update a local model. Only the diff will be pulled.
@@ -150,13 +149,13 @@ ollama pull llama3.1
 ### Remove a model
 
 ```
-ollama rm llama3.1
+ollama rm llama3
 ```
 
 ### Copy a model
 
 ```
-ollama cp llama3.1 my-model
+ollama cp llama3 my-model
 ```
 
 ### Multiline input
@@ -173,21 +172,21 @@ I'm a basic program that prints the famous "Hello, world!" message to the consol
 ### Multimodal models
 
 ```
-ollama run llava "What's in this image? /Users/jmorgan/Desktop/smile.png"
+>>> What's in this image? /Users/jmorgan/Desktop/smile.png
 The image features a yellow smiley face, which is likely the central focus of the picture.
 ```
 
 ### Pass the prompt as an argument
 
 ```
-$ ollama run llama3.1 "Summarize this file: $(cat README.md)"
+$ ollama run llama3 "Summarize this file: $(cat README.md)"
  Ollama is a lightweight, extensible framework for building and running language models on the local machine. It provides a simple API for creating, running, and managing models, as well as a library of pre-built models that can be easily used in a variety of applications.
 ```
 
 ### Show model information
 
 ```
-ollama show llama3.1
+ollama show llama3
 ```
 
 ### List models on your computer
@@ -215,7 +214,7 @@ Next, start the server:
 Finally, in a separate shell, run a model:
 
 ```
-./ollama run llama3.1
+./ollama run llama3
 ```
 
 ## REST API
@@ -226,7 +225,7 @@ Ollama has a REST API for running and managing models.
 
 ```
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3.1",
+  "model": "llama3",
   "prompt":"Why is the sky blue?"
 }'
 ```
@@ -235,7 +234,7 @@ curl http://localhost:11434/api/generate -d '{
 
 ```
 curl http://localhost:11434/api/chat -d '{
-  "model": "llama3.1",
+  "model": "llama3",
   "messages": [
     { "role": "user", "content": "why is the sky blue?" }
   ]
@@ -390,7 +389,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Llama Coder](https://github.com/ex3ndr/llama-coder) (Copilot alternative using Ollama)
 - [Ollama Copilot](https://github.com/bernardo-bruning/ollama-copilot) (Proxy that allows you to use ollama as a copilot like Github copilot)
 - [twinny](https://github.com/rjmacarthy/twinny) (Copilot and Copilot chat alternative using Ollama)
-- [Wingman-AI](https://github.com/RussellCanfield/wingman-ai) (Copilot code and chat alternative using Ollama and Hugging Face)
+- [Wingman-AI](https://github.com/RussellCanfield/wingman-ai) (Copilot code and chat alternative using Ollama and HuggingFace)
 - [Page Assist](https://github.com/n4ze3m/page-assist) (Chrome Extension)
 - [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)

api/types.go

@@ -114,11 +114,6 @@ func (t Tools) String() string {
 	return string(bts)
 }
 
-func (t Tool) String() string {
-	bts, _ := json.Marshal(t)
-	return string(bts)
-}
-
 // 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.
@@ -214,7 +209,6 @@ type Options struct {
 	NumPredict       int      `json:"num_predict,omitempty"`
 	TopK             int      `json:"top_k,omitempty"`
 	TopP             float32  `json:"top_p,omitempty"`
-	MinP             float32  `json:"min_p,omitempty"`
 	TFSZ             float32  `json:"tfs_z,omitempty"`
 	TypicalP         float32  `json:"typical_p,omitempty"`
 	RepeatLastN      int      `json:"repeat_last_n,omitempty"`
@@ -267,10 +261,6 @@ type EmbedRequest struct {
 type EmbedResponse struct {
 	Model      string      `json:"model"`
 	Embeddings [][]float32 `json:"embeddings"`
-
-	TotalDuration   time.Duration `json:"total_duration,omitempty"`
-	LoadDuration    time.Duration `json:"load_duration,omitempty"`
-	PromptEvalCount int           `json:"prompt_eval_count,omitempty"`
 }
 
 // EmbeddingRequest is the request passed to [Client.Embeddings].

app/ollama.iss

@@ -138,7 +138,7 @@ SetupAppRunningError=Another Ollama installer is running.%n%nPlease cancel or fi
 
 
 ;FinishedHeadingLabel=Run your first model
-;FinishedLabel=%nRun this command in a PowerShell or cmd terminal.%n%n%n    ollama run llama3.1
+;FinishedLabel=%nRun this command in a PowerShell or cmd terminal.%n%n%n    ollama run llama3
 ;ClickFinish=%n
 
 [Registry]

app/ollama_welcome.ps1

@@ -4,5 +4,5 @@ write-host "Welcome to Ollama!"
 write-host ""
 write-host "Run your first model:"
 write-host ""
-write-host "`tollama run llama3.1"
+write-host "`tollama run llama3"
 write-host ""

cmd/cmd.go

@@ -1341,7 +1341,6 @@ func NewCLI() *cobra.Command {
 				envVars["OLLAMA_NUM_PARALLEL"],
 				envVars["OLLAMA_NOPRUNE"],
 				envVars["OLLAMA_ORIGINS"],
-				envVars["OLLAMA_SCHED_SPREAD"],
 				envVars["OLLAMA_TMPDIR"],
 				envVars["OLLAMA_FLASH_ATTENTION"],
 				envVars["OLLAMA_LLM_LIBRARY"],

cmd/interactive.go

@@ -1,7 +1,6 @@
 package cmd
 
 import (
-	"cmp"
 	"errors"
 	"fmt"
 	"io"
@@ -10,14 +9,13 @@ import (
 	"path/filepath"
 	"regexp"
 	"slices"
+	"sort"
 	"strings"
 
 	"github.com/spf13/cobra"
-	"golang.org/x/exp/maps"
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/envconfig"
-	"github.com/ollama/ollama/parser"
 	"github.com/ollama/ollama/progress"
 	"github.com/ollama/ollama/readline"
 	"github.com/ollama/ollama/types/errtypes"
@@ -140,7 +138,6 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 		fmt.Fprintln(os.Stderr, "  /set parameter num_predict <int>      Max number of tokens to predict")
 		fmt.Fprintln(os.Stderr, "  /set parameter top_k <int>            Pick from top k num of tokens")
 		fmt.Fprintln(os.Stderr, "  /set parameter top_p <float>          Pick token based on sum of probabilities")
-		fmt.Fprintln(os.Stderr, "  /set parameter min_p <float>          Pick token based on top token probability * min_p")
 		fmt.Fprintln(os.Stderr, "  /set parameter num_ctx <int>          Set the context size")
 		fmt.Fprintln(os.Stderr, "  /set parameter temperature <float>    Set creativity level")
 		fmt.Fprintln(os.Stderr, "  /set parameter repeat_penalty <float> How strongly to penalize repetitions")
@@ -378,9 +375,9 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 					return err
 				}
 				req := &api.ShowRequest{
-					Name:    opts.Model,
-					System:  opts.System,
-					Options: opts.Options,
+					Name:     opts.Model,
+					System:   opts.System,
+					Options:  opts.Options,
 				}
 				resp, err := client.Show(cmd.Context(), req)
 				if err != nil {
@@ -509,35 +506,31 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 }
 
 func buildModelfile(opts runOptions) string {
-	var f parser.File
-	f.Commands = append(f.Commands, parser.Command{Name: "model", Args: cmp.Or(opts.ParentModel, opts.Model)})
-
+	var mf strings.Builder
+	model := opts.ParentModel
+	if model == "" {
+		model = opts.Model
+	}
+	fmt.Fprintf(&mf, "FROM %s\n", model)
 	if opts.System != "" {
-		f.Commands = append(f.Commands, parser.Command{Name: "system", Args: opts.System})
+		fmt.Fprintf(&mf, "SYSTEM \"\"\"%s\"\"\"\n", opts.System)
 	}
 
-	keys := maps.Keys(opts.Options)
-	slices.Sort(keys)
+	keys := make([]string, 0)
+	for k := range opts.Options {
+		keys = append(keys, k)
+	}
+	sort.Strings(keys)
 	for _, k := range keys {
-		v := opts.Options[k]
-		var cmds []parser.Command
-		switch t := v.(type) {
-		case []string:
-			for _, s := range t {
-				cmds = append(cmds, parser.Command{Name: k, Args: s})
-			}
-		default:
-			cmds = append(cmds, parser.Command{Name: k, Args: fmt.Sprintf("%v", t)})
-		}
-
-		f.Commands = append(f.Commands, cmds...)
+		fmt.Fprintf(&mf, "PARAMETER %s %v\n", k, opts.Options[k])
 	}
+	fmt.Fprintln(&mf)
 
 	for _, msg := range opts.Messages {
-		f.Commands = append(f.Commands, parser.Command{Name: "message", Args: fmt.Sprintf("%s: %s", msg.Role, msg.Content)})
+		fmt.Fprintf(&mf, "MESSAGE %s \"\"\"%s\"\"\"\n", msg.Role, msg.Content)
 	}
 
-	return f.String()
+	return mf.String()
 }
 
 func normalizeFilePath(fp string) string {

cmd/interactive_test.go

@@ -1,10 +1,12 @@
 package cmd
 
 import (
+	"bytes"
 	"testing"
+	"text/template"
 
-	"github.com/google/go-cmp/cmp"
 	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
 
 	"github.com/ollama/ollama/api"
 )
@@ -55,53 +57,58 @@ d:\path with\spaces\seven.svg inbetween7 c:\users\jdoe\eight.png inbetween8
 
 func TestModelfileBuilder(t *testing.T) {
 	opts := runOptions{
-		Model:  "hork",
-		System: "You are part horse and part shark, but all hork. Do horklike things",
+		Model:    "hork",
+		System:   "You are part horse and part shark, but all hork. Do horklike things",
 		Messages: []api.Message{
 			{Role: "user", Content: "Hey there hork!"},
 			{Role: "assistant", Content: "Yes it is true, I am half horse, half shark."},
 		},
-		Options: map[string]any{
-			"temperature":      0.9,
-			"seed":             42,
-			"penalize_newline": false,
-			"stop":             []string{"hi", "there"},
-		},
+		Options: map[string]interface{}{},
 	}
 
-	t.Run("model", func(t *testing.T) {
-		expect := `FROM hork
-SYSTEM You are part horse and part shark, but all hork. Do horklike things
+	opts.Options["temperature"] = 0.9
+	opts.Options["seed"] = 42
+	opts.Options["penalize_newline"] = false
+	opts.Options["stop"] = []string{"hi", "there"}
+
+	mf := buildModelfile(opts)
+	expectedModelfile := `FROM {{.Model}}
+SYSTEM """{{.System}}"""
 PARAMETER penalize_newline false
 PARAMETER seed 42
-PARAMETER stop hi
-PARAMETER stop there
+PARAMETER stop [hi there]
 PARAMETER temperature 0.9
-MESSAGE user Hey there hork!
-MESSAGE assistant Yes it is true, I am half horse, half shark.
+
+MESSAGE user """Hey there hork!"""
+MESSAGE assistant """Yes it is true, I am half horse, half shark."""
 `
 
-		actual := buildModelfile(opts)
-		if diff := cmp.Diff(expect, actual); diff != "" {
-			t.Errorf("mismatch (-want +got):\n%s", diff)
-		}
-	})
+	tmpl, err := template.New("").Parse(expectedModelfile)
+	require.NoError(t, err)
 
-	t.Run("parent model", func(t *testing.T) {
-		opts.ParentModel = "horseshark"
-		expect := `FROM horseshark
-SYSTEM You are part horse and part shark, but all hork. Do horklike things
+	var buf bytes.Buffer
+	err = tmpl.Execute(&buf, opts)
+	require.NoError(t, err)
+	assert.Equal(t, buf.String(), mf)
+
+	opts.ParentModel = "horseshark"
+	mf = buildModelfile(opts)
+	expectedModelfile = `FROM {{.ParentModel}}
+SYSTEM """{{.System}}"""
 PARAMETER penalize_newline false
 PARAMETER seed 42
-PARAMETER stop hi
-PARAMETER stop there
+PARAMETER stop [hi there]
 PARAMETER temperature 0.9
-MESSAGE user Hey there hork!
-MESSAGE assistant Yes it is true, I am half horse, half shark.
+
+MESSAGE user """Hey there hork!"""
+MESSAGE assistant """Yes it is true, I am half horse, half shark."""
 `
-		actual := buildModelfile(opts)
-		if diff := cmp.Diff(expect, actual); diff != "" {
-			t.Errorf("mismatch (-want +got):\n%s", diff)
-		}
-	})
+
+	tmpl, err = template.New("").Parse(expectedModelfile)
+	require.NoError(t, err)
+
+	var parentBuf bytes.Buffer
+	err = tmpl.Execute(&parentBuf, opts)
+	require.NoError(t, err)
+	assert.Equal(t, parentBuf.String(), mf)
 }

docs/api.md

@@ -336,7 +336,6 @@ curl http://localhost:11434/api/generate -d '{
     "num_predict": 100,
     "top_k": 20,
     "top_p": 0.9,
-    "min_p": 0.0,
     "tfs_z": 0.5,
     "typical_p": 0.7,
     "repeat_last_n": 33,
@@ -587,7 +586,7 @@ Final response:
 
 ##### Request
 
-Send a chat message with images. The images should be provided as an array, with the individual images encoded in Base64.
+Send a chat message with a conversation history.
 
 ```shell
 curl http://localhost:11434/api/chat -d '{

docs/docker.md

@@ -63,7 +63,7 @@ docker run -d --device /dev/kfd --device /dev/dri -v ollama:/root/.ollama -p 114
 Now you can run a model:
 
 ```
-docker exec -it ollama ollama run llama3.1
+docker exec -it ollama ollama run llama3
 ```
 
 ### Try different models

docs/faq.md

@@ -227,7 +227,7 @@ curl http://localhost:11434/api/chat -d '{"model": "mistral"}'
 
 To preload a model using the CLI, use the command:
 ```shell
-ollama run llama3.1 ""
+ollama run llama3 ""
 ```
 
 ## How do I keep a model loaded in memory or make it unload immediately?
@@ -272,8 +272,4 @@ The following server settings may be used to adjust how Ollama handles concurren
 - `OLLAMA_NUM_PARALLEL` - The maximum number of parallel requests each model will process at the same time.  The default will auto-select either 4 or 1 based on available memory.
 - `OLLAMA_MAX_QUEUE` - The maximum number of requests Ollama will queue when busy before rejecting additional requests. The default is 512
 
-Note: Windows with Radeon GPUs currently default to 1 model maximum due to limitations in ROCm v5.7 for available VRAM reporting.  Once ROCm v6.2 is available, Windows Radeon will follow the defaults above.  You may enable concurrent model loads on Radeon on Windows, but ensure you don't load more models than will fit into your GPUs VRAM.
-
-## How does Ollama load models on multiple GPUs?
-
-Installing multiple GPUs of the same brand can be a great way to increase your available VRAM to load larger models.  When you load a new model, Ollama evaluates the required VRAM for the model against what is currently available.  If the model will entirely fit on any single GPU, Ollama will load the model on that GPU.  This typically provides the best performance as it reduces the amount of data transfering across the PCI bus during inference.  If the model does not fit entirely on one GPU, then it will be spread across all the available GPUs.
+Note: Windows with Radeon GPUs currently default to 1 model maximum due to limitations in ROCm v5.7 for available VRAM reporting.  Once ROCm v6.2 is available, Windows Radeon will follow the defaults above.  You may enable concurrent model loads on Radeon on Windows, but ensure you don't load more models than will fit into your GPUs VRAM.

docs/modelfile.md

@@ -141,7 +141,6 @@ PARAMETER <parameter> <parametervalue>
 | num_predict    | Maximum number of tokens to predict when generating text. (Default: 128, -1 = infinite generation, -2 = fill context)                                                                                                                                   | int        | num_predict 42       |
 | top_k          | Reduces the probability of generating nonsense. A higher value (e.g. 100) will give more diverse answers, while a lower value (e.g. 10) will be more conservative. (Default: 40)                                                                        | int        | top_k 40             |
 | top_p          | Works together with top-k. A higher value (e.g., 0.95) will lead to more diverse text, while a lower value (e.g., 0.5) will generate more focused and conservative text. (Default: 0.9)                                                                 | float      | top_p 0.9            |
-| min_p          | Alternative to the top_p, and aims to ensure a balance of quality and variety. The parameter *p* represents the minimum probability for a token to be considered, relative to the probability of the most likely token. For example, with *p*=0.05 and the most likely token having a probability of 0.9, logits with a value less than 0.045 are filtered out. (Default: 0.0) | float      | min_p 0.05            |
 
 ### TEMPLATE
 

docs/tutorials/langchainjs.md

@@ -15,7 +15,7 @@ import { Ollama } from "@langchain/community/llms/ollama";
 
 const ollama = new Ollama({
   baseUrl: "http://localhost:11434",
-  model: "llama3.1",
+  model: "llama3",
 });
 
 const answer = await ollama.invoke(`why is the sky blue?`);
@@ -23,7 +23,7 @@ const answer = await ollama.invoke(`why is the sky blue?`);
 console.log(answer);
 ```
 
-That will get us the same thing as if we ran `ollama run llama3.1 "why is the sky blue"` in the terminal. But we want to load a document from the web to ask a question against. **Cheerio** is a great library for ingesting a webpage, and **LangChain** uses it in their **CheerioWebBaseLoader**. So let's install **Cheerio** and build that part of the app.
+That will get us the same thing as if we ran `ollama run llama3 "why is the sky blue"` in the terminal. But we want to load a document from the web to ask a question against. **Cheerio** is a great library for ingesting a webpage, and **LangChain** uses it in their **CheerioWebBaseLoader**. So let's install **Cheerio** and build that part of the app.
 
 ```bash
 npm install cheerio

docs/windows.md

@@ -23,8 +23,6 @@ Logs will often be helpful in diagnosing the problem (see
 * NVIDIA 452.39 or newer Drivers if you have an NVIDIA card
 * AMD Radeon Driver https://www.amd.com/en/support if you have a Radeon card
 
-Ollama uses unicode characters for progress indication, which may render as unknown squares in some older terminal fonts in Windows 10. If you see this, try changing your terminal font settings.
-
 ## API Access
 
 Here's a quick example showing API access from `powershell`

examples/go-chat/main.go

@@ -35,7 +35,7 @@ func main() {
 
 	ctx := context.Background()
 	req := &api.ChatRequest{
-		Model:    "llama3.1",
+		Model:    "llama3",
 		Messages: messages,
 	}
 

examples/go-generate-streaming/main.go

@@ -16,7 +16,7 @@ func main() {
 
 	// By default, GenerateRequest is streaming.
 	req := &api.GenerateRequest{
-		Model:  "gemma2",
+		Model:  "gemma",
 		Prompt: "how many planets are there?",
 	}
 

examples/go-generate/main.go

@@ -15,7 +15,7 @@ func main() {
 	}
 
 	req := &api.GenerateRequest{
-		Model:  "gemma2",
+		Model:  "gemma",
 		Prompt: "how many planets are there?",
 
 		// set streaming to false

examples/langchain-python-rag-document/README.md

@@ -4,14 +4,6 @@ This example provides an interface for asking questions to a PDF document.
 
 ## Setup
 
-1. Ensure you have the `llama3.1` model installed:
-
-```
-ollama pull llama3.1
-```
-
-2. Install the Python Requirements.
-
 ```
 pip install -r requirements.txt
 ```

examples/langchain-python-rag-document/main.py

@@ -51,7 +51,7 @@ while True:
         template=template,
     )
 
-    llm = Ollama(model="llama3.1", callback_manager=CallbackManager([StreamingStdOutCallbackHandler()]))
+    llm = Ollama(model="llama3:8b", callback_manager=CallbackManager([StreamingStdOutCallbackHandler()]))
     qa_chain = RetrievalQA.from_chain_type(
         llm,
         retriever=vectorstore.as_retriever(),

examples/langchain-python-rag-websummary/README.md

@@ -4,10 +4,10 @@ This example summarizes the website, [https://ollama.com/blog/run-llama2-uncenso
 
 ## Running the Example
 
-1. Ensure you have the `llama3.1` model installed:
+1. Ensure you have the `llama2` model installed:
 
    ```bash
-   ollama pull llama3.1
+   ollama pull llama2
    ```
 
 2. Install the Python Requirements.

examples/langchain-python-rag-websummary/main.py

@@ -5,8 +5,8 @@ from langchain.chains.summarize import load_summarize_chain
 loader = WebBaseLoader("https://ollama.com/blog/run-llama2-uncensored-locally")
 docs = loader.load()
 
-llm = Ollama(model="llama3.1")
+llm = Ollama(model="llama3")
 chain = load_summarize_chain(llm, chain_type="stuff")
 
-result = chain.invoke(docs)
+result = chain.invoke(docs) 
 print(result)

examples/langchain-python-simple/README.md

@@ -4,10 +4,10 @@ This example is a basic "hello world" of using LangChain with Ollama.
 
 ## Running the Example
 
-1. Ensure you have the `llama3.1` model installed:
+1. Ensure you have the `llama3` model installed:
 
    ```bash
-   ollama pull llama3.1
+   ollama pull llama3
    ```
 
 2. Install the Python Requirements.

examples/langchain-python-simple/main.py

@@ -1,6 +1,6 @@
 from langchain.llms import Ollama
 
 input = input("What is your question?")
-llm = Ollama(model="llama3.1")
+llm = Ollama(model="llama3")
 res = llm.predict(input)
 print (res)

examples/modelfile-mario/Modelfile

@@ -1,4 +1,4 @@
-FROM llama3.1
+FROM llama3
 PARAMETER temperature 1
 SYSTEM """
 You are Mario from super mario bros, acting as an assistant.

examples/modelfile-mario/readme.md

@@ -2,12 +2,12 @@
 
 # Example character: Mario
 
-This example shows how to create a basic character using Llama3.1 as the base model.
+This example shows how to create a basic character using Llama3 as the base model.
 
 To run this example:
 
 1. Download the Modelfile
-2. `ollama pull llama3.1` to get the base model used in the model file.
+2. `ollama pull llama3` to get the base model used in the model file.
 3. `ollama create NAME -f ./Modelfile`
 4. `ollama run NAME`
 
@@ -18,7 +18,7 @@ Ask it some questions like "Who are you?" or "Is Peach in trouble again?"
 What the model file looks like:
 
 ```
-FROM llama3.1
+FROM llama3
 PARAMETER temperature 1
 SYSTEM """
 You are Mario from Super Mario Bros, acting as an assistant.

examples/python-dockerit/dockerit.py

@@ -4,7 +4,7 @@ imageName = input("Enter the name of the image: ")
 client = docker.from_env()
 s = requests.Session()
 output=""
-with s.post('http://localhost:11434/api/generate', json={'model': 'mattw/dockerit', 'prompt': inputDescription}, stream=True) as r:
+with s.post('http://localhost:11434/api/generate', json={'model': 'dockerit', 'prompt': inputDescription}, stream=True) as r:
   for line in r.iter_lines():
     if line:
       j = json.loads(line)

examples/python-json-datagenerator/predefinedschema.py

@@ -2,7 +2,7 @@ import requests
 import json
 import random
 
-model = "llama3.1"
+model = "llama3"
 template = {
   "firstName": "",
   "lastName": "",

examples/python-json-datagenerator/randomaddresses.py

@@ -12,7 +12,7 @@ countries = [
     "France",
 ]
 country = random.choice(countries)
-model = "llama3.1"
+model = "llama3"
 
 prompt = f"generate one realistically believable sample data set of a persons first name, last name, address in {country}, and phone number. Do not use common names. Respond using JSON. Key names should have no backslashes, values should use plain ascii with no special characters."
 

examples/python-json-datagenerator/readme.md

@@ -6,10 +6,10 @@ There are two python scripts in this example. `randomaddresses.py` generates ran
 
 ## Running the Example
 
-1. Ensure you have the `llama3.1` model installed:
+1. Ensure you have the `llama3` model installed:
 
    ```bash
-   ollama pull llama3.1
+   ollama pull llama3
    ```
 
 2. Install the Python Requirements.

examples/python-simplechat/client.py

@@ -2,7 +2,7 @@ import json
 import requests
 
 # NOTE: ollama must be running for this to work, start the ollama app or run `ollama serve`
-model = "llama3.1"  # TODO: update this for whatever model you wish to use
+model = "llama3"  # TODO: update this for whatever model you wish to use
 
 
 def chat(messages):

examples/python-simplechat/readme.md

@@ -4,10 +4,10 @@ The **chat** endpoint is one of two ways to generate text from an LLM with Ollam
 
 ## Running the Example
 
-1. Ensure you have the `llama3.1` model installed:
+1. Ensure you have the `llama3` model installed:
 
    ```bash
-   ollama pull llama3.1
+   ollama pull llama3
    ```
 
 2. Install the Python Requirements.

examples/typescript-simplechat/client.ts

@@ -1,6 +1,6 @@
 import * as readline from "readline";
 
-const model = "llama3.1";
+const model = "llama3";
 type Message = {
   role: "assistant" | "user" | "system";
   content: string;

gpu/amd_linux.go

@@ -10,7 +10,6 @@ import (
 	"path/filepath"
 	"regexp"
 	"slices"
-	"sort"
 	"strconv"
 	"strings"
 
@@ -83,20 +82,6 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 	// The amdgpu driver always exposes the host CPU(s) first, but we have to skip them and subtract
 	// from the other IDs to get alignment with the HIP libraries expectations (zero is the first GPU, not the CPU)
 	matches, _ := filepath.Glob(GPUPropertiesFileGlob)
-	sort.Slice(matches, func(i, j int) bool {
-		// /sys/class/kfd/kfd/topology/nodes/<number>/properties
-		a, err := strconv.ParseInt(filepath.Base(filepath.Dir(matches[i])), 10, 64)
-		if err != nil {
-			slog.Debug("parse err", "error", err, "match", matches[i])
-			return false
-		}
-		b, err := strconv.ParseInt(filepath.Base(filepath.Dir(matches[j])), 10, 64)
-		if err != nil {
-			slog.Debug("parse err", "error", err, "match", matches[i])
-			return false
-		}
-		return a < b
-	})
 	cpuCount := 0
 	for _, match := range matches {
 		slog.Debug("evaluating amdgpu node " + match)

integration/embed_test.go

@@ -69,10 +69,6 @@ func TestAllMiniLMEmbed(t *testing.T) {
 	if !floatsEqual32(res.Embeddings[0][0], 0.010071031) {
 		t.Fatalf("expected 0.010071031, got %.8f", res.Embeddings[0][0])
 	}
-
-	if res.PromptEvalCount != 8 {
-		t.Fatalf("expected 8 prompt tokens, got %d", res.PromptEvalCount)
-	}
 }
 
 func TestAllMiniLMBatchEmbed(t *testing.T) {
@@ -101,10 +97,6 @@ func TestAllMiniLMBatchEmbed(t *testing.T) {
 	if !floatsEqual32(res.Embeddings[0][0], 0.010071031) || !floatsEqual32(res.Embeddings[1][0], -0.009802706) {
 		t.Fatalf("expected 0.010071031 and -0.009802706, got %.8f and %.8f", res.Embeddings[0][0], res.Embeddings[1][0])
 	}
-
-	if res.PromptEvalCount != 16 {
-		t.Fatalf("expected 16 prompt tokens, got %d", res.PromptEvalCount)
-	}
 }
 
 func TestAllMiniLMEmbedTruncate(t *testing.T) {

llm/ext_server/server.cpp

@@ -41,7 +41,6 @@
 
 #if defined(_WIN32)
 #include <windows.h>
-#include <errhandlingapi.h>
 #endif
 
 #include <cstddef>
@@ -1221,7 +1220,6 @@ struct llama_server_context
                 res.result_json = json
                 {
                     {"embedding", std::vector<float>(embd, embd + n_embd)},
-                    {"timings",             slot.get_formated_timings()},
                 };
             }
         }
@@ -2439,6 +2437,15 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
             params.lora_adapter.emplace_back(lora_adapter, std::stof(argv[i]));
             params.use_mmap = false;
         }
+        else if (arg == "--lora-base")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            params.lora_base = argv[i];
+        }
         else if (arg == "-v" || arg == "--verbose")
         {
             server_verbose = true;
@@ -2730,9 +2737,6 @@ int wmain(int argc, wchar_t **wargv) {
     for (int i = 0; i < argc; ++i) {
         argv[i] = wchar_to_char(wargv[i]);
     }
-
-    // Adjust error mode to avoid error dialog after we start.
-    SetErrorMode(SEM_FAILCRITICALERRORS);
 #else
 int main(int argc, char **argv) {
 #endif
@@ -3204,15 +3208,11 @@ int main(int argc, char **argv) {
 
                     responses = result.result_json.value("results", std::vector<json>{result.result_json});
                     json embeddings = json::array();
-
-                    int prompt_n = 0;
                     for (auto & elem : responses) {
                         embeddings.push_back(elem.at("embedding"));
-                        prompt_n += elem.at("timings").at("prompt_n").get<int>();
                     }
-
                     // send the result
-                    json embedding_res = json{{"embedding", embeddings}, {"prompt_n", prompt_n}};
+                    json embedding_res = json{{"embedding", embeddings}};
                     return res.set_content(embedding_res.dump(), "application/json; charset=utf-8");
                 }
             });

llm/llm_darwin_amd64.go

@@ -2,10 +2,7 @@ package llm
 
 import (
 	"embed"
-	"syscall"
 )
 
 //go:embed build/darwin/x86_64/*/bin/*
 var libEmbed embed.FS
-
-var LlamaServerSysProcAttr = &syscall.SysProcAttr{}

llm/llm_darwin_arm64.go

@@ -2,10 +2,7 @@ package llm
 
 import (
 	"embed"
-	"syscall"
 )
 
 //go:embed build/darwin/arm64/*/bin/*
 var libEmbed embed.FS
-
-var LlamaServerSysProcAttr = &syscall.SysProcAttr{}

llm/llm_linux.go

@@ -1,11 +1,6 @@
 package llm
 
-import (
-	"embed"
-	"syscall"
-)
+import "embed"
 
 //go:embed build/linux/*/*/bin/*
 var libEmbed embed.FS
-
-var LlamaServerSysProcAttr = &syscall.SysProcAttr{}

llm/llm_windows.go

@@ -1,20 +1,6 @@
 package llm
 
-import (
-	"embed"
-	"syscall"
-)
+import "embed"
 
 // unused on windows
 var libEmbed embed.FS
-
-const CREATE_DEFAULT_ERROR_MODE = 0x04000000
-
-var LlamaServerSysProcAttr = &syscall.SysProcAttr{
-	// Wire up the default error handling logic If for some reason a DLL is
-	// missing in the path this will pop up a GUI Dialog explaining the fault so
-	// the user can either fix their PATH, or report a bug. Without this
-	// setting, the process exits immediately with a generic exit status but no
-	// way to (easily) figure out what the actual missing DLL was.
-	CreationFlags: CREATE_DEFAULT_ERROR_MODE,
-}

llm/patches/05-default-pretokenizer.diff

@@ -1,8 +1,8 @@
 diff --git a/src/llama.cpp b/src/llama.cpp
-index a207451f..2ddf431d 100644
+index 8fe51971..7113ba64 100644
 --- a/src/llama.cpp
 +++ b/src/llama.cpp
-@@ -5347,16 +5347,7 @@ static void llm_load_vocab(
+@@ -5433,16 +5433,7 @@ static void llm_load_vocab(
          if (vocab.type == LLAMA_VOCAB_TYPE_BPE) {
              vocab.tokenizer_add_space_prefix = false;
              vocab.tokenizer_clean_spaces = true;
@@ -20,9 +20,9 @@ index a207451f..2ddf431d 100644
                  vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
              } else if (
                      tokenizer_pre == "llama3"   ||
-@@ -5443,7 +5434,8 @@ static void llm_load_vocab(
-                 tokenizer_pre == "codeshell") {
-                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_CODESHELL;
+@@ -5526,7 +5517,8 @@ static void llm_load_vocab(
+                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_SMOLLM;
+                 vocab.tokenizer_clean_spaces = false;
              } else {
 -                throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
 +                LLAMA_LOG_WARN("%s: missing or unrecognized pre-tokenizer type, using: 'default'\n", __func__);

llm/patches/09-lora.diff

@@ -2,7 +2,7 @@ diff --git a/common/common.cpp b/common/common.cpp
 index dbb724fb..c26fe6ee 100644
 --- a/common/common.cpp
 +++ b/common/common.cpp
-@@ -2087,14 +2087,27 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
+@@ -2087,14 +2087,29 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
      for (unsigned int i = 0; i < params.lora_adapter.size(); ++i) {
          const std::string & lora_adapter = std::get<0>(params.lora_adapter[i]);
          float lora_scale = std::get<1>(params.lora_adapter[i]);
@@ -20,7 +20,9 @@ index dbb724fb..c26fe6ee 100644
 +            int err = llama_model_apply_lora_from_file(model,
 +                                                    lora_adapter.c_str(),
 +                                                    lora_scale,
-+                                                    nullptr,
++                                                    ((i > 0) || params.lora_base.empty())
++                                                        ? NULL
++                                                        : params.lora_base.c_str(),
 +                                                    params.n_threads);
 +            if (err != 0) {
 +                fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__);

llm/patches/10-params.diff

@@ -1,20 +0,0 @@
-diff --git a/src/llama.cpp b/src/llama.cpp
-index a207451f..fba6b175 100644
---- a/src/llama.cpp
-+++ b/src/llama.cpp
-@@ -4969,6 +4969,7 @@ static void llm_load_hparams(
-                 hparams.attn_soft_cap = true;
- 
-                 switch (hparams.n_layer) {
-+                    case 26: model.type = e_model::MODEL_2B; break;
-                     case 42: model.type = e_model::MODEL_9B; break;
-                     case 46: model.type = e_model::MODEL_27B; break;
-                     default: model.type = e_model::MODEL_UNKNOWN;
-@@ -11736,6 +11737,7 @@ struct llm_build_context {
- 
-                 // ref: https://github.com/google/gemma_pytorch/commit/03e657582d17cb5a8617ebf333c1c16f3694670e
-                 switch (model.type) {
-+                    case e_model::MODEL_2B: Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k))); break;
-                     case e_model::MODEL_9B:  Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k)));   break;
-                     case e_model::MODEL_27B: Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd / n_head))); break;
-                     default: GGML_ABORT("fatal error");

llm/server.go

@@ -33,7 +33,7 @@ type LlamaServer interface {
 	Ping(ctx context.Context) error
 	WaitUntilRunning(ctx context.Context) error
 	Completion(ctx context.Context, req CompletionRequest, fn func(CompletionResponse)) error
-	Embed(ctx context.Context, input []string) (*EmbedResponse, error)
+	Embed(ctx context.Context, input []string) ([][]float32, error)
 	Tokenize(ctx context.Context, content string) ([]int, error)
 	Detokenize(ctx context.Context, tokens []int) (string, error)
 	Close() error
@@ -346,7 +346,6 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		s.cmd.Env = os.Environ()
 		s.cmd.Stdout = os.Stdout
 		s.cmd.Stderr = s.status
-		s.cmd.SysProcAttr = LlamaServerSysProcAttr
 
 		envWorkarounds := [][2]string{}
 		for _, gpu := range gpus {
@@ -727,7 +726,6 @@ func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn fu
 		"temperature":       req.Options.Temperature,
 		"top_k":             req.Options.TopK,
 		"top_p":             req.Options.TopP,
-		"min_p":             req.Options.MinP,
 		"tfs_z":             req.Options.TFSZ,
 		"typical_p":         req.Options.TypicalP,
 		"repeat_last_n":     req.Options.RepeatLastN,
@@ -879,11 +877,10 @@ type EmbedRequest struct {
 }
 
 type EmbedResponse struct {
-	Embedding       [][]float32 `json:"embedding"`
-	PromptEvalCount int         `json:"prompt_n"`
+	Embedding [][]float32 `json:"embedding"`
 }
 
-func (s *llmServer) Embed(ctx context.Context, input []string) (*EmbedResponse, error) {
+func (s *llmServer) Embed(ctx context.Context, input []string) ([][]float32, error) {
 	if err := s.sem.Acquire(ctx, 1); err != nil {
 		slog.Error("Failed to acquire semaphore", "error", err)
 		return nil, err
@@ -925,12 +922,12 @@ func (s *llmServer) Embed(ctx context.Context, input []string) (*EmbedResponse,
 		return nil, fmt.Errorf("%s", body)
 	}
 
-	var e EmbedResponse
-	if err := json.Unmarshal(body, &e); err != nil {
+	var embedding EmbedResponse
+	if err := json.Unmarshal(body, &embedding); err != nil {
 		return nil, fmt.Errorf("unmarshal tokenize response: %w", err)
 	}
 
-	return &e, nil
+	return embedding.Embedding, nil
 }
 
 type TokenizeRequest struct {

macapp/src/app.tsx

@@ -19,7 +19,7 @@ export default function () {
   const [step, setStep] = useState<Step>(Step.WELCOME)
   const [commandCopied, setCommandCopied] = useState<boolean>(false)
 
-  const command = 'ollama run llama3.1'
+  const command = 'ollama run llama3'
 
   return (
     <div className='drag'>

openai/openai.go

@@ -218,9 +218,6 @@ func toChatCompletion(id string, r api.ChatResponse) ChatCompletion {
 			Index:   0,
 			Message: Message{Role: r.Message.Role, Content: r.Message.Content, ToolCalls: toolCalls},
 			FinishReason: func(reason string) *string {
-				if len(toolCalls) > 0 {
-					reason = "tool_calls"
-				}
 				if len(reason) > 0 {
 					return &reason
 				}

parser/parser_test.go

@@ -451,7 +451,6 @@ func TestParseFileParameters(t *testing.T) {
 		"num_predict 1":                {"num_predict", "1"},
 		"top_k 1":                      {"top_k", "1"},
 		"top_p 1.0":                    {"top_p", "1.0"},
-		"min_p 0.05":                   {"min_p", "0.05"},
 		"tfs_z 1.0":                    {"tfs_z", "1.0"},
 		"typical_p 1.0":                {"typical_p", "1.0"},
 		"repeat_last_n 1":              {"repeat_last_n", "1"},

scripts/install.sh

@@ -198,29 +198,19 @@ if check_gpu lspci amdgpu || check_gpu lshw amdgpu; then
     exit 0
 fi
 
-CUDA_REPO_ERR_MSG="NVIDIA GPU detected, but your OS and Architecture are not supported by NVIDIA.  Please install the CUDA driver manually https://docs.nvidia.com/cuda/cuda-installation-guide-linux/"
 # ref: https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#rhel-7-centos-7
 # ref: https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#rhel-8-rocky-8
 # ref: https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#rhel-9-rocky-9
 # ref: https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#fedora
 install_cuda_driver_yum() {
     status 'Installing NVIDIA repository...'
-    
     case $PACKAGE_MANAGER in
         yum)
             $SUDO $PACKAGE_MANAGER -y install yum-utils
-            if curl -I --silent --fail --location "https://developer.download.nvidia.com/compute/cuda/repos/$1$2/$(uname -m)/cuda-$1$2.repo" >/dev/null ; then
-                $SUDO $PACKAGE_MANAGER-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/$1$2/$(uname -m)/cuda-$1$2.repo
-            else
-                error $CUDA_REPO_ERR_MSG
-            fi
+            $SUDO $PACKAGE_MANAGER-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/$1$2/$(uname -m)/cuda-$1$2.repo
             ;;
         dnf)
-            if curl -I --silent --fail --location "https://developer.download.nvidia.com/compute/cuda/repos/$1$2/$(uname -m)/cuda-$1$2.repo" >/dev/null ; then
-                $SUDO $PACKAGE_MANAGER config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/$1$2/$(uname -m)/cuda-$1$2.repo
-            else
-                error $CUDA_REPO_ERR_MSG
-            fi
+            $SUDO $PACKAGE_MANAGER config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/$1$2/$(uname -m)/cuda-$1$2.repo
             ;;
     esac
 
@@ -245,11 +235,7 @@ install_cuda_driver_yum() {
 # ref: https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#debian
 install_cuda_driver_apt() {
     status 'Installing NVIDIA repository...'
-    if curl -I --silent --fail --location "https://developer.download.nvidia.com/compute/cuda/repos/$1$2/$(uname -m)/cuda-keyring_1.1-1_all.deb" >/dev/null ; then
-        curl -fsSL -o $TEMP_DIR/cuda-keyring.deb https://developer.download.nvidia.com/compute/cuda/repos/$1$2/$(uname -m)/cuda-keyring_1.1-1_all.deb
-    else
-        error $CUDA_REPO_ERR_MSG
-    fi
+    curl -fsSL -o $TEMP_DIR/cuda-keyring.deb https://developer.download.nvidia.com/compute/cuda/repos/$1$2/$(uname -m)/cuda-keyring_1.1-1_all.deb
 
     case $1 in
         debian)

server/auth.go

@@ -67,7 +67,7 @@ func getAuthorizationToken(ctx context.Context, challenge registryChallenge) (st
 
 	headers.Add("Authorization", signature)
 
-	response, err := makeRequest(ctx, http.MethodGet, redirectURL, headers, nil, &registryOptions{})
+	response, err := makeRequest(ctx, http.MethodGet, redirectURL, headers, nil, nil)
 	if err != nil {
 		return "", err
 	}

server/download.go

@@ -44,19 +44,17 @@ type blobDownload struct {
 
 	context.CancelFunc
 
-	done       chan struct{}
+	done       bool
 	err        error
 	references atomic.Int32
 }
 
 type blobDownloadPart struct {
-	N         int
-	Offset    int64
-	Size      int64
-	Completed atomic.Int64
-
-	lastUpdatedMu sync.Mutex
-	lastUpdated   time.Time
+	N           int
+	Offset      int64
+	Size        int64
+	Completed   int64
+	lastUpdated time.Time
 
 	*blobDownload `json:"-"`
 }
@@ -74,7 +72,7 @@ func (p *blobDownloadPart) Name() string {
 }
 
 func (p *blobDownloadPart) StartsAt() int64 {
-	return p.Offset + p.Completed.Load()
+	return p.Offset + p.Completed
 }
 
 func (p *blobDownloadPart) StopsAt() int64 {
@@ -84,9 +82,7 @@ func (p *blobDownloadPart) StopsAt() int64 {
 func (p *blobDownloadPart) Write(b []byte) (n int, err error) {
 	n = len(b)
 	p.blobDownload.Completed.Add(int64(n))
-	p.lastUpdatedMu.Lock()
 	p.lastUpdated = time.Now()
-	p.lastUpdatedMu.Unlock()
 	return n, nil
 }
 
@@ -96,8 +92,6 @@ func (b *blobDownload) Prepare(ctx context.Context, requestURL *url.URL, opts *r
 		return err
 	}
 
-	b.done = make(chan struct{})
-
 	for _, partFilePath := range partFilePaths {
 		part, err := b.readPart(partFilePath)
 		if err != nil {
@@ -105,7 +99,7 @@ func (b *blobDownload) Prepare(ctx context.Context, requestURL *url.URL, opts *r
 		}
 
 		b.Total += part.Size
-		b.Completed.Add(part.Completed.Load())
+		b.Completed.Add(part.Completed)
 		b.Parts = append(b.Parts, part)
 	}
 
@@ -145,7 +139,6 @@ func (b *blobDownload) Prepare(ctx context.Context, requestURL *url.URL, opts *r
 }
 
 func (b *blobDownload) Run(ctx context.Context, requestURL *url.URL, opts *registryOptions) {
-	defer close(b.done)
 	b.err = b.run(ctx, requestURL, opts)
 }
 
@@ -237,7 +230,7 @@ func (b *blobDownload) run(ctx context.Context, requestURL *url.URL, opts *regis
 	g.SetLimit(numDownloadParts)
 	for i := range b.Parts {
 		part := b.Parts[i]
-		if part.Completed.Load() == part.Size {
+		if part.Completed == part.Size {
 			continue
 		}
 
@@ -245,7 +238,7 @@ func (b *blobDownload) run(ctx context.Context, requestURL *url.URL, opts *regis
 			var err error
 			for try := 0; try < maxRetries; try++ {
 				w := io.NewOffsetWriter(file, part.StartsAt())
-				err = b.downloadChunk(inner, directURL, w, part)
+				err = b.downloadChunk(inner, directURL, w, part, opts)
 				switch {
 				case errors.Is(err, context.Canceled), errors.Is(err, syscall.ENOSPC):
 					// return immediately if the context is canceled or the device is out of space
@@ -286,31 +279,29 @@ func (b *blobDownload) run(ctx context.Context, requestURL *url.URL, opts *regis
 		return err
 	}
 
+	b.done = true
 	return nil
 }
 
-func (b *blobDownload) downloadChunk(ctx context.Context, requestURL *url.URL, w io.Writer, part *blobDownloadPart) error {
+func (b *blobDownload) downloadChunk(ctx context.Context, requestURL *url.URL, w io.Writer, part *blobDownloadPart, opts *registryOptions) error {
 	g, ctx := errgroup.WithContext(ctx)
 	g.Go(func() error {
-		req, err := http.NewRequestWithContext(ctx, http.MethodGet, requestURL.String(), nil)
-		if err != nil {
-			return err
-		}
-		req.Header.Set("Range", fmt.Sprintf("bytes=%d-%d", part.StartsAt(), part.StopsAt()-1))
-		resp, err := http.DefaultClient.Do(req)
+		headers := make(http.Header)
+		headers.Set("Range", fmt.Sprintf("bytes=%d-%d", part.StartsAt(), part.StopsAt()-1))
+		resp, err := makeRequestWithRetry(ctx, http.MethodGet, requestURL, headers, nil, opts)
 		if err != nil {
 			return err
 		}
 		defer resp.Body.Close()
 
-		n, err := io.CopyN(w, io.TeeReader(resp.Body, part), part.Size-part.Completed.Load())
+		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)
 			return err
 		}
 
-		part.Completed.Add(n)
+		part.Completed += n
 		if err := b.writePart(part.Name(), part); err != nil {
 			return err
 		}
@@ -324,21 +315,15 @@ func (b *blobDownload) downloadChunk(ctx context.Context, requestURL *url.URL, w
 		for {
 			select {
 			case <-ticker.C:
-				if part.Completed.Load() >= part.Size {
+				if part.Completed >= part.Size {
 					return nil
 				}
 
-				part.lastUpdatedMu.Lock()
-				lastUpdated := part.lastUpdated
-				part.lastUpdatedMu.Unlock()
-
-				if !lastUpdated.IsZero() && time.Since(lastUpdated) > 5*time.Second {
+				if !part.lastUpdated.IsZero() && time.Since(part.lastUpdated) > 5*time.Second {
 					const msg = "%s part %d stalled; retrying. If this persists, press ctrl-c to exit, then 'ollama pull' to find a faster connection."
 					slog.Info(fmt.Sprintf(msg, b.Digest[7:19], part.N))
 					// reset last updated
-					part.lastUpdatedMu.Lock()
 					part.lastUpdated = time.Time{}
-					part.lastUpdatedMu.Unlock()
 					return errPartStalled
 				}
 			case <-ctx.Done():
@@ -403,8 +388,6 @@ func (b *blobDownload) Wait(ctx context.Context, fn func(api.ProgressResponse))
 	ticker := time.NewTicker(60 * time.Millisecond)
 	for {
 		select {
-		case <-b.done:
-			return b.err
 		case <-ticker.C:
 			fn(api.ProgressResponse{
 				Status:    fmt.Sprintf("pulling %s", b.Digest[7:19]),
@@ -412,6 +395,10 @@ func (b *blobDownload) Wait(ctx context.Context, fn func(api.ProgressResponse))
 				Total:     b.Total,
 				Completed: b.Completed.Load(),
 			})
+
+			if b.done || b.err != nil {
+				return b.err
+			}
 		case <-ctx.Done():
 			return ctx.Err()
 		}

server/routes.go

@@ -284,7 +284,6 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 }
 
 func (s *Server) EmbedHandler(c *gin.Context) {
-	checkpointStart := time.Now()
 	var req api.EmbedRequest
 	err := c.ShouldBindJSON(&req)
 	switch {
@@ -333,8 +332,6 @@ func (s *Server) EmbedHandler(c *gin.Context) {
 		return
 	}
 
-	checkpointLoaded := time.Now()
-
 	kvData, err := getKVData(m.ModelPath, false)
 	if err != nil {
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
@@ -373,16 +370,13 @@ func (s *Server) EmbedHandler(c *gin.Context) {
 		return
 	}
 
-	for i, e := range embeddings.Embedding {
-		embeddings.Embedding[i] = normalize(e)
+	for i, e := range embeddings {
+		embeddings[i] = normalize(e)
 	}
 
 	resp := api.EmbedResponse{
-		Model:           req.Model,
-		Embeddings:      embeddings.Embedding,
-		TotalDuration:   time.Since(checkpointStart),
-		LoadDuration:    checkpointLoaded.Sub(checkpointStart),
-		PromptEvalCount: embeddings.PromptEvalCount,
+		Model:      req.Model,
+		Embeddings: embeddings,
 	}
 	c.JSON(http.StatusOK, resp)
 }
@@ -434,9 +428,9 @@ func (s *Server) EmbeddingsHandler(c *gin.Context) {
 		return
 	}
 
-	embedding := make([]float64, len(embeddings.Embedding[0]))
+	embedding := make([]float64, len(embeddings[0]))
 
-	for i, v := range embeddings.Embedding[0] {
+	for i, v := range embeddings[0] {
 		embedding[i] = float64(v)
 	}
 

server/sched.go

@@ -132,6 +132,8 @@ func (s *Scheduler) processPending(ctx context.Context) {
 			if len(pending.model.ProjectorPaths) > 0 && numParallel != 1 {
 				numParallel = 1
 				slog.Warn("multimodal models don't support parallel requests yet")
+			} else if strings.Contains(pending.model.Config.ModelFamily, "bert") {
+				numParallel = runtime.NumCPU()
 			}
 
 			for {
@@ -212,12 +214,9 @@ func (s *Scheduler) processPending(ctx context.Context) {
 					} else if loadedCount == 0 {
 						// No models loaded. Load the model but prefer the best fit.
 						slog.Debug("loading first model", "model", pending.model.ModelPath)
-						g := pickBestFullFitByLibrary(pending, ggml, gpus, &numParallel)
+						g := pickBestFitGPUs(pending, ggml, gpus, &numParallel)
 						if g != nil {
 							gpus = g
-						} else {
-							// Only allow partial loads when this is the first model
-							gpus = pickBestPartialFitByLibrary(pending, ggml, gpus, &numParallel)
 						}
 						s.loadFn(pending, ggml, gpus, numParallel)
 						break
@@ -234,7 +233,7 @@ func (s *Scheduler) processPending(ctx context.Context) {
 
 						// Update free memory from currently loaded models
 						s.updateFreeSpace(availGpus)
-						fitGpus := pickBestFullFitByLibrary(pending, ggml, availGpus, &numParallel)
+						fitGpus := pickBestFitGPUs(pending, ggml, availGpus, &numParallel)
 						if fitGpus != nil {
 							slog.Debug("new model fits with existing models, loading")
 							s.loadFn(pending, ggml, fitGpus, numParallel)
@@ -671,12 +670,11 @@ func (a ByDuration) Less(i, j int) bool {
 // func (a BySize) Swap(i, j int)      { a[i], a[j] = a[j], a[i] }
 // func (a BySize) Less(i, j int) bool { return a[i].estimatedVRAM < a[j].estimatedVRAM }
 
-// pickBestFullFitByLibrary will try to find the optimal placement of the model in the available GPUs where the model fully fits
-// The list of GPUs returned will always be the same brand (library)
+// pickBestFitGPUs will try to find the optimal placement of the model in the available GPUs where the model fully fits
 // If the model can not be fit fully within the available GPU(s) nil is returned
 // If numParallel is <= 0, this will attempt try to optimize parallism based on available VRAM, and adjust
 // opts.NumCtx accordingly
-func pickBestFullFitByLibrary(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel *int) gpu.GpuInfoList {
+func pickBestFitGPUs(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel *int) gpu.GpuInfoList {
 	var estimatedVRAM uint64
 
 	var numParallelToTry []int
@@ -727,25 +725,6 @@ func pickBestFullFitByLibrary(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoL
 	return nil
 }
 
-// If multiple Libraries are detected, pick the Library which loads the most layers for the model
-func pickBestPartialFitByLibrary(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel *int) gpu.GpuInfoList {
-	*numParallel = 1
-	byLibrary := gpus.ByLibrary()
-	if len(byLibrary) <= 1 {
-		return gpus
-	}
-	var bestEstimate uint64
-	var bestFit int
-	for i, gl := range byLibrary {
-		_, estimatedVRAM := llm.PredictServerFit(gl, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts)
-		if estimatedVRAM > bestEstimate {
-			bestEstimate = estimatedVRAM
-			bestFit = i
-		}
-	}
-	return byLibrary[bestFit]
-}
-
 // findRunnerToUnload finds a runner to unload to make room for a new model
 func (s *Scheduler) findRunnerToUnload() *runnerRef {
 	s.loadedMu.Lock()

server/sched_test.go

@@ -666,50 +666,11 @@ func TestAlreadyCanceled(t *testing.T) {
 	require.Empty(t, scenario1a.req.successCh)
 }
 
-func TestHomogeneousGPUs(t *testing.T) {
-	ctx, done := context.WithTimeout(context.Background(), 100*time.Millisecond)
-	defer done()
-	s := InitScheduler(ctx)
-
-	s.getGpuFn = func() gpu.GpuInfoList {
-		// Set memory values to require the model to be spread
-		gpus := []gpu.GpuInfo{
-			{Library: "cuda"},
-			{Library: "rocm"},
-		}
-		gpus[0].TotalMemory = 1 * format.GibiByte
-		gpus[0].FreeMemory = 256 * format.MebiByte
-		gpus[1].TotalMemory = 1 * format.GibiByte
-		gpus[1].FreeMemory = 256 * format.MebiByte
-		return gpus
-	}
-	s.getCpuFn = getCpuFn
-	a := newScenarioRequest(t, ctx, "ollama-model-1", 10, &api.Duration{Duration: 5 * time.Millisecond})
-	s.newServerFn = func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options, numParallel int) (llm.LlamaServer, error) {
-		require.Len(t, gpus, 1)
-		return a.newServer(gpus, model, ggml, adapters, projectors, opts, numParallel)
-	}
-	slog.Info("a")
-	s.pendingReqCh <- a.req
-	require.Len(t, s.pendingReqCh, 1)
-	s.Run(ctx)
-	select {
-	case resp := <-a.req.successCh:
-		require.Equal(t, resp.llama, a.srv)
-		require.Empty(t, s.pendingReqCh)
-		require.Empty(t, a.req.errCh)
-	case err := <-a.req.errCh:
-		t.Fatal(err.Error())
-	case <-ctx.Done():
-		t.Fatal("timeout")
-	}
-}
-
 type mockLlm struct {
 	pingResp           error
 	waitResp           error
 	completionResp     error
-	embedResp          *llm.EmbedResponse
+	embedResp          [][]float32
 	embedRespErr       error
 	tokenizeResp       []int
 	tokenizeRespErr    error
@@ -727,7 +688,7 @@ func (s *mockLlm) WaitUntilRunning(ctx context.Context) error { return s.waitRes
 func (s *mockLlm) Completion(ctx context.Context, req llm.CompletionRequest, fn func(llm.CompletionResponse)) error {
 	return s.completionResp
 }
-func (s *mockLlm) Embed(ctx context.Context, input []string) (*llm.EmbedResponse, error) {
+func (s *mockLlm) Embed(ctx context.Context, input []string) ([][]float32, error) {
 	return s.embedResp, s.embedRespErr
 }
 func (s *mockLlm) Tokenize(ctx context.Context, content string) ([]int, error) {

server/upload.go

@@ -254,7 +254,7 @@ func (b *blobUpload) uploadPart(ctx context.Context, method string, requestURL *
 
 		// retry uploading to the redirect URL
 		for try := range maxRetries {
-			err = b.uploadPart(ctx, http.MethodPut, redirectURL, part, &registryOptions{})
+			err = b.uploadPart(ctx, http.MethodPut, redirectURL, part, nil)
 			switch {
 			case errors.Is(err, context.Canceled):
 				return err