Merge https://github.com/ollama/ollama

Signed-off-by: baalajimaestro <me@baalajimaestro.me>

.github/workflows/release.yaml

@@ -31,7 +31,7 @@ jobs:
           security set-keychain-settings -lut 3600 build.keychain
       - uses: actions/setup-go@v5
         with:
-          go-version-file: go.mod
+          go-version: "stable"
           cache: true
       - name: Build Darwin
         env:
@@ -87,7 +87,7 @@ jobs:
           write-host "plugin installed"
       - uses: actions/setup-go@v5
         with:
-          go-version-file: go.mod
+          go-version: "stable"
           cache: true
       - run: go get ./...
       - run: |
@@ -141,7 +141,7 @@ jobs:
           write-host "plugin installed"
       - uses: actions/setup-go@v5
         with:
-          go-version-file: go.mod
+          go-version: "stable"
           cache: true
       - name: 'Install ROCm'
         run: |
@@ -218,7 +218,7 @@ jobs:
           write-host "plugin installed"
       - uses: actions/setup-go@v5
         with:
-          go-version-file: go.mod
+          go-version: "stable"
           cache: true
       - name: 'Install CUDA'
         run: |
@@ -306,7 +306,7 @@ jobs:
           write-host "plugin installed"
       - uses: actions/setup-go@v5
         with:
-          go-version-file: go.mod
+          go-version: "stable"
           cache: true
       - run: go get
       - uses: actions/download-artifact@v4

.github/workflows/test.yaml

@@ -63,7 +63,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions/setup-go@v5
         with:
-          go-version-file: go.mod
+          go-version: "stable"
           cache: true
       - run: go get ./...
       - run: |
@@ -163,7 +163,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions/setup-go@v5
         with:
-          go-version-file: go.mod
+          go-version: "stable"
           cache: true
       - name: 'Install ROCm'
         run: |
@@ -200,7 +200,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions/setup-go@v5
         with:
-          go-version-file: go.mod
+          go-version: "stable"
           cache: true
       - name: 'Install CUDA'
         run: |
@@ -255,7 +255,7 @@ jobs:
           submodules: recursive
       - uses: actions/setup-go@v5
         with:
-          go-version-file: go.mod
+          go-version: "stable"
           cache: false
       - run: |
           case ${{ matrix.arch }} in
@@ -297,7 +297,7 @@ jobs:
           submodules: recursive
       - uses: actions/setup-go@v5
         with:
-          go-version-file: go.mod
+          go-version: "stable"
           cache: true
       - run: |
           case ${{ matrix.arch }} in

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](https://ollama.com/library/llama3):
+To run and chat with [Llama 3.1](https://ollama.com/library/llama3.1):
 
 ```
-ollama run llama3
+ollama run llama3.1
 ```
 
 ## Model library
@@ -49,8 +49,9 @@ Here are some example models that can be downloaded:
 
 | Model              | Parameters | Size  | Download                       |
 | ------------------ | ---------- | ----- | ------------------------------ |
-| Llama 3            | 8B         | 4.7GB | `ollama run llama3`            |
+| Llama 3.1          | 8B         | 4.7GB | `ollama run llama3.1`          |
-| Llama 3            | 70B        | 40GB  | `ollama run llama3:70b`        |
+| Llama 3.1          | 70B        | 40GB  | `ollama run llama3.1:70b`      |
+| Llama 3.1          | 405B       | 231GB | `ollama run llama3.1:405b`     |
 | 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`            |
@@ -64,7 +65,8 @@ Here are some example models that can be downloaded:
 | LLaVA              | 7B         | 4.5GB | `ollama run llava`             |
 | 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.
+> [!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.
 
 ## Customize a model
 
@@ -96,16 +98,16 @@ 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` model:
+Models from the Ollama library can be customized with a prompt. For example, to customize the `llama3.1` model:
 
 ```
-ollama pull llama3
+ollama pull llama3.1
 ```
 
 Create a `Modelfile`:
 
 ```
-FROM llama3
+FROM llama3.1
 
 # set the temperature to 1 [higher is more creative, lower is more coherent]
 PARAMETER temperature 1
@@ -140,7 +142,7 @@ ollama create mymodel -f ./Modelfile
 ### Pull a model
 
 ```
-ollama pull llama3
+ollama pull llama3.1
 ```
 
 > This command can also be used to update a local model. Only the diff will be pulled.
@@ -148,13 +150,13 @@ ollama pull llama3
 ### Remove a model
 
 ```
-ollama rm llama3
+ollama rm llama3.1
 ```
 
 ### Copy a model
 
 ```
-ollama cp llama3 my-model
+ollama cp llama3.1 my-model
 ```
 
 ### Multiline input
@@ -171,21 +173,21 @@ I'm a basic program that prints the famous "Hello, world!" message to the consol
 ### Multimodal models
 
 ```
->>> What's in this image? /Users/jmorgan/Desktop/smile.png
+ollama run llava "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 "Summarize this file: $(cat README.md)"
+$ ollama run llama3.1 "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
+ollama show llama3.1
 ```
 
 ### List models on your computer
@@ -213,7 +215,7 @@ Next, start the server:
 Finally, in a separate shell, run a model:
 
 ```
-./ollama run llama3
+./ollama run llama3.1
 ```
 
 ## REST API
@@ -224,7 +226,7 @@ Ollama has a REST API for running and managing models.
 
 ```
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3",
+  "model": "llama3.1",
   "prompt":"Why is the sky blue?"
 }'
 ```
@@ -233,7 +235,7 @@ curl http://localhost:11434/api/generate -d '{
 
 ```
 curl http://localhost:11434/api/chat -d '{
-  "model": "llama3",
+  "model": "llama3.1",
   "messages": [
     { "role": "user", "content": "why is the sky blue?" }
   ]
@@ -296,6 +298,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Kerlig AI](https://www.kerlig.com/) (AI writing assistant for macOS)
 - [AI Studio](https://github.com/MindWorkAI/AI-Studio)
 - [Sidellama](https://github.com/gyopak/sidellama) (browser-based LLM client)
+- [LLMStack](https://github.com/trypromptly/LLMStack) (No-code multi-agent framework to build LLM agents and workflows)
 
 ### Terminal
 
@@ -334,6 +337,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 ### Libraries
 
 - [LangChain](https://python.langchain.com/docs/integrations/llms/ollama) and [LangChain.js](https://js.langchain.com/docs/modules/model_io/models/llms/integrations/ollama) with [example](https://js.langchain.com/docs/use_cases/question_answering/local_retrieval_qa)
+- [Firebase Genkit](https://firebase.google.com/docs/genkit/plugins/ollama)
 - [LangChainGo](https://github.com/tmc/langchaingo/) with [example](https://github.com/tmc/langchaingo/tree/main/examples/ollama-completion-example)
 - [LangChain4j](https://github.com/langchain4j/langchain4j) with [example](https://github.com/langchain4j/langchain4j-examples/tree/main/ollama-examples/src/main/java)
 - [LangChainRust](https://github.com/Abraxas-365/langchain-rust) with [example](https://github.com/Abraxas-365/langchain-rust/blob/main/examples/llm_ollama.rs)

SECURITY.md

@@ -0,0 +1,25 @@
+# Security
+
+The Ollama maintainer team takes security seriously and will actively work to resolve security issues.
+
+## Reporting a vulnerability
+
+If you discover a security vulnerability, please do not open a public issue. Instead, please report it by emailing hello@ollama.com. We ask that you give us sufficient time to investigate and address the vulnerability before disclosing it publicly.
+
+Please include the following details in your report:
+- A description of the vulnerability
+- Steps to reproduce the issue
+- Your assessment of the potential impact
+- Any possible mitigations
+
+## Security best practices
+
+While the maintainer team does their best to secure Ollama, users are encouraged to implement their own security best practices, such as:
+
+- Regularly updating to the latest version of Ollama
+- Securing access to hosted instances of Ollama
+- Monitoring systems for unusual activity
+
+## Contact
+
+For any other questions or concerns related to security, please contact us at hello@ollama.com

api/types.go

@@ -114,6 +114,11 @@ 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.
@@ -209,6 +214,7 @@ 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"`
@@ -261,6 +267,10 @@ 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
+;FinishedLabel=%nRun this command in a PowerShell or cmd terminal.%n%n%n    ollama run llama3.1
 ;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"
+write-host "`tollama run llama3.1"
 write-host ""

cmd/cmd.go

@@ -1341,10 +1341,10 @@ 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"],
-				envVars["OLLAMA_MAX_VRAM"],
 			})
 		default:
 			appendEnvDocs(cmd, envs)

cmd/interactive.go

@@ -1,6 +1,7 @@
 package cmd
 
 import (
+	"cmp"
 	"errors"
 	"fmt"
 	"io"
@@ -9,13 +10,14 @@ 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"
@@ -138,6 +140,7 @@ 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")
@@ -506,31 +509,35 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 }
 
 func buildModelfile(opts runOptions) string {
-	var mf strings.Builder
+	var f parser.File
-	model := opts.ParentModel
+	f.Commands = append(f.Commands, parser.Command{Name: "model", Args: cmp.Or(opts.ParentModel, opts.Model)})
-	if model == "" {
+
-		model = opts.Model
-	}
-	fmt.Fprintf(&mf, "FROM %s\n", model)
 	if opts.System != "" {
-		fmt.Fprintf(&mf, "SYSTEM \"\"\"%s\"\"\"\n", opts.System)
+		f.Commands = append(f.Commands, parser.Command{Name: "system", Args: opts.System})
 	}
 
-	keys := make([]string, 0)
+	keys := maps.Keys(opts.Options)
-	for k := range opts.Options {
+	slices.Sort(keys)
-		keys = append(keys, k)
-	}
-	sort.Strings(keys)
 	for _, k := range keys {
-		fmt.Fprintf(&mf, "PARAMETER %s %v\n", k, opts.Options[k])
+		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.Fprintln(&mf)
 
 	for _, msg := range opts.Messages {
-		fmt.Fprintf(&mf, "MESSAGE %s \"\"\"%s\"\"\"\n", msg.Role, msg.Content)
+		f.Commands = append(f.Commands, parser.Command{Name: "message", Args: fmt.Sprintf("%s: %s", msg.Role, msg.Content)})
 	}
 
-	return mf.String()
+	return f.String()
 }
 
 func normalizeFilePath(fp string) string {

cmd/interactive_test.go

@@ -1,12 +1,10 @@
 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"
 )
@@ -63,52 +61,47 @@ func TestModelfileBuilder(t *testing.T) {
 			{Role: "user", Content: "Hey there hork!"},
 			{Role: "assistant", Content: "Yes it is true, I am half horse, half shark."},
 		},
-		Options: map[string]interface{}{},
+		Options: map[string]any{
+			"temperature":      0.9,
+			"seed":             42,
+			"penalize_newline": false,
+			"stop":             []string{"hi", "there"},
+		},
 	}
 
-	opts.Options["temperature"] = 0.9
+	t.Run("model", func(t *testing.T) {
-	opts.Options["seed"] = 42
+		expect := `FROM hork
-	opts.Options["penalize_newline"] = false
+SYSTEM You are part horse and part shark, but all hork. Do horklike things
-	opts.Options["stop"] = []string{"hi", "there"}
-
-	mf := buildModelfile(opts)
-	expectedModelfile := `FROM {{.Model}}
-SYSTEM """{{.System}}"""
 PARAMETER penalize_newline false
 PARAMETER seed 42
-PARAMETER stop [hi there]
+PARAMETER stop hi
+PARAMETER stop there
 PARAMETER temperature 0.9
-
+MESSAGE user Hey there hork!
-MESSAGE user """Hey there hork!"""
+MESSAGE assistant Yes it is true, I am half horse, half shark.
-MESSAGE assistant """Yes it is true, I am half horse, half shark."""
 `
 
-	tmpl, err := template.New("").Parse(expectedModelfile)
+		actual := buildModelfile(opts)
-	require.NoError(t, err)
+		if diff := cmp.Diff(expect, actual); diff != "" {
-
+			t.Errorf("mismatch (-want +got):\n%s", diff)
-	var buf bytes.Buffer
+		}
-	err = tmpl.Execute(&buf, opts)
+	})
-	require.NoError(t, err)
-	assert.Equal(t, buf.String(), mf)
 
+	t.Run("parent model", func(t *testing.T) {
 		opts.ParentModel = "horseshark"
-	mf = buildModelfile(opts)
+		expect := `FROM horseshark
-	expectedModelfile = `FROM {{.ParentModel}}
+SYSTEM You are part horse and part shark, but all hork. Do horklike things
-SYSTEM """{{.System}}"""
 PARAMETER penalize_newline false
 PARAMETER seed 42
-PARAMETER stop [hi there]
+PARAMETER stop hi
+PARAMETER stop there
 PARAMETER temperature 0.9
-
+MESSAGE user Hey there hork!
-MESSAGE user """Hey there hork!"""
+MESSAGE assistant Yes it is true, I am half horse, half shark.
-MESSAGE assistant """Yes it is true, I am half horse, half shark."""
 `
-
+		actual := buildModelfile(opts)
-	tmpl, err = template.New("").Parse(expectedModelfile)
+		if diff := cmp.Diff(expect, actual); diff != "" {
-	require.NoError(t, err)
+			t.Errorf("mismatch (-want +got):\n%s", diff)
-
+		}
-	var parentBuf bytes.Buffer
+	})
-	err = tmpl.Execute(&parentBuf, opts)
-	require.NoError(t, err)
-	assert.Equal(t, parentBuf.String(), mf)
 }

convert/mistral.go

@@ -71,6 +71,11 @@ func (m *MistralModel) WriteGGUF(ws io.WriteSeeker) error {
 		"tokenizer.ggml.unknown_token_id": uint32(0),
 	}
 
+	if m.Params.HeadDimension > 0 {
+		kv["llama.attention.key_length"] = uint32(m.Params.HeadDimension)
+		kv["llama.attention.value_length"] = uint32(m.Params.HeadDimension)
+	}
+
 	return llm.NewGGUFV3(m.Params.ByteOrder).Encode(ws, kv, m.Tensors)
 }
 

docs/api.md

@@ -40,6 +40,7 @@ Generate a response for a given prompt with a provided model. This is a streamin
 
 - `model`: (required) the [model name](#model-names)
 - `prompt`: the prompt to generate a response for
+- `suffix`: the text after the model response
 - `images`: (optional) a list of base64-encoded images (for multimodal models such as `llava`)
 
 Advanced parameters (optional):
@@ -57,7 +58,8 @@ Advanced parameters (optional):
 
 Enable JSON mode by setting the `format` parameter to `json`. This will structure the response as a valid JSON object. See the JSON mode [example](#request-json-mode) below.
 
-> Note: it's important to instruct the model to use JSON in the `prompt`. Otherwise, the model may generate large amounts whitespace.
+> [!IMPORTANT]
+> It's important to instruct the model to use JSON in the `prompt`. Otherwise, the model may generate large amounts whitespace.
 
 ### Examples
 
@@ -148,8 +150,44 @@ If `stream` is set to `false`, the response will be a single JSON object:
 }
 ```
 
+#### Request (with suffix)
+
+##### Request
+
+```shell
+curl http://localhost:11434/api/generate -d '{
+  "model": "codellama:code",
+  "prompt": "def compute_gcd(a, b):",
+  "suffix": "    return result",
+  "options": {
+    "temperature": 0
+  },
+  "stream": false
+}'
+```
+
+##### Response
+
+```json
+{
+  "model": "codellama:code",
+  "created_at": "2024-07-22T20:47:51.147561Z",
+  "response": "\n  if a == 0:\n    return b\n  else:\n    return compute_gcd(b % a, a)\n\ndef compute_lcm(a, b):\n  result = (a * b) / compute_gcd(a, b)\n",
+  "done": true,
+  "done_reason": "stop",
+  "context": [...],
+  "total_duration": 1162761250,
+  "load_duration": 6683708,
+  "prompt_eval_count": 17,
+  "prompt_eval_duration": 201222000,
+  "eval_count": 63,
+  "eval_duration": 953997000
+}
+```
+
 #### Request (JSON mode)
 
+> [!IMPORTANT]
 > When `format` is set to `json`, the output will always be a well-formed JSON object. It's important to also instruct the model to respond in JSON.
 
 ##### Request
@@ -298,6 +336,7 @@ 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,
@@ -380,12 +419,14 @@ Generate the next message in a chat with a provided model. This is a streaming e
 
 - `model`: (required) the [model name](#model-names)
 - `messages`: the messages of the chat, this can be used to keep a chat memory
+- `tools`: tools for the model to use if supported. Requires `stream` to be set to `false`
 
 The `message` object has the following fields:
 
-- `role`: the role of the message, either `system`, `user` or `assistant`
+- `role`: the role of the message, either `system`, `user`, `assistant`, or `tool`
 - `content`: the content of the message
 - `images` (optional): a list of images to include in the message (for multimodal models such as `llava`)
+- `tool_calls` (optional): a list of tools the model wants to use
 
 Advanced parameters (optional):
 
@@ -546,7 +587,7 @@ Final response:
 
 ##### Request
 
-Send a chat message with a conversation history.
+Send a chat message with images. The images should be provided as an array, with the individual images encoded in Base64.
 
 ```shell
 curl http://localhost:11434/api/chat -d '{
@@ -622,6 +663,79 @@ curl http://localhost:11434/api/chat -d '{
 }
 ```
 
+#### Chat request (with tools)
+
+##### Request
+
+```
+curl http://localhost:11434/api/chat -d '{
+  "model": "mistral",
+  "messages": [
+    {
+      "role": "user",
+      "content": "What is the weather today in Paris?"
+    }
+  ],
+  "stream": false,
+  "tools": [
+    {
+      "type": "function",
+      "function": {
+        "name": "get_current_weather",
+        "description": "Get the current weather for a location",
+        "parameters": {
+          "type": "object",
+          "properties": {
+            "location": {
+              "type": "string",
+              "description": "The location to get the weather for, e.g. San Francisco, CA"
+            },
+            "format": {
+              "type": "string",
+              "description": "The format to return the weather in, e.g. 'celsius' or 'fahrenheit'",
+              "enum": ["celsius", "fahrenheit"]
+            }
+          },
+          "required": ["location", "format"]
+        }
+      }
+    }
+  ]
+}'
+```
+
+##### Response
+
+```json
+{
+  "model": "mistral:7b-instruct-v0.3-q4_K_M",
+  "created_at": "2024-07-22T20:33:28.123648Z",
+  "message": {
+    "role": "assistant",
+    "content": "",
+    "tool_calls": [
+      {
+        "function": {
+          "name": "get_current_weather",
+          "arguments": {
+            "format": "celsius",
+            "location": "Paris, FR"
+          }
+        }
+      }
+    ]
+  },
+  "done_reason": "stop",
+  "done": true,
+  "total_duration": 885095291,
+  "load_duration": 3753500,
+  "prompt_eval_count": 122,
+  "prompt_eval_duration": 328493000,
+  "eval_count": 33,
+  "eval_duration": 552222000
+}
+```
+
 ## Create a Model
 
 ```shell
@@ -1026,7 +1140,7 @@ If `stream` is set to `false`, then the response is a single JSON object:
 ## Generate Embeddings
 
 ```shell
-POST /api/embeddings
+POST /api/embed
 ```
 
 Generate embeddings from a model
@@ -1034,10 +1148,11 @@ Generate embeddings from a model
 ### Parameters
 
 - `model`: name of model to generate embeddings from
-- `prompt`: text to generate embeddings for
+- `input`: text or list of text to generate embeddings for
 
 Advanced parameters:
 
+- `truncate`: truncates the end of each input to fit within context length. Returns error if `false` and context length is exceeded. Defaults to `true`
 - `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
 - `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
 
@@ -1046,9 +1161,9 @@ Advanced parameters:
 #### Request
 
 ```shell
-curl http://localhost:11434/api/embeddings -d '{
+curl http://localhost:11434/api/embed -d '{
   "model": "all-minilm",
-  "prompt": "Here is an article about llamas..."
+  "input": "Why is the sky blue?"
 }'
 ```
 
@@ -1056,10 +1171,35 @@ curl http://localhost:11434/api/embeddings -d '{
 
 ```json
 {
-  "embedding": [
+  "model": "all-minilm",
-    0.5670403838157654, 0.009260174818336964, 0.23178744316101074, -0.2916173040866852, -0.8924556970596313,
+  "embeddings": [[
-    0.8785552978515625, -0.34576427936553955, 0.5742510557174683, -0.04222835972905159, -0.137906014919281
+    0.010071029, -0.0017594862, 0.05007221, 0.04692972, 0.054916814,
-  ]
+    0.008599704, 0.105441414, -0.025878139, 0.12958129, 0.031952348
+  ]]
+}
+```
+
+#### Request (Multiple input)
+
+```shell
+curl http://localhost:11434/api/embed -d '{
+  "model": "all-minilm",
+  "input": ["Why is the sky blue?", "Why is the grass green?"]
+}'
+```
+
+#### Response
+
+```json
+{
+  "model": "all-minilm",
+  "embeddings": [[
+    0.010071029, -0.0017594862, 0.05007221, 0.04692972, 0.054916814,
+    0.008599704, 0.105441414, -0.025878139, 0.12958129, 0.031952348
+  ],[
+    -0.0098027075, 0.06042469, 0.025257962, -0.006364387, 0.07272725,
+    0.017194884, 0.09032035, -0.051705178, 0.09951512, 0.09072481
+  ]]
 }
 ```
 
@@ -1106,3 +1246,45 @@ A single JSON object will be returned.
   ]
 }
 ```
+
+## Generate Embedding
+
+> Note: this endpoint has been superseded by `/api/embed`
+
+```shell
+POST /api/embeddings
+```
+
+Generate embeddings from a model
+
+### Parameters
+
+- `model`: name of model to generate embeddings from
+- `prompt`: text to generate embeddings for
+
+Advanced parameters:
+
+- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
+- `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
+
+### Examples
+
+#### Request
+
+```shell
+curl http://localhost:11434/api/embeddings -d '{
+  "model": "all-minilm",
+  "prompt": "Here is an article about llamas..."
+}'
+```
+
+#### Response
+
+```json
+{
+  "embedding": [
+    0.5670403838157654, 0.009260174818336964, 0.23178744316101074, -0.2916173040866852, -0.8924556970596313,
+    0.8785552978515625, -0.34576427936553955, 0.5742510557174683, -0.04222835972905159, -0.137906014919281
+  ]
+}
+```

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
+docker exec -it ollama ollama run llama3.1
 ```
 
 ### 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 ""
+ollama run llama3.1 ""
 ```
 
 ## How do I keep a model loaded in memory or make it unload immediately?
@@ -273,3 +273,7 @@ The following server settings may be used to adjust how Ollama handles concurren
 - `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.

docs/modelfile.md

@@ -1,6 +1,7 @@
 # Ollama Model File
 
-> Note: `Modelfile` syntax is in development
+> [!NOTE]
+> `Modelfile` syntax is in development
 
 A model file is the blueprint to create and share models with Ollama.
 
@@ -140,6 +141,7 @@ 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/openai.md

@@ -78,8 +78,8 @@ curl http://localhost:11434/v1/chat/completions \
 - [x] Streaming
 - [x] JSON mode
 - [x] Reproducible outputs
+- [x] Tools (streaming support coming soon)
 - [ ] Vision
-- [ ] Function calling
 - [ ] Logprobs
 
 #### Supported request fields
@@ -97,9 +97,9 @@ curl http://localhost:11434/v1/chat/completions \
 - [x] `temperature`
 - [x] `top_p`
 - [x] `max_tokens`
-- [ ] `logit_bias`
+- [x] `tools`
-- [ ] `tools`
 - [ ] `tool_choice`
+- [ ] `logit_bias`
 - [ ] `user`
 - [ ] `n`
 

docs/template.md

@@ -0,0 +1,173 @@
+# Template
+
+Ollama provides a powerful templating engine backed by Go's built-in templating engine to construct prompts for your large language model. This feature is a valuable tool to get the most out of your models.
+
+## Basic Template Structure
+
+A basic Go template consists of three main parts:
+
+* **Layout**: The overall structure of the template.
+* **Variables**: Placeholders for dynamic data that will be replaced with actual values when the template is rendered.
+* **Functions**: Custom functions or logic that can be used to manipulate the template's content.
+
+Here's an example of a simple chat template:
+
+```gotmpl
+{{- range .Messages }}
+{{ .Role }}: {{ .Content }}
+{{- end }}
+```
+
+In this example, we have:
+
+* A basic messages structure (layout)
+* Three variables: `Messages`, `Role`, and `Content` (variables)
+* A custom function (action) that iterates over an array of items (`range .Messages`) and displays each item
+
+## Adding templates to your model
+
+By default, models imported into Ollama have a default template of `{{ .Prompt }}`, i.e. user inputs are sent verbatim to the LLM. This is appropriate for text or code completion models but lacks essential markers for chat or instruction models.
+
+Omitting a template in these models puts the responsibility of correctly templating input onto the user. Adding a template allows users to easily get the best results from the model.
+
+To add templates in your model, you'll need to add a `TEMPLATE` command to the Modelfile. Here's an example using Meta's Llama 3.
+
+```dockerfile
+FROM llama3
+
+TEMPLATE """{{- if .System }}<|start_header_id|>system<|end_header_id|>
+
+{{ .System }}<|eot_id|>
+{{- end }}
+{{- range .Messages }}<|start_header_id|>{{ .Role }}<|end_header_id|>
+
+{{ .Content }}<|eot_id|>
+{{- end }}<|start_header_id|>assistant<|end_header_id|>
+
+"""
+```
+
+## Variables
+
+`System` (string): system prompt
+
+`Prompt` (string): user prompt
+
+`Response` (string): assistant response
+
+`Suffix` (string): text inserted after the assistant's response
+
+`Messages` (list): list of messages
+
+`Messages[].Role` (string): role which can be one of `system`, `user`, `assistant`, or `tool`
+
+`Messages[].Content` (string):  message content
+
+`Messages[].ToolCalls` (list): list of tools the model wants to call
+
+`Messages[].ToolCalls[].Function` (object): function to call
+
+`Messages[].ToolCalls[].Function.Name` (string): function name
+
+`Messages[].ToolCalls[].Function.Arguments` (map): mapping of argument name to argument value
+
+`Tools` (list): list of tools the model can access
+
+`Tools[].Type` (string): schema type. `type` is always `function`
+
+`Tools[].Function` (object): function definition
+
+`Tools[].Function.Name` (string): function name
+
+`Tools[].Function.Description` (string): function description
+
+`Tools[].Function.Parameters` (object): function parameters
+
+`Tools[].Function.Parameters.Type` (string): schema type. `type` is always `object`
+
+`Tools[].Function.Parameters.Required` (list): list of required properties
+
+`Tools[].Function.Parameters.Properties` (map): mapping of property name to property definition
+
+`Tools[].Function.Parameters.Properties[].Type` (string): property type
+
+`Tools[].Function.Parameters.Properties[].Description` (string): property description
+
+`Tools[].Function.Parameters.Properties[].Enum` (list): list of valid values
+
+## Tips and Best Practices
+
+Keep the following tips and best practices in mind when working with Go templates:
+
+* **Be mindful of dot**: Control flow structures like `range` and `with` changes the value `.`
+* **Out-of-scope variables**: Use `$.` to reference variables not currently in scope, starting from the root
+* **Whitespace control**: Use `-` to trim leading (`{{-`) and trailing (`-}}`) whitespace
+
+## Examples
+
+### Example Messages
+
+#### ChatML
+
+ChatML is a popular template format. It can be used for models such as Databrick's DBRX, Intel's Neural Chat, and Microsoft's Orca 2.
+
+```gotmpl
+{{- if .System }}<|im_start|>system
+{{ .System }}<|im_end|>
+{{ end }}
+{{- range .Messages }}<|im_start|>{{ .Role }}
+{{ .Content }}<|im_end|>
+{{ end }}<|im_start|>assistant
+{{ else }}
+{{ if .System }}<|im_start|>system
+{{ .System }}<|im_end|>
+```
+
+### Example Tools
+
+Tools support can be added to a model by adding a `{{ .Tools }}` node to the template. This feature is useful for models trained to call external tools and can a powerful tool for retrieving real-time data or performing complex tasks.
+
+#### Mistral
+
+Mistral v0.3 and Mixtral 8x22B supports tool calling.
+
+```gotmpl
+{{- range $index, $_ := .Messages }}
+{{- if eq .Role "user" }}
+{{- if and (le (len (slice $.Messages $index)) 2) $.Tools }}[AVAILABLE_TOOLS] {{ json $.Tools }}[/AVAILABLE_TOOLS]
+{{- end }}[INST] {{ if and (eq (len (slice $.Messages $index)) 1) $.System }}{{ $.System }}
+
+{{ end }}{{ .Content }}[/INST]
+{{- else if eq .Role "assistant" }}
+{{- if .Content }} {{ .Content }}</s>
+{{- else if .ToolCalls }}[TOOL_CALLS] [
+{{- range .ToolCalls }}{"name": "{{ .Function.Name }}", "arguments": {{ json .Function.Arguments }}}
+{{- end }}]</s>
+{{- end }}
+{{- else if eq .Role "tool" }}[TOOL_RESULTS] {"content": {{ .Content }}}[/TOOL_RESULTS]
+{{- end }}
+{{- end }}
+```
+
+### Example Fill-in-Middle
+
+Fill-in-middle support can be added to a model by adding a `{{ .Suffix }}` node to the template. This feature is useful for models that are trained to generate text in the middle of user input, such as code completion models.
+
+#### CodeLlama
+
+CodeLlama [7B](https://ollama.com/library/codellama:7b-code) and [13B](https://ollama.com/library/codellama:13b-code) code completion models support fill-in-middle.
+
+```gotmpl
+<PRE> {{ .Prompt }} <SUF>{{ .Suffix }} <MID>
+```
+
+> [!NOTE]
+> CodeLlama 34B and 70B code completion and all instruct and Python fine-tuned models do not support fill-in-middle.
+
+#### Codestral
+
+Codestral [22B](https://ollama.com/library/codestral:22b) supports fill-in-middle.
+
+```gotmpl
+[SUFFIX]{{ .Suffix }}[PREFIX] {{ .Prompt }}
+```

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",
+  model: "llama3.1",
 });
 
 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 "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 llama3.1 "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.
 
 ```bash
 npm install cheerio

docs/windows.md

@@ -23,6 +23,8 @@ 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`

envconfig/config.go

@@ -43,8 +43,6 @@ var (
 	MaxRunners int
 	// Set via OLLAMA_MAX_QUEUE in the environment
 	MaxQueuedRequests int
-	// Set via OLLAMA_MAX_VRAM in the environment
-	MaxVRAM uint64
 	// Set via OLLAMA_MODELS in the environment
 	ModelsDir string
 	// Set via OLLAMA_NOHISTORY in the environment
@@ -89,7 +87,6 @@ func AsMap() map[string]EnvVar {
 		"OLLAMA_LLM_LIBRARY":       {"OLLAMA_LLM_LIBRARY", LLMLibrary, "Set LLM library to bypass autodetection"},
 		"OLLAMA_MAX_LOADED_MODELS": {"OLLAMA_MAX_LOADED_MODELS", MaxRunners, "Maximum number of loaded models per GPU"},
 		"OLLAMA_MAX_QUEUE":         {"OLLAMA_MAX_QUEUE", MaxQueuedRequests, "Maximum number of queued requests"},
-		"OLLAMA_MAX_VRAM":          {"OLLAMA_MAX_VRAM", MaxVRAM, "Maximum VRAM"},
 		"OLLAMA_MODELS":            {"OLLAMA_MODELS", ModelsDir, "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"},
@@ -194,16 +191,6 @@ func LoadConfig() {
 
 	TmpDir = clean("OLLAMA_TMPDIR")
 
-	userLimit := clean("OLLAMA_MAX_VRAM")
-	if userLimit != "" {
-		avail, err := strconv.ParseUint(userLimit, 10, 64)
-		if err != nil {
-			slog.Error("invalid setting, ignoring", "OLLAMA_MAX_VRAM", userLimit, "error", err)
-		} else {
-			MaxVRAM = avail
-		}
-	}
-
 	LLMLibrary = clean("OLLAMA_LLM_LIBRARY")
 
 	if onp := clean("OLLAMA_NUM_PARALLEL"); onp != "" {

examples/go-chat/main.go

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

examples/go-generate-streaming/main.go

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

examples/go-generate/main.go

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

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

@@ -4,6 +4,14 @@ 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="llama3:8b", callback_manager=CallbackManager([StreamingStdOutCallbackHandler()]))
+    llm = Ollama(model="llama3.1", 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 `llama2` model installed:
+1. Ensure you have the `llama3.1` model installed:
 
    ```bash
-   ollama pull llama2
+   ollama pull llama3.1
    ```
 
 2. Install the Python Requirements.

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

@@ -5,7 +5,7 @@ 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")
+llm = Ollama(model="llama3.1")
 chain = load_summarize_chain(llm, chain_type="stuff")
 
 result = chain.invoke(docs)

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` model installed:
+1. Ensure you have the `llama3.1` model installed:
 
    ```bash
-   ollama pull llama3
+   ollama pull llama3.1
    ```
 
 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")
+llm = Ollama(model="llama3.1")
 res = llm.predict(input)
 print (res)

examples/modelfile-mario/Modelfile

@@ -1,4 +1,4 @@
-FROM llama3
+FROM llama3.1
 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 as the base model.
+This example shows how to create a basic character using Llama3.1 as the base model.
 
 To run this example:
 
 1. Download the Modelfile
-2. `ollama pull llama3` to get the base model used in the model file.
+2. `ollama pull llama3.1` 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
+FROM llama3.1
 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': 'dockerit', 'prompt': inputDescription}, stream=True) as r:
+with s.post('http://localhost:11434/api/generate', json={'model': 'mattw/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"
+model = "llama3.1"
 template = {
   "firstName": "",
   "lastName": "",

examples/python-json-datagenerator/randomaddresses.py

@@ -12,7 +12,7 @@ countries = [
     "France",
 ]
 country = random.choice(countries)
-model = "llama3"
+model = "llama3.1"
 
 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` model installed:
+1. Ensure you have the `llama3.1` model installed:
 
    ```bash
-   ollama pull llama3
+   ollama pull llama3.1
    ```
 
 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"  # TODO: update this for whatever model you wish to use
+model = "llama3.1"  # 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` model installed:
+1. Ensure you have the `llama3.1` model installed:
 
    ```bash
-   ollama pull llama3
+   ollama pull llama3.1
    ```
 
 2. Install the Python Requirements.

examples/typescript-simplechat/client.ts

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

gpu/amd_linux.go

@@ -10,6 +10,7 @@ import (
 	"path/filepath"
 	"regexp"
 	"slices"
+	"sort"
 	"strconv"
 	"strings"
 
@@ -82,6 +83,20 @@ 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/concurrency_test.go

@@ -69,7 +69,7 @@ func TestIntegrationConcurrentPredictOrcaMini(t *testing.T) {
 	reqLimit := len(req)
 	iterLimit := 5
 
-	vram := os.Getenv("OLLAMA_MAX_VRAM")
+	vram := os.Getenv("OLLAMA_MAX_VRAM") // TODO - discover actual VRAM
 	if vram != "" {
 		max, err := strconv.ParseUint(vram, 10, 64)
 		require.NoError(t, err)
@@ -106,7 +106,7 @@ func TestIntegrationConcurrentPredictOrcaMini(t *testing.T) {
 
 // Stress the system if we know how much VRAM it has, and attempt to load more models than will fit
 func TestMultiModelStress(t *testing.T) {
-	vram := os.Getenv("OLLAMA_MAX_VRAM")
+	vram := os.Getenv("OLLAMA_MAX_VRAM") // TODO - discover actual VRAM
 	if vram == "" {
 		t.Skip("OLLAMA_MAX_VRAM not specified, can't pick the right models for the stress test")
 	}

integration/embed_test.go

@@ -4,12 +4,45 @@ package integration
 
 import (
 	"context"
+	"math"
 	"testing"
 	"time"
 
 	"github.com/ollama/ollama/api"
 )
 
+func floatsEqual32(a, b float32) bool {
+	return math.Abs(float64(a-b)) <= 1e-4
+}
+
+func floatsEqual64(a, b float64) bool {
+	return math.Abs(a-b) <= 1e-4
+}
+
+func TestAllMiniLMEmbeddings(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
+	defer cancel()
+
+	req := api.EmbeddingRequest{
+		Model:  "all-minilm",
+		Prompt: "why is the sky blue?",
+	}
+
+	res, err := embeddingTestHelper(ctx, t, req)
+
+	if err != nil {
+		t.Fatalf("error: %v", err)
+	}
+
+	if len(res.Embedding) != 384 {
+		t.Fatalf("expected 384 floats, got %d", len(res.Embedding))
+	}
+
+	if !floatsEqual64(res.Embedding[0], 0.06642947345972061) {
+		t.Fatalf("expected 0.06642947345972061, got %.16f", res.Embedding[0])
+	}
+}
+
 func TestAllMiniLMEmbed(t *testing.T) {
 	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
 	defer cancel()
@@ -33,8 +66,12 @@ func TestAllMiniLMEmbed(t *testing.T) {
 		t.Fatalf("expected 384 floats, got %d", len(res.Embeddings[0]))
 	}
 
-	if res.Embeddings[0][0] != 0.010071031 {
+	if !floatsEqual32(res.Embeddings[0][0], 0.010071031) {
-		t.Fatalf("expected 0.010071031, got %f", res.Embeddings[0][0])
+		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)
 	}
 }
 
@@ -61,12 +98,16 @@ func TestAllMiniLMBatchEmbed(t *testing.T) {
 		t.Fatalf("expected 384 floats, got %d", len(res.Embeddings[0]))
 	}
 
-	if res.Embeddings[0][0] != 0.010071031 || res.Embeddings[1][0] != -0.009802706 {
+	if !floatsEqual32(res.Embeddings[0][0], 0.010071031) || !floatsEqual32(res.Embeddings[1][0], -0.009802706) {
-		t.Fatalf("expected 0.010071031 and -0.009802706, got %f and %f", res.Embeddings[0][0], res.Embeddings[1][0])
+		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) {
+func TestAllMiniLMEmbedTruncate(t *testing.T) {
 	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
 	defer cancel()
 
@@ -135,6 +176,22 @@ func TestAllMiniLmEmbedTruncate(t *testing.T) {
 	}
 }
 
+func embeddingTestHelper(ctx context.Context, t *testing.T, req api.EmbeddingRequest) (*api.EmbeddingResponse, error) {
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+	if err := PullIfMissing(ctx, client, req.Model); err != nil {
+		t.Fatalf("failed to pull model %s: %v", req.Model, err)
+	}
+
+	response, err := client.Embeddings(ctx, &req)
+
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
 func embedTestHelper(ctx context.Context, t *testing.T, req api.EmbedRequest) (*api.EmbedResponse, error) {
 	client, _, cleanup := InitServerConnection(ctx, t)
 	defer cleanup()

llm/ext_server/server.cpp

@@ -41,6 +41,7 @@
 
 #if defined(_WIN32)
 #include <windows.h>
+#include <errhandlingapi.h>
 #endif
 
 #include <cstddef>
@@ -1220,6 +1221,7 @@ struct llama_server_context
                 res.result_json = json
                 {
                     {"embedding", std::vector<float>(embd, embd + n_embd)},
+                    {"timings",             slot.get_formated_timings()},
                 };
             }
         }
@@ -2437,15 +2439,6 @@ 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;
@@ -2737,6 +2730,9 @@ 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
@@ -3208,11 +3204,15 @@ 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}};
+                    json embedding_res = json{{"embedding", embeddings}, {"prompt_n", prompt_n}};
                     return res.set_content(embedding_res.dump(), "application/json; charset=utf-8");
                 }
             });

llm/llama.cpp

@@ -1 +1 @@
-Subproject commit a8db2a9ce64cd4417f6a312ab61858f17f0f8584
+Subproject commit 6eeaeba126ff701f3e8f79f246805b7023709972

llm/llm_darwin_amd64.go

@@ -2,7 +2,10 @@ 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

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

llm/llm_linux.go

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

llm/llm_windows.go

@@ -1,6 +1,20 @@
 package llm
 
-import "embed"
+import (
+	"embed"
+	"syscall"
+)
 
 // 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 2b9ace28..172640e2 100644
+index a207451f..2ddf431d 100644
 --- a/src/llama.cpp
 +++ b/src/llama.cpp
-@@ -5357,16 +5357,7 @@ static void llm_load_vocab(
+@@ -5347,16 +5347,7 @@ 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 2b9ace28..172640e2 100644
                  vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
              } else if (
                      tokenizer_pre == "llama3"   ||
-@@ -5439,7 +5430,8 @@ static void llm_load_vocab(
+@@ -5443,7 +5434,8 @@ static void llm_load_vocab(
-                 tokenizer_pre == "jais") {
+                 tokenizer_pre == "codeshell") {
-                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_JAIS;
+                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_CODESHELL;
              } 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/06-qwen2.diff

@@ -1,13 +0,0 @@
-diff --git a/src/llama.cpp b/src/llama.cpp
-index 40d2ec2c..f34eb79a 100644
---- a/src/llama.cpp
-+++ b/src/llama.cpp
-@@ -6943,7 +6943,7 @@ static struct ggml_tensor * llm_build_kqv(
-         struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
-         cb(kq, "kq", il);
- 
--        if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX) {
-+        if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX || model.arch == LLM_ARCH_QWEN2) {
-             // for this arch, we need to perform the KQ multiplication with F32 precision, otherwise we get NaNs
-             // ref: https://github.com/ggerganov/llama.cpp/pull/4490#issuecomment-1859055847
-             ggml_mul_mat_set_prec(kq, GGML_PREC_F32);

llm/patches/09-lora.diff

@@ -0,0 +1,358 @@
+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
+     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]);
++
++        // try to load as gguf
+         auto adapter = llama_lora_adapter_init(model, lora_adapter.c_str());
+         if (adapter == nullptr) {
+-            fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__);
+-            llama_free(lctx);
+-            llama_free_model(model);
+-            return std::make_tuple(nullptr, nullptr);
++            fprintf(stderr, "%s: error: failed to apply lora adapter, trying ggla\n", __func__);
++
++            // if that fails, try loading as ggla for compatibility
++            int err = llama_model_apply_lora_from_file(model,
++                                                    lora_adapter.c_str(),
++                                                    lora_scale,
++                                                    nullptr,
++                                                    params.n_threads);
++            if (err != 0) {
++                fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__);
++                llama_free(lctx);
++                llama_free_model(model);
++                return std::make_tuple(nullptr, nullptr);
++            }
++        } else {
++            llama_lora_adapter_set(lctx, adapter, lora_scale);
+         }
+-        llama_lora_adapter_set(lctx, adapter, lora_scale);
+     }
+ 
+     if (params.ignore_eos) {
+diff --git a/include/llama.h b/include/llama.h
+index 93fd77ca..b0fb37a6 100644
+--- a/include/llama.h
++++ b/include/llama.h
+@@ -1160,6 +1160,20 @@ extern "C" {
+ 
+     LLAMA_API void llama_dump_timing_info_yaml(FILE * stream, const struct llama_context * ctx);
+ 
++    // Apply a LoRA adapter to a loaded model
++    // path_base_model is the path to a higher quality model to use as a base for
++    // the layers modified by the adapter. Can be NULL to use the current loaded model.
++    // The model needs to be reloaded before applying a new adapter, otherwise the adapter
++    // will be applied on top of the previous one
++    // Returns 0 on success
++    LLAMA_API int32_t llama_model_apply_lora_from_file(
++            const struct llama_model * model,
++                            const char * path_lora,
++                                float   scale,
++                            const char * path_base_model,
++                                int32_t   n_threads);
++
++
+ #ifdef __cplusplus
+ }
+ #endif
+diff --git a/src/llama.cpp b/src/llama.cpp
+index 80a0dd0f..9d7b0e17 100644
+--- a/src/llama.cpp
++++ b/src/llama.cpp
+@@ -21880,3 +21880,290 @@ static void llama_log_callback_default(ggml_log_level level, const char * text,
+     fputs(text, stderr);
+     fflush(stderr);
+ }
++
++static int llama_apply_lora_from_file_internal(
++    const struct llama_model & model, const char * path_lora, float scale, const char * path_base_model, int n_threads
++) {
++    LLAMA_LOG_INFO("%s: applying lora adapter from '%s' - please wait ...\n", __func__, path_lora);
++
++    const int64_t t_start_lora_us = ggml_time_us();
++
++    llama_file fin(path_lora, "rb");
++
++    // verify magic and version
++    {
++        uint32_t magic = fin.read_u32();
++        if (magic != LLAMA_FILE_MAGIC_GGLA) {
++            LLAMA_LOG_ERROR("%s: bad file magic\n", __func__);
++            return 1;
++        }
++
++        uint32_t format_version = fin.read_u32();
++        if (format_version != 1) {
++            LLAMA_LOG_ERROR("%s: unsupported file version\n", __func__ );
++            return 1;
++        }
++    }
++
++    int32_t lora_r = fin.read_u32();
++    int32_t lora_alpha = fin.read_u32();
++    float scaling = scale * (float)lora_alpha / (float)lora_r;
++
++    LLAMA_LOG_INFO("%s: r = %d, alpha = %d, scaling = %.2f\n", __func__, lora_r, lora_alpha, scaling);
++
++    // load base model
++    std::unique_ptr<llama_model_loader> ml;
++    if (path_base_model) {
++        LLAMA_LOG_INFO("%s: loading base model from '%s'\n", __func__, path_base_model);
++        ml.reset(new llama_model_loader(path_base_model, /*use_mmap*/ true, /*check_tensors*/ false, /*kv_overrides*/ nullptr));
++        ml->init_mappings(/*prefetch*/ false); // no prefetching
++    }
++
++    struct tensor_meta {
++        std::string name;
++        ggml_type type;
++        int32_t ne[2];
++        size_t offset;
++    };
++    std::map<std::string, tensor_meta> tensor_meta_map;
++
++    // load all tensor meta
++    while (true) {
++        if (fin.tell() == fin.size) {
++            // eof
++            break;
++        }
++
++        int32_t n_dims;
++        int32_t name_len;
++        int32_t ftype;
++
++        fin.read_raw(&n_dims, sizeof(n_dims));
++        fin.read_raw(&name_len, sizeof(name_len));
++        fin.read_raw(&ftype, sizeof(ftype));
++
++        if (n_dims != 1 && n_dims != 2) {
++            LLAMA_LOG_ERROR("%s: unsupported tensor dimension %d\n", __func__, n_dims);
++            return 1;
++        }
++
++        int32_t ne[2] = { 1, 1 };
++        for (int i = 0; i < n_dims; ++i) {
++            fin.read_raw(&ne[i], sizeof(ne[i]));
++        }
++
++        std::string name;
++        {
++            GGML_ASSERT(name_len < GGML_MAX_NAME);
++            char buf[GGML_MAX_NAME];
++            fin.read_raw(buf, name_len);
++            name = std::string(buf, name_len);
++        }
++
++        // check for lora suffix
++        std::string lora_suffix;
++        if (name.length() > 6) {
++            lora_suffix = name.substr(name.length() - 6);
++        }
++        if (lora_suffix != ".loraA" && lora_suffix != ".loraB") {
++            LLAMA_LOG_ERROR("%s: error: '%s' is not a lora tensor\n", __func__, name.c_str());
++            return 1;
++        }
++
++        // tensor type
++        ggml_type wtype;
++        switch (ftype) {
++            case 0: wtype = GGML_TYPE_F32;  break;
++            case 1: wtype = GGML_TYPE_F16;  break;
++            default:
++                    {
++                        LLAMA_LOG_ERROR("%s: invalid tensor data type '%d'\n",
++                                __func__, ftype);
++                        return 1;
++                    }
++        }
++
++        // data offset
++        size_t offset = fin.tell();
++        offset = (offset + 31) & -32;
++
++        // skip tensor data
++        fin.seek(offset + ggml_row_size(wtype, ne[0]) * ne[1], SEEK_SET);
++
++        tensor_meta_map.emplace(name, tensor_meta{ name, wtype, { ne[0], ne[1] }, offset });
++    }
++
++    bool warned = false;
++    int n_tensors = 0;
++
++    // apply
++    ggml_backend_t backend_cpu = ggml_backend_cpu_init();
++    if (backend_cpu == nullptr) {
++        LLAMA_LOG_ERROR("%s: error: failed to initialize cpu backend\n", __func__);
++        return 1;
++    }
++    ggml_backend_cpu_set_n_threads(backend_cpu, n_threads);
++
++    std::vector<no_init<uint8_t>> read_buf;
++    for (const auto & it : model.tensors_by_name) {
++        const std::string & base_name = it.first;
++        ggml_tensor * model_t = it.second;
++
++        if (tensor_meta_map.find(base_name + ".loraA") == tensor_meta_map.end() ||
++            tensor_meta_map.find(base_name + ".loraB") == tensor_meta_map.end()) {
++            continue;
++        }
++
++        tensor_meta & metaA = tensor_meta_map.at(base_name + ".loraA");
++        tensor_meta & metaB = tensor_meta_map.at(base_name + ".loraB");
++
++        ggml_init_params lora_init_params = {
++            /* .mem_size   */ ggml_tensor_overhead()*128 + ggml_graph_overhead(),
++            /* .mem_buffer */ nullptr,
++            /* .no_alloc   */ true,
++        };
++        ggml_context * lora_ctx = ggml_init(lora_init_params);
++        if (lora_ctx == nullptr) {
++            LLAMA_LOG_ERROR("%s: error: failed to initialize lora context\n", __func__);
++            ggml_backend_free(backend_cpu);
++            return 1;
++        }
++
++        // create tensors
++        ggml_tensor * loraA = ggml_new_tensor_2d(lora_ctx, metaA.type, metaA.ne[0], metaA.ne[1]);
++        ggml_tensor * loraB = ggml_new_tensor_2d(lora_ctx, metaB.type, metaB.ne[0], metaB.ne[1]);
++        ggml_set_name(loraA, metaA.name.c_str());
++        ggml_set_name(loraB, metaB.name.c_str());
++
++        ggml_tensor * base_t;
++        if (ml) {
++            if (!ml->get_tensor_meta(base_name.c_str())) {
++                LLAMA_LOG_ERROR("%s: error: tensor '%s' not found in base model\n", __func__, base_name.c_str());
++                return 1;
++            }
++            base_t = ggml_dup_tensor(lora_ctx, ml->get_tensor_meta(base_name.c_str()));
++        } else {
++            base_t = ggml_dup_tensor(lora_ctx, model_t);
++        }
++        ggml_set_name(base_t, base_name.c_str());
++
++        // allocate in backend buffer
++        ggml_backend_buffer_t lora_buf = ggml_backend_alloc_ctx_tensors_from_buft(lora_ctx, ggml_backend_cpu_buffer_type());
++        if (lora_buf == nullptr) {
++            LLAMA_LOG_ERROR("%s: error: failed to allocate lora tensors\n", __func__);
++            return 1;
++        }
++
++        // load tensor data
++        auto load_tensor = [&read_buf, &fin](const tensor_meta & tensor_meta, ggml_tensor * tensor) {
++            read_buf.resize(ggml_nbytes(tensor));
++            fin.seek(tensor_meta.offset, SEEK_SET);
++            fin.read_raw(read_buf.data(), ggml_nbytes(tensor));
++            ggml_backend_tensor_set(tensor, read_buf.data(), 0, read_buf.size());
++        };
++        load_tensor(metaA, loraA);
++        load_tensor(metaB, loraB);
++
++        // load base model tensor data
++        if (ml) {
++            ml->load_data_for(base_t);
++        } else {
++            ggml_backend_tensor_copy(model_t, base_t);
++        }
++
++        if (ggml_is_quantized(base_t->type) && !warned) {
++            LLAMA_LOG_WARN("%s: warning: using a lora adapter with a quantized model may result in poor quality, "
++                            "use a f16 or f32 base model with --lora-base\n", __func__);
++            warned = true;
++        }
++
++        if (base_t->ne[0] != loraA->ne[1] || base_t->ne[1] != loraB->ne[1]) {
++            LLAMA_LOG_ERROR("%s: incompatible tensor dimensions (%" PRId64 " and %" PRId64 ");"
++                            " are you sure that this adapter is for this model?\n", __func__, base_t->ne[0], loraA->ne[1]);
++            ggml_free(lora_ctx);
++            ggml_backend_buffer_free(lora_buf);
++            ggml_backend_free(backend_cpu);
++            return 1;
++        }
++
++        auto build_lora_graph = [&]() {
++            // w = w + BA*s
++            ggml_tensor * BA = ggml_mul_mat(lora_ctx, loraA, loraB);
++            ggml_set_name(BA, "BA");
++
++            if (scaling != 1.0f) {
++                BA = ggml_scale(lora_ctx, BA, scaling);
++                ggml_set_name(BA, "BA_scaled");
++            }
++
++            ggml_tensor * r;
++            r = ggml_add_inplace(lora_ctx, base_t, BA);
++            ggml_set_name(r, "r_add");
++
++            if (base_t->type != model_t->type) {
++                // convert the result to the model type
++                r = ggml_cast(lora_ctx, r, model_t->type);
++                ggml_set_name(r, "r_cast");
++            }
++
++            return r;
++        };
++
++        ggml_cgraph * gf = ggml_new_graph(lora_ctx);
++        ggml_tensor * r = build_lora_graph();
++        ggml_build_forward_expand(gf, r);
++
++        ggml_backend_buffer_t graph_buf = ggml_backend_alloc_ctx_tensors_from_buft(lora_ctx, ggml_backend_cpu_buffer_type());
++        if (graph_buf == nullptr) {
++            LLAMA_LOG_ERROR("%s: error: failed to allocate graph tensors\n", __func__);
++            ggml_free(lora_ctx);
++            ggml_backend_buffer_free(lora_buf);
++            ggml_backend_free(backend_cpu);
++            return 1;
++        }
++
++        ggml_backend_graph_compute(backend_cpu, gf);
++
++        ggml_backend_tensor_set(model_t, r->data, 0, ggml_nbytes(r));
++
++#if 0
++        // TODO: use scheduler with fallback to CPU for less copies between CPU and GPU
++        //ggml_backend_sched_t sched = ggml_backend_sched_new(backends.data(), backends.size(), GGML_DEFAULT_GRAPH_SIZE);
++
++        // sched compute
++        ggml_build_forward_expand(gf, build_graph());
++        ggml_backend_sched_init_measure(sched, gf);
++
++        // create the graph again, since the previous one was destroyed by the measure
++        ggml_graph_clear(gf);
++        ggml_build_forward_expand(gf, build_graph());
++        ggml_backend_sched_graph_compute(sched, gf);
++        ggml_backend_sched_free(sched);
++#endif
++
++        ggml_backend_buffer_free(lora_buf);
++        ggml_backend_buffer_free(graph_buf);
++        ggml_free(lora_ctx);
++
++        n_tensors++;
++        if (n_tensors % 4 == 0) {
++            LLAMA_LOG_INFO(".");
++        }
++    }
++
++    ggml_backend_free(backend_cpu);
++
++    const int64_t t_lora_us = ggml_time_us() - t_start_lora_us;
++    LLAMA_LOG_INFO(" done (%.2f ms)\n", t_lora_us / 1000.0);
++
++    return 0;
++}
++
++int32_t llama_model_apply_lora_from_file(const struct llama_model * model, const char * path_lora, float scale, const char * path_base_model, int32_t n_threads) {
++    try {
++        return llama_apply_lora_from_file_internal(*model, path_lora, scale, path_base_model, n_threads);
++    } catch (const std::exception & err) {
++        LLAMA_LOG_ERROR("%s: failed to apply lora adapter: %s\n", __func__, err.what());
++        return 1;
++    }
++}
+\ No newline at end of file

llm/patches/10-params.diff

@@ -0,0 +1,20 @@
+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/patches/10-tekken.diff

@@ -1,43 +0,0 @@
-diff --git a/include/llama.h b/include/llama.h
-index bb4b05ba..a92174e0 100644
---- a/include/llama.h
-+++ b/include/llama.h
-@@ -92,6 +92,7 @@ extern "C" {
-         LLAMA_VOCAB_PRE_TYPE_CHATGLM4       = 17,
-         LLAMA_VOCAB_PRE_TYPE_VIKING         = 18,
-         LLAMA_VOCAB_PRE_TYPE_JAIS           = 19,
-+        LLAMA_VOCAB_PRE_TYPE_TEKKEN         = 20,
-     };
- 
-     // note: these values should be synchronized with ggml_rope
-diff --git a/src/llama.cpp b/src/llama.cpp
-index 18364976..435b6fe5 100644
---- a/src/llama.cpp
-+++ b/src/llama.cpp
-@@ -5429,6 +5429,12 @@ static void llm_load_vocab(
-             } else if (
-                 tokenizer_pre == "jais") {
-                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_JAIS;
-+            } else if (
-+                tokenizer_pre == "tekken") {
-+                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_TEKKEN;
-+                vocab.tokenizer_clean_spaces = false;
-+                vocab.tokenizer_ignore_merges = true;
-+                vocab.tokenizer_add_bos = true;
-             } else {
-                 LLAMA_LOG_WARN("%s: missing or unrecognized pre-tokenizer type, using: 'default'\n", __func__);
-                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
-@@ -15448,6 +15454,13 @@ struct llm_tokenizer_bpe {
-                     " ?[^(\\s|.,!?…。，、।۔،)]+",
-                 };
-                 break;
-+            case LLAMA_VOCAB_PRE_TYPE_TEKKEN:
-+                    // original regex from tokenizer.json
-+                    // "[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]*[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]+|[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]+[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]*|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+"
-+                regex_exprs = {
-+                    "[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))*((?=[\\p{L}])([^A-Z]))+|[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))+((?=[\\p{L}])([^A-Z]))*|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
-+                };
-+                break;
-             default:
-                 // default regex for BPE tokenization pre-processing
-                 regex_exprs = {

llm/patches/11-embd_kv.diff

@@ -1,19 +0,0 @@
-diff --git a/src/llama.cpp b/src/llama.cpp
-index 2b9ace28..e60d3d8d 100644
---- a/src/llama.cpp
-+++ b/src/llama.cpp
-@@ -6052,10 +6052,10 @@ static bool llm_load_tensors(
- 
-                         layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
- 
--                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd});
--                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa});
--                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa});
--                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
-+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd,  n_embd_head_k * n_head});
-+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa});
-+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa});
-+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd});
- 
-                         // optional bias tensors
-                         layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd},     llama_model_loader::TENSOR_NOT_REQUIRED);

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) ([][]float32, error)
+	Embed(ctx context.Context, input []string) (*EmbedResponse, error)
 	Tokenize(ctx context.Context, content string) ([]int, error)
 	Detokenize(ctx context.Context, tokens []int) (string, error)
 	Close() error
@@ -346,6 +346,7 @@ 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 {
@@ -417,7 +418,17 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 
 		// reap subprocess when it exits
 		go func() {
-			s.done <- s.cmd.Wait()
+			err := s.cmd.Wait()
+			// Favor a more detailed message over the process exit status
+			if err != nil && s.status != nil && s.status.LastErrMsg != "" {
+				slog.Debug("llama runner terminated", "error", err)
+				if strings.Contains(s.status.LastErrMsg, "unknown model") {
+					s.status.LastErrMsg = "this model is not supported by your version of Ollama. You may need to upgrade"
+				}
+				s.done <- fmt.Errorf(s.status.LastErrMsg)
+			} else {
+				s.done <- err
+			}
 		}()
 
 		return s, nil
@@ -580,14 +591,7 @@ func (s *llmServer) WaitUntilRunning(ctx context.Context) error {
 			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 := ""
+			return fmt.Errorf("llama runner process has terminated: %w", err)
-			if s.status != nil && s.status.LastErrMsg != "" {
-				msg = s.status.LastErrMsg
-			}
-			if strings.Contains(msg, "unknown model") {
-				return fmt.Errorf("this model is not supported by your version of Ollama. You may need to upgrade")
-			}
-			return fmt.Errorf("llama runner process has terminated: %v %s", err, msg)
 		default:
 		}
 		if time.Now().After(stallTimer) {
@@ -723,6 +727,7 @@ 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,
@@ -875,9 +880,10 @@ type EmbedRequest struct {
 
 type EmbedResponse struct {
 	Embedding       [][]float32 `json:"embedding"`
+	PromptEvalCount int         `json:"prompt_n"`
 }
 
-func (s *llmServer) Embed(ctx context.Context, input []string) ([][]float32, error) {
+func (s *llmServer) Embed(ctx context.Context, input []string) (*EmbedResponse, error) {
 	if err := s.sem.Acquire(ctx, 1); err != nil {
 		slog.Error("Failed to acquire semaphore", "error", err)
 		return nil, err
@@ -919,12 +925,12 @@ func (s *llmServer) Embed(ctx context.Context, input []string) ([][]float32, err
 		return nil, fmt.Errorf("%s", body)
 	}
 
-	var embedding EmbedResponse
+	var e EmbedResponse
-	if err := json.Unmarshal(body, &embedding); err != nil {
+	if err := json.Unmarshal(body, &e); err != nil {
 		return nil, fmt.Errorf("unmarshal tokenize response: %w", err)
 	}
 
-	return embedding.Embedding, nil
+	return &e, 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'
+  const command = 'ollama run llama3.1'
 
   return (
     <div className='drag'>

openai/openai.go

@@ -218,6 +218,9 @@ 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

@@ -451,6 +451,7 @@ 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

@@ -198,19 +198,29 @@ 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
             ;;
         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
             ;;
     esac
 
@@ -235,7 +245,11 @@ 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
 
     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, nil)
+	response, err := makeRequest(ctx, http.MethodGet, redirectURL, headers, nil, &registryOptions{})
 	if err != nil {
 		return "", err
 	}

server/download.go

@@ -8,6 +8,7 @@ import (
 	"io"
 	"log/slog"
 	"math"
+	"math/rand/v2"
 	"net/http"
 	"net/url"
 	"os"
@@ -43,7 +44,7 @@ type blobDownload struct {
 
 	context.CancelFunc
 
-	done       bool
+	done       chan struct{}
 	err        error
 	references atomic.Int32
 }
@@ -52,7 +53,9 @@ type blobDownloadPart struct {
 	N         int
 	Offset    int64
 	Size      int64
-	Completed   int64
+	Completed atomic.Int64
+
+	lastUpdatedMu sync.Mutex
 	lastUpdated   time.Time
 
 	*blobDownload `json:"-"`
@@ -71,7 +74,7 @@ func (p *blobDownloadPart) Name() string {
 }
 
 func (p *blobDownloadPart) StartsAt() int64 {
-	return p.Offset + p.Completed
+	return p.Offset + p.Completed.Load()
 }
 
 func (p *blobDownloadPart) StopsAt() int64 {
@@ -81,7 +84,9 @@ 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
 }
 
@@ -91,6 +96,8 @@ 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 {
@@ -98,7 +105,7 @@ func (b *blobDownload) Prepare(ctx context.Context, requestURL *url.URL, opts *r
 		}
 
 		b.Total += part.Size
-		b.Completed.Add(part.Completed)
+		b.Completed.Add(part.Completed.Load())
 		b.Parts = append(b.Parts, part)
 	}
 
@@ -138,9 +145,36 @@ 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)
 }
 
+func newBackoff(maxBackoff time.Duration) func(ctx context.Context) error {
+	var n int
+	return func(ctx context.Context) error {
+		if ctx.Err() != nil {
+			return ctx.Err()
+		}
+
+		n++
+
+		// n^2 backoff timer is a little smoother than the
+		// common choice of 2^n.
+		d := min(time.Duration(n*n)*10*time.Millisecond, maxBackoff)
+		// Randomize the delay between 0.5-1.5 x msec, in order
+		// to prevent accidental "thundering herd" problems.
+		d = time.Duration(float64(d) * (rand.Float64() + 0.5))
+		t := time.NewTimer(d)
+		defer t.Stop()
+		select {
+		case <-ctx.Done():
+			return ctx.Err()
+		case <-t.C:
+			return nil
+		}
+	}
+}
+
 func (b *blobDownload) run(ctx context.Context, requestURL *url.URL, opts *registryOptions) error {
 	defer blobDownloadManager.Delete(b.Digest)
 	ctx, b.CancelFunc = context.WithCancel(ctx)
@@ -153,11 +187,57 @@ func (b *blobDownload) run(ctx context.Context, requestURL *url.URL, opts *regis
 
 	_ = file.Truncate(b.Total)
 
+	directURL, err := func() (*url.URL, error) {
+		ctx, cancel := context.WithTimeout(ctx, 30*time.Second)
+		defer cancel()
+
+		backoff := newBackoff(10 * time.Second)
+		for {
+			// shallow clone opts to be used in the closure
+			// without affecting the outer opts.
+			newOpts := new(registryOptions)
+			*newOpts = *opts
+
+			newOpts.CheckRedirect = func(req *http.Request, via []*http.Request) error {
+				if len(via) > 10 {
+					return errors.New("maxium redirects exceeded (10) for directURL")
+				}
+
+				// if the hostname is the same, allow the redirect
+				if req.URL.Hostname() == requestURL.Hostname() {
+					return nil
+				}
+
+				// stop at the first redirect that is not
+				// the same hostname as the original
+				// request.
+				return http.ErrUseLastResponse
+			}
+
+			resp, err := makeRequestWithRetry(ctx, http.MethodGet, requestURL, nil, nil, newOpts)
+			if err != nil {
+				slog.Warn("failed to get direct URL; backing off and retrying", "err", err)
+				if err := backoff(ctx); err != nil {
+					return nil, err
+				}
+				continue
+			}
+			defer resp.Body.Close()
+			if resp.StatusCode != http.StatusTemporaryRedirect {
+				return nil, fmt.Errorf("unexpected status code %d", resp.StatusCode)
+			}
+			return resp.Location()
+		}
+	}()
+	if err != nil {
+		return err
+	}
+
 	g, inner := errgroup.WithContext(ctx)
 	g.SetLimit(numDownloadParts)
 	for i := range b.Parts {
 		part := b.Parts[i]
-		if part.Completed == part.Size {
+		if part.Completed.Load() == part.Size {
 			continue
 		}
 
@@ -165,7 +245,7 @@ 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, requestURL, w, part, opts)
+				err = b.downloadChunk(inner, directURL, w, part)
 				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
@@ -206,29 +286,31 @@ 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, opts *registryOptions) error {
+func (b *blobDownload) downloadChunk(ctx context.Context, requestURL *url.URL, w io.Writer, part *blobDownloadPart) error {
 	g, ctx := errgroup.WithContext(ctx)
 	g.Go(func() error {
-		headers := make(http.Header)
+		req, err := http.NewRequestWithContext(ctx, http.MethodGet, requestURL.String(), nil)
-		headers.Set("Range", fmt.Sprintf("bytes=%d-%d", part.StartsAt(), part.StopsAt()-1))
+		if err != nil {
-		resp, err := makeRequestWithRetry(ctx, http.MethodGet, requestURL, headers, nil, opts)
+			return err
+		}
+		req.Header.Set("Range", fmt.Sprintf("bytes=%d-%d", part.StartsAt(), part.StopsAt()-1))
+		resp, err := http.DefaultClient.Do(req)
 		if err != nil {
 			return err
 		}
 		defer resp.Body.Close()
 
-		n, err := io.CopyN(w, io.TeeReader(resp.Body, part), part.Size-part.Completed)
+		n, err := io.CopyN(w, io.TeeReader(resp.Body, part), part.Size-part.Completed.Load())
 		if err != nil && !errors.Is(err, context.Canceled) && !errors.Is(err, io.ErrUnexpectedEOF) {
 			// rollback progress
 			b.Completed.Add(-n)
 			return err
 		}
 
-		part.Completed += n
+		part.Completed.Add(n)
 		if err := b.writePart(part.Name(), part); err != nil {
 			return err
 		}
@@ -242,15 +324,21 @@ func (b *blobDownload) downloadChunk(ctx context.Context, requestURL *url.URL, w
 		for {
 			select {
 			case <-ticker.C:
-				if part.Completed >= part.Size {
+				if part.Completed.Load() >= part.Size {
 					return nil
 				}
 
-				if !part.lastUpdated.IsZero() && time.Since(part.lastUpdated) > 5*time.Second {
+				part.lastUpdatedMu.Lock()
+				lastUpdated := part.lastUpdated
+				part.lastUpdatedMu.Unlock()
+
+				if !lastUpdated.IsZero() && time.Since(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():
@@ -315,6 +403,8 @@ 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]),
@@ -322,10 +412,6 @@ 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/images.go

@@ -54,6 +54,8 @@ type registryOptions struct {
 	Username string
 	Password string
 	Token    string
+
+	CheckRedirect func(req *http.Request, via []*http.Request) error
 }
 
 type Model struct {
@@ -1131,7 +1133,9 @@ func makeRequest(ctx context.Context, method string, requestURL *url.URL, header
 		req.ContentLength = contentLength
 	}
 
-	resp, err := http.DefaultClient.Do(req)
+	resp, err := (&http.Client{
+		CheckRedirect: regOpts.CheckRedirect,
+	}).Do(req)
 	if err != nil {
 		return nil, err
 	}

server/model.go

@@ -263,13 +263,27 @@ func detectChatTemplate(layers []*layerGGML) ([]*layerGGML, error) {
 			if t, err := template.Named(s); err != nil {
 				slog.Debug("template detection", "error", err)
 			} else {
-				tmpl, err := NewLayer(t.Reader(), "application/vnd.ollama.image.template")
+				layer, err := NewLayer(t.Reader(), "application/vnd.ollama.image.template")
 				if err != nil {
 					return nil, err
 				}
 
-				tmpl.status = fmt.Sprintf("using autodetected template %s", t.Name)
+				layer.status = fmt.Sprintf("using autodetected template %s", t.Name)
-				layers = append(layers, &layerGGML{tmpl, nil})
+				layers = append(layers, &layerGGML{layer, nil})
+
+				if t.Parameters != nil {
+					var b bytes.Buffer
+					if err := json.NewEncoder(&b).Encode(t.Parameters); err != nil {
+						return nil, err
+					}
+
+					layer, err := NewLayer(&b, "application/vnd.ollama.image.params")
+					if err != nil {
+						return nil, err
+					}
+
+					layers = append(layers, &layerGGML{layer, nil})
+				}
 			}
 		}
 	}
@@ -344,6 +358,10 @@ func (m *Model) parseToolCalls(s string) ([]api.ToolCall, bool) {
 		}
 	}
 
+	if name == "" || arguments == "" {
+		return nil, false
+	}
+
 	var objs []map[string]any
 	for offset := 0; offset < len(s); {
 		var obj map[string]any
@@ -361,24 +379,41 @@ func (m *Model) parseToolCalls(s string) ([]api.ToolCall, bool) {
 			return nil, false
 		} else {
 			offset += int(decoder.InputOffset())
-			objs = append(objs, obj)
+
+			// collect all nested objects
+			var collect func(any) []map[string]any
+			collect = func(obj any) (all []map[string]any) {
+				switch o := obj.(type) {
+				case map[string]any:
+					all = append(all, o)
+					for _, v := range o {
+						all = append(all, collect(v)...)
+					}
+				case []any:
+					for _, v := range o {
+						all = append(all, collect(v)...)
+					}
+				}
+
+				return all
+			}
+			objs = append(objs, collect(obj)...)
 		}
 	}
 
 	var toolCalls []api.ToolCall
 	for _, kv := range objs {
-		var call api.ToolCall
+		n, nok := kv[name].(string)
-		for k, v := range kv {
+		a, aok := kv[arguments].(map[string]any)
-			switch k {
+		if nok && aok {
-			case name:
+			toolCalls = append(toolCalls, api.ToolCall{
-				call.Function.Name = v.(string)
+				Function: api.ToolCallFunction{
-			case arguments:
+					Name:      n,
-				call.Function.Arguments = v.(map[string]any)
+					Arguments: a,
+				},
+			})
 		}
 	}
 
-		toolCalls = append(toolCalls, call)
-	}
-
 	return toolCalls, len(toolCalls) > 0
 }

server/model_test.go

@@ -166,6 +166,7 @@ The temperature in San Francisco, CA is 70°F and in Toronto, Canada is 20°C.`,
 {"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}}
 {"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}}
 </tool_call>`, true},
+		{"xlam", `{"tool_calls": [{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}},{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}}]}`, true},
 	}
 
 	var tools []api.Tool

server/routes.go

@@ -284,6 +284,7 @@ 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 {
@@ -332,6 +333,8 @@ 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()})
@@ -370,13 +373,16 @@ func (s *Server) EmbedHandler(c *gin.Context) {
 		return
 	}
 
-	for i, e := range embeddings {
+	for i, e := range embeddings.Embedding {
-		embeddings[i] = normalize(e)
+		embeddings.Embedding[i] = normalize(e)
 	}
 
 	resp := api.EmbedResponse{
 		Model:           req.Model,
-		Embeddings: embeddings,
+		Embeddings:      embeddings.Embedding,
+		TotalDuration:   time.Since(checkpointStart),
+		LoadDuration:    checkpointLoaded.Sub(checkpointStart),
+		PromptEvalCount: embeddings.PromptEvalCount,
 	}
 	c.JSON(http.StatusOK, resp)
 }
@@ -428,9 +434,9 @@ func (s *Server) EmbeddingsHandler(c *gin.Context) {
 		return
 	}
 
-	embedding := make([]float64, len(embeddings[0]))
+	embedding := make([]float64, len(embeddings.Embedding[0]))
 
-	for i, v := range embeddings[0] {
+	for i, v := range embeddings.Embedding[0] {
 		embedding[i] = float64(v)
 	}
 
@@ -609,10 +615,9 @@ func (s *Server) CreateModelHandler(c *gin.Context) {
 		defer cancel()
 
 		quantization := cmp.Or(r.Quantize, r.Quantization)
-		if err := CreateModel(ctx, name, filepath.Dir(r.Path), strings.ToUpper(quantization), f, fn); err != nil {
+		if err := CreateModel(ctx, name, filepath.Dir(r.Path), strings.ToUpper(quantization), f, fn); errors.Is(err, errBadTemplate) {
-			if errors.Is(err, errBadTemplate) {
 			ch <- gin.H{"error": err.Error(), "status": http.StatusBadRequest}
-			}
+		} else if err != nil {
 			ch <- gin.H{"error": err.Error()}
 		}
 	}()

server/routes_create_test.go

@@ -599,9 +599,10 @@ func TestCreateDetectTemplate(t *testing.T) {
 		}
 
 		checkFileExists(t, filepath.Join(p, "blobs", "*"), []string{
+			filepath.Join(p, "blobs", "sha256-0d79f567714c62c048378f2107fb332dabee0135d080c302d884317da9433cc5"),
 			filepath.Join(p, "blobs", "sha256-553c4a3f747b3d22a4946875f1cc8ed011c2930d83f864a0c7265f9ec0a20413"),
 			filepath.Join(p, "blobs", "sha256-c608dc615584cd20d9d830363dabf8a4783ae5d34245c3d8c115edb3bc7b28e4"),
-			filepath.Join(p, "blobs", "sha256-f836ee110db21567f826332e4cedd746c06d10664fd5a9ea3659e3683a944510"),
+			filepath.Join(p, "blobs", "sha256-ea34c57ba5b78b740aafe2aeb74dc6507fc3ad14170b64c26a04fb9e36c88d75"),
 		})
 	})
 

server/sched.go

@@ -212,9 +212,12 @@ 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 := pickBestFitGPUs(pending, ggml, gpus, &numParallel)
+						g := pickBestFullFitByLibrary(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
@@ -231,7 +234,7 @@ func (s *Scheduler) processPending(ctx context.Context) {
 
 						// Update free memory from currently loaded models
 						s.updateFreeSpace(availGpus)
-						fitGpus := pickBestFitGPUs(pending, ggml, availGpus, &numParallel)
+						fitGpus := pickBestFullFitByLibrary(pending, ggml, availGpus, &numParallel)
 						if fitGpus != nil {
 							slog.Debug("new model fits with existing models, loading")
 							s.loadFn(pending, ggml, fitGpus, numParallel)
@@ -668,11 +671,12 @@ 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 }
 
-// pickBestFitGPUs will try to find the optimal placement of the model in the available GPUs where the model fully fits
+// 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)
 // 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 pickBestFitGPUs(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel *int) gpu.GpuInfoList {
+func pickBestFullFitByLibrary(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel *int) gpu.GpuInfoList {
 	var estimatedVRAM uint64
 
 	var numParallelToTry []int
@@ -723,6 +727,25 @@ func pickBestFitGPUs(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numP
 	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

@@ -666,11 +666,50 @@ 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          [][]float32
+	embedResp          *llm.EmbedResponse
 	embedRespErr       error
 	tokenizeResp       []int
 	tokenizeRespErr    error
@@ -688,7 +727,7 @@ 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) ([][]float32, error) {
+func (s *mockLlm) Embed(ctx context.Context, input []string) (*llm.EmbedResponse, error) {
 	return s.embedResp, s.embedRespErr
 }
 func (s *mockLlm) Tokenize(ctx context.Context, content string) ([]int, error) {

server/testdata/tools/xlam.gotmpl

@@ -0,0 +1,45 @@
+{{- if .System }}{{ .System }}
+{{ end }}
+{{- range $i, $_ := .Messages }}
+{{- if eq .Role "user" }}### Instruction:
+{{- if and $.Tools (le (len (slice $.Messages $i)) 2) }}
+[BEGIN OF TASK INSTRUCTION]
+You are an expert in composing functions. You are given a question and a set of possible functions. 
+Based on the question, you will need to make one or more function/tool calls to achieve the purpose. 
+If none of the functions can be used, point it out and refuse to answer. 
+If the given question lacks the parameters required by the function, also point it out.
+[END OF TASK INSTRUCTION]
+
+[BEGIN OF AVAILABLE TOOLS]
+{{ $.Tools }}
+[END OF AVAILABLE TOOLS]
+
+[BEGIN OF FORMAT INSTRUCTION]
+The output MUST strictly adhere to the following JSON format, and NO other text MUST be included.
+The example format is as follows. Please make sure the parameter type is correct. If no function call is needed, please make tool_calls an empty list '[]'.
+```
+{
+    "tool_calls": [
+    {"name": "func_name1", "arguments": {"argument1": "value1", "argument2": "value2"}},
+    ... (more tool calls as required)
+    ]
+}
+```
+[END OF FORMAT INSTRUCTION]
+
+[BEGIN OF QUERY]
+{{ .Content }}
+[END OF QUERY]
+
+
+{{ else }}
+{{ .Content }}
+{{ end }}
+{{- else if .ToolCalls }}### Response:
+{"tool_calls": [{{ range .ToolCalls }}{"name": "{{ .Function.Name }}", "arguments": {{ .Function.Arguments }}}{{ end }}]}
+<|EOT|>
+{{ else if eq .Role "assistant" }}### Response:
+{{ .Content }}
+<|EOT|>
+{{ end }}
+{{- end }}### Response:

server/testdata/tools/xlam.out

@@ -0,0 +1,40 @@
+You are a knowledgable assistant. You can answer questions and perform tasks.
+### Instruction:
+What's the weather like today in Paris?
+### Response:
+{"tool_calls": [{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Paris, France"}}]}
+<|EOT|>
+### Response:
+The current temperature in Paris, France is 22 degrees Celsius.
+<|EOT|>
+### Instruction:
+[BEGIN OF TASK INSTRUCTION]
+You are an expert in composing functions. You are given a question and a set of possible functions. 
+Based on the question, you will need to make one or more function/tool calls to achieve the purpose. 
+If none of the functions can be used, point it out and refuse to answer. 
+If the given question lacks the parameters required by the function, also point it out.
+[END OF TASK INSTRUCTION]
+
+[BEGIN OF AVAILABLE TOOLS]
+[{"type":"function","function":{"name":"get_current_weather","description":"Get the current weather","parameters":{"type":"object","required":["location","format"],"properties":{"format":{"type":"string","description":"The temperature unit to use. Infer this from the users location.","enum":["celsius","fahrenheit"]},"location":{"type":"string","description":"The city and state, e.g. San Francisco, CA"}}}}}]
+[END OF AVAILABLE TOOLS]
+
+[BEGIN OF FORMAT INSTRUCTION]
+The output MUST strictly adhere to the following JSON format, and NO other text MUST be included.
+The example format is as follows. Please make sure the parameter type is correct. If no function call is needed, please make tool_calls an empty list '[]'.
+```
+{
+    "tool_calls": [
+    {"name": "func_name1", "arguments": {"argument1": "value1", "argument2": "value2"}},
+    ... (more tool calls as required)
+    ]
+}
+```
+[END OF FORMAT INSTRUCTION]
+
+[BEGIN OF QUERY]
+What's the weather like today in San Francisco and Toronto?
+[END OF QUERY]
+
+
+### Response:

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, nil)
+			err = b.uploadPart(ctx, http.MethodPut, redirectURL, part, &registryOptions{})
 			switch {
 			case errors.Is(err, context.Canceled):
 				return err

template/alfred.json

@@ -0,0 +1,8 @@
+{
+  "stop": [
+    "<start_system>",
+    "<end_message>",
+    "<start_user>",
+    "<start_assistant>"
+  ]
+}

template/alpaca.json

@@ -0,0 +1,6 @@
+{
+  "stop": [
+    "### Instruction:",
+    "### Response"
+  ]
+}

template/chatml.json

@@ -0,0 +1,6 @@
+{
+  "stop": [
+    "<|im_start|>",
+    "<|im_end|>"
+  ]
+}

template/chatqa.json

@@ -0,0 +1,8 @@
+{
+  "stop": [
+    "System:",
+    "User:",
+    "Assistant:",
+    "<|begin_of_text|>"
+  ]
+}

template/codellama-70b-instruct.json

@@ -0,0 +1,7 @@
+{
+  "stop": [
+    "Source:",
+    "Destination:",
+    "<step>"
+  ]
+}

template/falcon-instruct.json

@@ -0,0 +1,6 @@
+{
+  "stop": [
+    "User:",
+    "Assistant:"
+  ]
+}

template/gemma-instruct.json

@@ -0,0 +1,6 @@
+{
+  "stop": [
+    "<start_of_turn>",
+    "<end_of_turn>"
+  ]
+}

template/granite-instruct.json

@@ -0,0 +1,7 @@
+{
+  "stop": [
+    "System:",
+    "Question:",
+    "Answer:"
+  ]
+}

template/llama2-chat.json

@@ -0,0 +1,8 @@
+{
+  "stop": [
+    "[INST]",
+    "[/INST]",
+    "<<SYS>>",
+    "<</SYS>>"
+  ]
+}

template/llama3-instruct.json

@@ -0,0 +1,7 @@
+{
+  "stop": [
+    "<|start_header_id|>",
+    "<|end_header_id|>",
+    "<|eot_id|>"
+  ]
+}

template/magicoder.json

@@ -0,0 +1,6 @@
+{
+  "stop": [
+    "@@ Instruction",
+    "@@ Response"
+  ]
+}

template/mistral-instruct.json

@@ -0,0 +1,6 @@
+{
+  "stop": [
+    "<|im_start|>",
+    "<|im_end|>"
+  ]
+}

template/openchat.json

@@ -0,0 +1,5 @@
+{
+  "stop": [
+    "<|end_of_turn|>"
+  ]
+}

template/phi-3.json

@@ -0,0 +1,8 @@
+{
+  "stop": [
+    "<|end|>",
+    "<|system|>",
+    "<|user|>",
+    "<|assistant|>"
+  ]
+}

template/solar-instruct.json

@@ -0,0 +1,7 @@
+{
+  "stop": [
+    "### System:",
+    "### User:",
+    "### Assistant"
+  ]
+}

template/starcoder2-instruct.json

@@ -0,0 +1,7 @@
+{
+  "stop": [
+    "### Instruction",
+    "### Response",
+    "<|endoftext|>"
+  ]
+}

template/template.go

@@ -23,6 +23,7 @@ import (
 var indexBytes []byte
 
 //go:embed *.gotmpl
+//go:embed *.json
 var templatesFS embed.FS
 
 var templatesOnce = sync.OnceValues(func() ([]*named, error) {
@@ -39,6 +40,15 @@ var templatesOnce = sync.OnceValues(func() ([]*named, error) {
 
 		// normalize line endings
 		t.Bytes = bytes.ReplaceAll(bts, []byte("\r\n"), []byte("\n"))
+
+		params, err := templatesFS.ReadFile(t.Name + ".json")
+		if err != nil {
+			continue
+		}
+
+		if err := json.Unmarshal(params, &t.Parameters); err != nil {
+			return nil, err
+		}
 	}
 
 	return templates, nil
@@ -48,6 +58,10 @@ type named struct {
 	Name     string `json:"name"`
 	Template string `json:"template"`
 	Bytes    []byte
+
+	Parameters *struct {
+		Stop []string `json:"stop"`
+	}
 }
 
 func (t named) Reader() io.Reader {

template/vicuna.json

@@ -0,0 +1,6 @@
+{
+  "stop": [
+    "USER:",
+    "ASSISTANT:"
+  ]
+}

template/zephyr.json

@@ -0,0 +1,8 @@
+{
+  "stop": [
+    "<|system|>",
+    "</s>",
+    "<|user|>",
+    "<|assistant|>"
+  ]
+}