num parallel embed

autodetect stop parameters from template

.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

Dockerfile

@@ -1,4 +1,4 @@
-ARG GOLANG_VERSION=1.22.1
+ARG GOLANG_VERSION=1.22.5
 ARG CMAKE_VERSION=3.22.1
 # this CUDA_VERSION corresponds with the one specified in docs/gpu.md
 ARG CUDA_VERSION=11.3.1

README.md

@@ -64,7 +64,8 @@ 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
 
@@ -296,6 +297,7 @@ 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
 

cmd/cmd.go

@@ -1344,7 +1344,6 @@ func NewCLI() *cobra.Command {
 				envVars["OLLAMA_TMPDIR"],
 				envVars["OLLAMA_FLASH_ATTENTION"],
 				envVars["OLLAMA_LLM_LIBRARY"],
-				envVars["OLLAMA_MAX_VRAM"],
 			})
 		default:
 			appendEnvDocs(cmd, envs)

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
@@ -380,12 +418,14 @@ 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):
 
@@ -622,6 +662,79 @@ 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 +1139,7 @@ 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 +1147,11 @@ 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 +1160,9 @@ 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 +1170,35 @@ curl http://localhost:11434/api/embeddings -d '{
 
 ```json
 {
-  "embedding": [
-    0.5670403838157654, 0.009260174818336964, 0.23178744316101074, -0.2916173040866852, -0.8924556970596313,
-    0.8785552978515625, -0.34576427936553955, 0.5742510557174683, -0.04222835972905159, -0.137906014919281
-  ]
+  "model": "all-minilm",
+  "embeddings": [[
+    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 +1245,45 @@ 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/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.
 

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`
-- [ ] `tools`
+- [x] `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 }}
+```

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 != "" {

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 +66,8 @@ func TestAllMiniLMEmbed(t *testing.T) {
 		t.Fatalf("expected 384 floats, got %d", len(res.Embeddings[0]))
 	}
 
-	if res.Embeddings[0][0] != 0.010071031 {
-		t.Fatalf("expected 0.010071031, got %f", res.Embeddings[0][0])
+	if !floatsEqual32(res.Embeddings[0][0], 0.010071031) {
+		t.Fatalf("expected 0.010071031, got %.8f", res.Embeddings[0][0])
 	}
 }
 
@@ -61,12 +94,12 @@ 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 {
-		t.Fatalf("expected 0.010071031 and -0.009802706, got %f and %f", res.Embeddings[0][0], res.Embeddings[1][0])
+	if !floatsEqual32(res.Embeddings[0][0], 0.010071031) || !floatsEqual32(res.Embeddings[1][0], -0.009802706) {
+		t.Fatalf("expected 0.010071031 and -0.009802706, got %.8f and %.8f", res.Embeddings[0][0], res.Embeddings[1][0])
 	}
 }
 
-func TestAllMiniLmEmbedTruncate(t *testing.T) {
+func TestAllMiniLMEmbedTruncate(t *testing.T) {
 	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
 	defer cancel()
 
@@ -135,6 +168,22 @@ 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/server.go

@@ -417,7 +417,17 @@ 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 +590,7 @@ 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 := ""
-			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)
+			return fmt.Errorf("llama runner process has terminated: %w", err)
 		default:
 		}
 		if time.Now().After(stallTimer) {

server/download.go

@@ -8,6 +8,7 @@ import (
 	"io"
 	"log/slog"
 	"math"
+	"math/rand/v2"
 	"net/http"
 	"net/url"
 	"os"
@@ -141,6 +142,32 @@ func (b *blobDownload) Run(ctx context.Context, requestURL *url.URL, opts *regis
 	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)
@@ -153,6 +180,52 @@ 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 {
@@ -165,7 +238,7 @@ func (b *blobDownload) run(ctx context.Context, requestURL *url.URL, opts *regis
 			var err error
 			for try := 0; try < maxRetries; try++ {
 				w := io.NewOffsetWriter(file, part.StartsAt())
-				err = b.downloadChunk(inner, requestURL, w, part, opts)
+				err = b.downloadChunk(inner, directURL, w, part, opts)
 				switch {
 				case errors.Is(err, context.Canceled), errors.Is(err, syscall.ENOSPC):
 					// return immediately if the context is canceled or the device is out of space

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)
-				layers = append(layers, &layerGGML{tmpl, nil})
+				layer.status = fmt.Sprintf("using autodetected template %s", t.Name)
+				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})
+				}
 			}
 		}
 	}

server/routes.go

@@ -609,10 +609,9 @@ 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 errors.Is(err, errBadTemplate) {
-				ch <- gin.H{"error": err.Error(), "status": http.StatusBadRequest}
-			}
+		if err := CreateModel(ctx, name, filepath.Dir(r.Path), strings.ToUpper(quantization), f, fn); 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

@@ -132,6 +132,8 @@ func (s *Scheduler) processPending(ctx context.Context) {
 			if len(pending.model.ProjectorPaths) > 0 && numParallel != 1 {
 				numParallel = 1
 				slog.Warn("multimodal models don't support parallel requests yet")
+			} else if strings.Contains(pending.model.Config.ModelFamily, "bert") {
+				numParallel = runtime.NumCPU()
 			}
 
 			for {

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|>"
+  ]
+}