Merge branch 'ollama:main' into main

Author: likelovewant

.github/workflows/test.yaml

@@ -243,7 +243,7 @@ jobs:
           $env:PATH="$gopath;$gccpath;$env:PATH"
           echo $env:PATH
           if (!(gcc --version | select-string -quiet clang)) { throw "wrong gcc compiler detected - must be clang" }
-          make -j 4      
+          make -j 4
       - name: 'Build Unix Go Runners'
         if: ${{ ! startsWith(matrix.os, 'windows-') }}
         run: make -j 4
@@ -310,8 +310,7 @@ jobs:
             arm64) echo ARCH=arm64 ;;
           esac >>$GITHUB_ENV
         shell: bash
-      - run: go build
-      - run: go test -v ./...
+      - run: go test ./...
 
   patches:
     needs: [changes]
@@ -323,4 +322,4 @@ jobs:
           submodules: recursive
       - name: Verify patches carry all the changes
         run: |
-          make apply-patches sync && git diff --compact-summary --exit-code llama
+          make apply-patches sync && git diff --compact-summary --exit-code llama

README.md

@@ -67,12 +67,12 @@ Here are some example models that can be downloaded:
 
 | Model              | Parameters | Size  | Download                         |
 | ------------------ | ---------- | ----- | -------------------------------- |
+| Llama 3.3          | 70B        | 43GB  | `ollama run llama3.3`            |
 | Llama 3.2          | 3B         | 2.0GB | `ollama run llama3.2`            |
 | Llama 3.2          | 1B         | 1.3GB | `ollama run llama3.2:1b`         |
 | Llama 3.2 Vision   | 11B        | 7.9GB | `ollama run llama3.2-vision`     |
 | Llama 3.2 Vision   | 90B        | 55GB  | `ollama run llama3.2-vision:90b` |
 | Llama 3.1          | 8B         | 4.7GB | `ollama run llama3.1`            |
-| Llama 3.1          | 70B        | 40GB  | `ollama run llama3.1:70b`        |
 | Llama 3.1          | 405B       | 231GB | `ollama run llama3.1:405b`       |
 | Phi 3 Mini         | 3.8B       | 2.3GB | `ollama run phi3`                |
 | Phi 3 Medium       | 14B        | 7.9GB | `ollama run phi3:medium`         |

api/types.go

@@ -67,7 +67,7 @@ type GenerateRequest struct {
 	Raw bool `json:"raw,omitempty"`
 
 	// Format specifies the format to return a response in.
-	Format string `json:"format"`
+	Format json.RawMessage `json:"format,omitempty"`
 
 	// KeepAlive controls how long the model will stay loaded in memory following
 	// this request.
@@ -94,7 +94,7 @@ type ChatRequest struct {
 	Stream *bool `json:"stream,omitempty"`
 
 	// Format is the format to return the response in (e.g. "json").
-	Format string `json:"format"`
+	Format json.RawMessage `json:"format,omitempty"`
 
 	// KeepAlive controls how long the model will stay loaded into memory
 	// following the request.

cmd/cmd.go

@@ -8,6 +8,7 @@ import (
 	"crypto/ed25519"
 	"crypto/rand"
 	"crypto/sha256"
+	"encoding/json"
 	"encoding/pem"
 	"errors"
 	"fmt"
@@ -1035,10 +1036,14 @@ func chat(cmd *cobra.Command, opts runOptions) (*api.Message, error) {
 		return nil
 	}
 
+	if opts.Format == "json" {
+		opts.Format = `"` + opts.Format + `"`
+	}
+
 	req := &api.ChatRequest{
 		Model:    opts.Model,
 		Messages: opts.Messages,
-		Format:   opts.Format,
+		Format:   json.RawMessage(opts.Format),
 		Options:  opts.Options,
 	}
 
@@ -1120,12 +1125,16 @@ func generate(cmd *cobra.Command, opts runOptions) error {
 		}
 	}
 
+	if opts.Format == "json" {
+		opts.Format = `"` + opts.Format + `"`
+	}
+
 	request := api.GenerateRequest{
 		Model:     opts.Model,
 		Prompt:    opts.Prompt,
 		Context:   generateContext,
 		Images:    opts.Images,
-		Format:    opts.Format,
+		Format:    json.RawMessage(opts.Format),
 		System:    opts.System,
 		Options:   opts.Options,
 		KeepAlive: opts.KeepAlive,
@@ -1445,6 +1454,7 @@ func NewCLI() *cobra.Command {
 				envVars["OLLAMA_SCHED_SPREAD"],
 				envVars["OLLAMA_TMPDIR"],
 				envVars["OLLAMA_FLASH_ATTENTION"],
+				envVars["OLLAMA_KV_CACHE_TYPE"],
 				envVars["OLLAMA_LLM_LIBRARY"],
 				envVars["OLLAMA_GPU_OVERHEAD"],
 				envVars["OLLAMA_LOAD_TIMEOUT"],

convert/tokenizer.go

@@ -10,6 +10,7 @@ import (
 	"log/slog"
 	"os"
 	"slices"
+	"strings"
 
 	"golang.org/x/exp/maps"
 )
@@ -60,7 +61,25 @@ func parseTokenizer(fsys fs.FS, specialTokenTypes []string) (*Tokenizer, error)
 			addedTokens[t.Content] = t
 		}
 
-		t.Merges = tt.Model.Merges
+		if len(tt.Model.Merges) == 0 {
+			// noop; merges is empty
+		} else if err := json.Unmarshal(tt.Model.Merges, &t.Merges); err == nil {
+			// noop; merges is []string
+		} else if merges, err := func() ([][]string, error) {
+			var merges [][]string
+			if err := json.Unmarshal(tt.Model.Merges, &merges); err != nil {
+				return nil, err
+			}
+
+			return merges, nil
+		}(); err == nil {
+			t.Merges = make([]string, len(merges))
+			for i := range merges {
+				t.Merges[i] = strings.Join(merges[i], " ")
+			}
+		} else {
+			return nil, fmt.Errorf("could not parse tokenizer merges. expected []string or [][]string: %w", err)
+		}
 
 		sha256sum := sha256.New()
 		for _, pt := range tt.PreTokenizer.PreTokenizers {
@@ -156,9 +175,9 @@ func parseTokenizer(fsys fs.FS, specialTokenTypes []string) (*Tokenizer, error)
 type tokenizer struct {
 	AddedTokens []token `json:"added_tokens"`
 	Model       struct {
-		Type   string         `json:"type"`
-		Vocab  map[string]int `json:"vocab"`
-		Merges []string       `json:"merges"`
+		Type   string          `json:"type"`
+		Vocab  map[string]int  `json:"vocab"`
+		Merges json.RawMessage `json:"merges"`
 	} `json:"model"`
 
 	PreTokenizer struct {

convert/tokenizer_test.go

@@ -191,6 +191,62 @@ func TestParseTokenizer(t *testing.T) {
 				Pre: "default",
 			},
 		},
+		{
+			name: "list string merges",
+			fsys: createTokenizerFS(t, t.TempDir(), map[string]io.Reader{
+				"tokenizer.json": strings.NewReader(`{
+					"model": {
+						"merges": [
+							"a b",
+							"c d",
+							"e f"
+						]
+					}
+				}`),
+			}),
+			want: &Tokenizer{
+				Vocabulary: &Vocabulary{
+					Model: "gpt2",
+				},
+				Merges: []string{
+					"a b",
+					"c d",
+					"e f",
+				},
+				Pre: "default",
+			},
+		},
+		{
+			name: "list list string merges",
+			fsys: createTokenizerFS(t, t.TempDir(), map[string]io.Reader{
+				"tokenizer.json": strings.NewReader(`{
+					"model": {
+						"merges": [
+							[
+								"a", "b"
+							],
+							[
+								"c", "d"
+							],
+							[
+								"e", "f"
+							]
+						]
+					}
+				}`),
+			}),
+			want: &Tokenizer{
+				Vocabulary: &Vocabulary{
+					Model: "gpt2",
+				},
+				Merges: []string{
+					"a b",
+					"c d",
+					"e f",
+				},
+				Pre: "default",
+			},
+		},
 	}
 
 	for _, tt := range cases {

discover/types.go

@@ -183,3 +183,17 @@ func (si SystemInfo) GetOptimalThreadCount() int {
 
 	return coreCount
 }
+
+// For each GPU, check if it does NOT support flash attention
+func (l GpuInfoList) FlashAttentionSupported() bool {
+	for _, gpu := range l {
+		supportsFA := gpu.Library == "metal" ||
+			(gpu.Library == "cuda" && gpu.DriverMajor >= 7) ||
+			gpu.Library == "rocm"
+
+		if !supportsFA {
+			return false
+		}
+	}
+	return true
+}

docs/api.md

@@ -45,7 +45,7 @@ Generate a response for a given prompt with a provided model. This is a streamin
 
 Advanced parameters (optional):
 
-- `format`: the format to return a response in. Currently the only accepted value is `json`
+- `format`: the format to return a response in. Format can be `json` or a JSON schema
 - `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
 - `system`: system message to (overrides what is defined in the `Modelfile`)
 - `template`: the prompt template to use (overrides what is defined in the `Modelfile`)
@@ -54,6 +54,10 @@ Advanced parameters (optional):
 - `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
 - `context` (deprecated): the context parameter returned from a previous request to `/generate`, this can be used to keep a short conversational memory
 
+#### Structured outputs
+
+Structured outputs are supported by providing a JSON schema in the `format` parameter. The model will generate a response that matches the schema. See the [structured outputs](#request-structured-outputs) example below.
+
 #### JSON mode
 
 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.
@@ -185,6 +189,52 @@ curl http://localhost:11434/api/generate -d '{
 }
 ```
 
+#### Request (Structured outputs)
+
+##### Request
+
+```shell
+curl -X POST http://localhost:11434/api/generate -H "Content-Type: application/json" -d '{
+  "model": "llama3.1:8b",
+  "prompt": "Ollama is 22 years old and is busy saving the world. Respond using JSON",
+  "stream": false,
+  "format": {
+    "type": "object",
+    "properties": {
+      "age": {
+        "type": "integer"
+      },
+      "available": {
+        "type": "boolean"
+      }
+    },
+    "required": [
+      "age",
+      "available"
+    ]
+  }
+}'
+```
+
+##### Response
+
+```json
+{
+  "model": "llama3.1:8b",
+  "created_at": "2024-12-06T00:48:09.983619Z",
+  "response": "{\n  \"age\": 22,\n  \"available\": true\n}",
+  "done": true,
+  "done_reason": "stop",
+  "context": [1, 2, 3],
+  "total_duration": 1075509083,
+  "load_duration": 567678166,
+  "prompt_eval_count": 28,
+  "prompt_eval_duration": 236000000,
+  "eval_count": 16,
+  "eval_duration": 269000000
+}
+```
+
 #### Request (JSON mode)
 
 > [!IMPORTANT]
@@ -456,11 +506,15 @@ The `message` object has the following fields:
 
 Advanced parameters (optional):
 
-- `format`: the format to return a response in. Currently the only accepted value is `json`
+- `format`: the format to return a response in. Format can be `json` or a JSON schema. 
 - `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
 - `stream`: if `false` the response will be returned as a single response object, rather than a stream of objects
 - `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
 
+### Structured outputs
+
+Structured outputs are supported by providing a JSON schema in the `format` parameter. The model will generate a response that matches the schema. See the [Chat request (Structured outputs)](#chat-request-structured-outputs) example below.
+
 ### Examples
 
 #### Chat Request (Streaming)
@@ -551,6 +605,54 @@ curl http://localhost:11434/api/chat -d '{
 }
 ```
 
+#### Chat request (Structured outputs)
+
+##### Request
+
+```shell
+curl -X POST http://localhost:11434/api/chat -H "Content-Type: application/json" -d '{
+  "model": "llama3.1",
+  "messages": [{"role": "user", "content": "Ollama is 22 years old and busy saving the world. Return a JSON object with the age and availability."}],
+  "stream": false,
+  "format": {
+    "type": "object",
+    "properties": {
+      "age": {
+        "type": "integer"
+      },
+      "available": {
+        "type": "boolean"
+      }
+    },
+    "required": [
+      "age",
+      "available"
+    ]
+  },
+  "options": {
+    "temperature": 0
+  }
+}'
+```
+
+##### Response
+
+```json
+{
+  "model": "llama3.1",
+  "created_at": "2024-12-06T00:46:58.265747Z",
+  "message": { "role": "assistant", "content": "{\"age\": 22, \"available\": false}" },
+  "done_reason": "stop",
+  "done": true,
+  "total_duration": 2254970291,
+  "load_duration": 574751416,
+  "prompt_eval_count": 34,
+  "prompt_eval_duration": 1502000000,
+  "eval_count": 12,
+  "eval_duration": 175000000
+}
+```
+
 #### Chat request (With History)
 
 Send a chat message with a conversation history. You can use this same approach to start the conversation using multi-shot or chain-of-thought prompting.