feat: adds SmolLM (#78)

* feat: add SmolLM 1.7B Q8_0 * feat: add SmolLM2 1.7B Q4_0_4_4 * chore: bump version
a-ghorbani · Nov 3, 2024 · 118652b · 118652b
1 parent c742f1d
commit 118652b
Show file tree

Hide file tree

Showing 7 changed files with 126 additions and 32 deletions.
diff --git a/android/app/build.gradle b/android/app/build.gradle
@@ -81,8 +81,8 @@ android {
         applicationId "com.pocketpalai"
         minSdkVersion rootProject.ext.minSdkVersion
         targetSdkVersion rootProject.ext.targetSdkVersion
-        versionCode 12 
-        versionName "1.4.5"
+        versionCode 13 
+        versionName "1.4.6"
         ndk {
             abiFilters "arm64-v8a", "x86_64"
         }

diff --git a/ios/PocketPal.xcodeproj/project.pbxproj b/ios/PocketPal.xcodeproj/project.pbxproj
@@ -479,7 +479,7 @@
 			buildSettings = {
 				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
 				CLANG_ENABLE_MODULES = YES;
-				CURRENT_PROJECT_VERSION = 12;
+				CURRENT_PROJECT_VERSION = 13;
 				DEVELOPMENT_TEAM = MYXGXY23Y6;
 				ENABLE_BITCODE = NO;
 				INFOPLIST_FILE = PocketPal/Info.plist;
@@ -488,7 +488,7 @@
 					"$(inherited)",
 					"@executable_path/Frameworks",
 				);
-				MARKETING_VERSION = 1.4.5;
+				MARKETING_VERSION = 1.4.6;
 				OTHER_CPLUSPLUSFLAGS = (
 					"$(OTHER_CFLAGS)",
 					"-DFOLLY_NO_CONFIG",
@@ -521,15 +521,15 @@
 			buildSettings = {
 				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
 				CLANG_ENABLE_MODULES = YES;
-				CURRENT_PROJECT_VERSION = 12;
+				CURRENT_PROJECT_VERSION = 13;
 				DEVELOPMENT_TEAM = MYXGXY23Y6;
 				INFOPLIST_FILE = PocketPal/Info.plist;
 				IPHONEOS_DEPLOYMENT_TARGET = 14.0;
 				LD_RUNPATH_SEARCH_PATHS = (
 					"$(inherited)",
 					"@executable_path/Frameworks",
 				);
-				MARKETING_VERSION = 1.4.5;
+				MARKETING_VERSION = 1.4.6;
 				OTHER_CPLUSPLUSFLAGS = (
 					"$(OTHER_CFLAGS)",
 					"-DFOLLY_NO_CONFIG",

diff --git a/package.json b/package.json
@@ -1,6 +1,6 @@
 {
   "name": "PocketPal",
-  "version": "1.4.5",
+  "version": "1.4.6",
   "private": true,
   "scripts": {
     "prepare": "husky",

diff --git a/src/hooks/useChatSession.ts b/src/hooks/useChatSession.ts
@@ -9,11 +9,7 @@ import {L10nContext} from '../utils';
 import {chatSessionStore, modelStore} from '../store';
 
 import {MessageType, User} from '../utils/types';
-import {
-  applyChatTemplate,
-  chatTemplates,
-  convertToChatMessages,
-} from '../utils/chat';
+import {applyChatTemplate, convertToChatMessages} from '../utils/chat';
 
 export const useChatSession = (
   context: LlamaContext | undefined,
@@ -113,10 +109,10 @@ export const useChatSession = (
       ]),
     ];
 
-    const prompt = applyChatTemplate(
-      modelStore.activeModel?.chatTemplate || chatTemplates.default,
+    const prompt = await applyChatTemplate(
       chatMessages,
-      1000, // This is not used.
+      modelStore.activeModel ?? null,
+      context,
     );
 
     const completionParams = toJS(modelStore.activeModel?.completionSettings);

diff --git a/src/store/defaultModels.ts b/src/store/defaultModels.ts
@@ -2,9 +2,10 @@ import {Model} from '../utils/types';
 import {chatTemplates, defaultCompletionParams} from '../utils/chat';
 import {Platform} from 'react-native';
 
-export const MODEL_LIST_VERSION = 6;
+export const MODEL_LIST_VERSION = 8;
 
 export const defaultModels: Model[] = [
+  // -------- Gemma --------
   {
     id: 'google/gemma-2-2b-it-GGUF',
     name: 'gemma-2-2b-it-GGUF (Q6_K)',
@@ -97,6 +98,7 @@ export const defaultModels: Model[] = [
       stop: ['<end_of_turn>'],
     },
   },
+  // -------- Danube --------
   {
     id: 'h2o-danube3-4b-chat-Q4_K_M.gguf',
     name: 'H2O.ai Danube 3 (Q4_K_M)',
@@ -193,6 +195,7 @@ export const defaultModels: Model[] = [
       penalty_repeat: 1.075,
     },
   },
+  // -------- Phi --------
   {
     id: 'Phi-3.5-mini-instruct.Q4_K_M.gguf',
     name: 'Phi-3.5 mini 4k instruct (Q4_K_M)',
@@ -249,6 +252,7 @@ export const defaultModels: Model[] = [
       stop: ['<|end|>'],
     },
   },
+  // -------- Qwen --------
   {
     id: 'qwen2-1_5b-instruct-q8_0.gguf',
     name: 'Qwen2-1.5B-Instruct (Q8_0)',
@@ -333,6 +337,7 @@ export const defaultModels: Model[] = [
       stop: ['<|im_end|>'],
     },
   },
+  // -------- Llama --------
   ...(Platform.OS === 'android'
     ? [
         {
@@ -537,4 +542,65 @@ export const defaultModels: Model[] = [
       stop: ['<|eot_id|>'],
     },
   },
+  // -------- SmolLM --------
+  {
+    id: 'default-bartowski/SmolLM2-1.7B-Instruct-Q8_0.gguf',
+    name: 'SmolLM2-1.7B-Instruct (Q8_0)',
+    type: 'SmolLM',
+    size: '1.82',
+    params: '1.7',
+    isDownloaded: false,
+    downloadUrl:
+      'https://huggingface.co/bartowski/SmolLM2-1.7B-Instruct-GGUF/resolve/main/SmolLM2-1.7B-Instruct-Q8_0.gguf?download=true',
+    hfUrl: 'https://huggingface.co/bartowski/SmolLM2-1.7B-Instruct-GGUF',
+    progress: 0,
+    filename: 'default-SmolLM2-1.7B-Instruct-Q8_0.gguf',
+    isLocal: false,
+    defaultChatTemplate: chatTemplates.smolLM,
+    chatTemplate: chatTemplates.smolLM,
+    defaultCompletionSettings: {
+      ...defaultCompletionParams,
+      n_predict: 500,
+      temperature: 0.7,
+      stop: ['<|endoftext|>', '<|im_end|>'],
+    },
+    completionSettings: {
+      ...defaultCompletionParams,
+      n_predict: 500,
+      temperature: 0.7,
+      stop: ['<|endoftext|>', '<|im_end|>'],
+    },
+  },
+  ...(Platform.OS === 'android'
+    ? [
+        {
+          id: 'default-bartowski/SmolLM2-1.7B-Instruct-Q4_0_4_4.gguf',
+          name: 'SmolLM2-1.7B-Instruct (Q4_0_4_4)',
+          type: 'SmolLM',
+          size: '0.99',
+          params: '1.7',
+          isDownloaded: false,
+          downloadUrl:
+            'https://huggingface.co/bartowski/SmolLM2-1.7B-Instruct-GGUF/resolve/main/SmolLM2-1.7B-Instruct-Q4_0_4_4.gguf?download=true',
+          hfUrl: 'https://huggingface.co/bartowski/SmolLM2-1.7B-Instruct-GGUF',
+          progress: 0,
+          filename: 'default-SmolLM2-1.7B-Instruct-Q4_0_4_4.gguf',
+          isLocal: false,
+          defaultChatTemplate: chatTemplates.smolLM,
+          chatTemplate: chatTemplates.smolLM,
+          defaultCompletionSettings: {
+            ...defaultCompletionParams,
+            n_predict: 1000,
+            temperature: 0.7,
+            stop: ['<|endoftext|>', '<|im_end|>'],
+          },
+          completionSettings: {
+            ...defaultCompletionParams,
+            n_predict: 1000,
+            temperature: 0.7,
+            stop: ['<|endoftext|>', '<|im_end|>'],
+          },
+        },
+      ]
+    : []),
 ];
diff --git a/src/utils/__tests__/chat.test.ts b/src/utils/__tests__/chat.test.ts
@@ -1,6 +1,7 @@
 import {Templates} from 'chat-formatter';
 import {applyChatTemplate} from '../chat';
 import {ChatMessage, ChatTemplateConfig} from '../types';
+import {createModel} from '../../../jest/fixtures/models';
 
 const conversationWSystem: ChatMessage[] = [
   {role: 'system', content: 'System prompt. '},
@@ -18,7 +19,8 @@ describe('Test Danube2 Chat Templates', () => {
       addGenerationPrompt: true,
       name: 'danube2',
     };
-    const result = applyChatTemplate(chatTemplate, conversationWSystem, -1);
+    const model = createModel({chatTemplate: chatTemplate});
+    const result = await applyChatTemplate(conversationWSystem, model, null);
     expect(result).toBe(
       'System prompt. </s><|prompt|>Hi there!</s><|answer|>Nice to meet you!</s><|prompt|>Can I ask a question?</s><|answer|>',
     );

diff --git a/src/utils/chat.ts b/src/utils/chat.ts
@@ -1,7 +1,6 @@
 import {applyTemplate, Templates} from 'chat-formatter';
-import {ChatMessage, ChatTemplateConfig, MessageType} from './types';
-//import {assistant} from '../store/ChatSessionStore';
-import {CompletionParams} from '@pocketpalai/llama.rn';
+import {ChatMessage, ChatTemplateConfig, MessageType, Model} from './types';
+import {CompletionParams, LlamaContext} from '@pocketpalai/llama.rn';
 
 export const userId = 'y9d7f8pgn';
 export const assistantId = 'h3o3lc5xj';
@@ -22,21 +21,42 @@ export function convertToChatMessages(
     .reverse();
 }
 
-export function applyChatTemplate(
-  template: ChatTemplateConfig,
-  chat: ChatMessage[],
-  // eslint-disable-next-line @typescript-eslint/no-unused-vars
-  length: number, //TODO: inforce length of formattedChat to fit the context.
-): string {
-  const formattedChat: string = applyTemplate(chat, {
-    customTemplate: template,
-    addGenerationPrompt: template.addGenerationPrompt,
-  }) as string;
+export async function applyChatTemplate(
+  messages: ChatMessage[],
+  model: Model | null,
+  context: LlamaContext | null,
+): Promise<string> {
+  const modelChatTemplate = model?.chatTemplate;
+  const contextChatTemplate = (context?.model as any)?.metadata?.[
+    'tokenizer.chat_template'
+  ];
 
-  return formattedChat;
+  let formattedChat: string | undefined;
+
+  try {
+    if (modelChatTemplate?.chatTemplate) {
+      formattedChat = applyTemplate(messages, {
+        customTemplate: modelChatTemplate,
+        addGenerationPrompt: modelChatTemplate.addGenerationPrompt,
+      }) as string;
+    } else if (contextChatTemplate) {
+      formattedChat = await context?.getFormattedChat(messages);
+    }
+
+    if (!formattedChat) {
+      formattedChat = applyTemplate(messages, {
+        customTemplate: chatTemplates.default,
+        addGenerationPrompt: chatTemplates.default.addGenerationPrompt,
+      }) as string;
+    }
+  } catch (error) {
+    console.error('Error applying chat template:', error); // TODO: handle error
+  }
+
+  return formattedChat || ' ';
 }
 
-export const chatTemplates = {
+export const chatTemplates: Record<string, ChatTemplateConfig> = {
   danube3: {
     ...Templates.templates.danube2,
     name: 'danube3',
@@ -112,6 +132,16 @@ export const chatTemplates = {
     systemPrompt:
       'You are Qwen, created by Alibaba Cloud. You are a helpful assistant.',
   },
+  smolLM: {
+    name: 'smolLM',
+    addGenerationPrompt: true,
+    systemPrompt: 'You are a helpful assistant.',
+    bosToken: '<|im_start|>',
+    eosToken: '<|im_end|>',
+    addBosToken: false,
+    addEosToken: false,
+    chatTemplate: '',
+  },
 };
 
 export const defaultCompletionParams: CompletionParams = {