Add distinction between chat and instruct models

src/lib/ChatRequest.svelte

@@ -8,7 +8,7 @@
     import { getDefaultModel, getRequestSettingList } from './Settings.svelte'
     import { v4 as uuidv4 } from 'uuid'
     import { get } from 'svelte/store'
-    import { getModelDetail } from './Models.svelte'
+    import { getLeadPrompt, getModelDetail } from './Models.svelte'
 
 export class ChatRequest {
       constructor () {
@@ -238,9 +238,10 @@ export class ChatRequest {
         const lastMessage = messages[messages.length - 1]
         const isContinue = lastMessage?.role === 'assistant' && lastMessage.finish_reason === 'length'
         const isUserPrompt = lastMessage?.role === 'user'
+        let results: Message[] = []
+        let injectedPrompt = false
         if (hiddenPromptPrefix && (isUserPrompt || isContinue)) {
-          let injectedPrompt = false
-          const results = hiddenPromptPrefix.split(/[\s\r\n]*::EOM::[\s\r\n]*/).reduce((a, m) => {
+          results = hiddenPromptPrefix.split(/[\s\r\n]*::EOM::[\s\r\n]*/).reduce((a, m) => {
             m = m.trim()
             if (m.length) {
               if (m.match(/\[\[USER_PROMPT\]\]/)) {
@@ -265,9 +266,21 @@ export class ChatRequest {
             }
           }
           if (injectedPrompt) messages.pop()
-          return results
         }
-        return []
+        const model = this.getModel()
+        const messageDetail = getModelDetail(model)
+        if (getLeadPrompt(this.getChat()).trim() && messageDetail.type === 'chat') {
+          const lastMessage = (results.length && injectedPrompt && !isContinue) ? results[results.length - 1] : messages[messages.length - 1]
+          if (lastMessage?.role !== 'assistant') {
+            const leadMessage = { role: 'assistant', content: getLeadPrompt(this.getChat()) } as Message
+            if (insert) {
+              messages.push(leadMessage)
+            } else {
+              results.push(leadMessage)
+            }
+          }
+        }
+        return results
       }
 
       /**

src/lib/Settings.svelte

@@ -597,17 +597,6 @@ const chatSettingsList: ChatSetting[] = [
         },
         hide: hideModelSetting
       },
-      {
-        key: 'leadPrompt',
-        name: 'Completion Lead Sequence ',
-        title: 'Sequence to hint the LLM should answer as assistant.',
-        type: 'textarea',
-        placeholder: (chatId) => {
-          const val = getModelDetail(getChatSettings(chatId).model).leadPrompt
-          return val || ''
-        },
-        hide: hideModelSetting
-      },
       {
         key: 'systemMessageStart',
         name: 'System Message Start Sequence',
@@ -630,6 +619,17 @@ const chatSettingsList: ChatSetting[] = [
         },
         hide: hideModelSetting
       },
+      {
+        key: 'leadPrompt',
+        name: 'Completion Lead Sequence',
+        title: 'Sequence to hint to answer as assistant.',
+        type: 'textarea',
+        placeholder: (chatId) => {
+          const val = getModelDetail(getChatSettings(chatId).model).leadPrompt
+          return val || ''
+        },
+        hide: hideModelSetting
+      },
       {
         // logit bias editor not implemented yet
         key: 'logit_bias',

src/lib/Types.svelte

@@ -6,7 +6,7 @@
 
 export type Model = typeof supportedChatModelKeys[number];
 
-export type RequestType = 'chat' | 'image'
+export type RequestType = 'chat' | 'instruct' | 'image'
 
 export type Usage = {
     completion_tokens: number;

src/lib/providers/openai/models.svelte

@@ -18,10 +18,10 @@ const hiddenSettings = {
       userMessageEnd: true,
       assistantMessageStart: true,
       assistantMessageEnd: true,
-      leadPrompt: true,
       systemMessageStart: true,
       systemMessageEnd: true,
       repititionPenalty: true
+      // leadPrompt: true
 }
 
 const chatModelBase = {

src/lib/providers/openai/util.svelte

@@ -49,7 +49,7 @@ const getSupportedModels = async (): Promise<Record<string, boolean>> => {
 
 export const checkModel = async (modelDetail: ModelDetail) => {
   const supportedModels = await getSupportedModels()
-  if (modelDetail.type === 'chat') {
+  if (modelDetail.type === 'chat' || modelDetail.type === 'instruct') {
         modelDetail.enabled = !!supportedModels[modelDetail.modelQuery || '']
   } else {
         // image request.  If we have any models, allow image endpoint

src/lib/providers/petals/models.svelte

@@ -17,7 +17,7 @@ const hideSettings = {
 }
 
 const chatModelBase = {
-  type: 'chat',
+  type: 'instruct', // Used for chat, but these models operate like instruct models -- you have to manually structure the messages sent to them
   help: 'Below are the settings that can be changed for the API calls. See <a target="_blank" href="https://platform.openai.com/docs/api-reference/chat/create">this overview</a> to start, though not all settings translate to Petals.',
   check: checkModel,
   start: '<s>',
@@ -45,7 +45,7 @@ const chatModelBase = {
         return prompts.reduce((a, m) => {
           a += countMessageTokens(m, model, chat)
           return a
-        }, 0) + countTokens(model, getStartSequence(chat)) + countTokens(model, getLeadPrompt(chat))
+        }, 0) + countTokens(model, getStartSequence(chat)) + ((prompts[prompts.length - 1] || {}).role !== 'assistant' ? countTokens(model, getLeadPrompt(chat)) : 0)
   }
 } as ModelDetail
 
@@ -74,6 +74,7 @@ export const chatModels : Record<string, ModelDetail> = {
         assistantStart: '[[SYSTEM_PROMPT]][[USER_PROMPT]]',
         systemStart: '<<SYS>>\n',
         systemEnd: '\n<</SYS>>\n\n'
+        // leadPrompt: ''
       },
       'stabilityai/StableBeluga2': {
         ...chatModelBase,

src/lib/providers/petals/util.svelte

@@ -8,7 +8,7 @@ export const set = (opt: Record<string, any>) => {
 }
 
 export const checkModel = async (modelDetail: ModelDetail) => {
-  if (modelDetail.type === 'chat') {
+  if (modelDetail.type === 'chat' || modelDetail.type === 'instruct') {
         modelDetail.enabled = get(globalStorage).enablePetals
   }
 }