feat: preserve model settings (janhq#3427)

* feat: preserve model settings

* feat: preserve model settings across new threads

* chore: lint fix

* fix: feature toggle off should also affect default value retrieve

core/src/types/model/modelEntity.ts

@@ -107,6 +107,9 @@ export type ModelMetadata = {
   tags: string[]
   size: number
   cover?: string
+  // These settings to preserve model settings across threads
+  default_ctx_len?: number
+  default_max_tokens?: number
 }
 
 /**

extensions/model-extension/src/index.ts

@@ -351,7 +351,7 @@ export default class JanModelExtension extends ModelExtension {
   }
 
   /**
-   * Saves a machine learning model.
+   * Saves a model file.
    * @param model - The model to save.
    * @returns A Promise that resolves when the model is saved.
    */

web/containers/ModelDropdown/index.tsx

@@ -45,6 +45,7 @@ import {
 
 import { extensionManager } from '@/extension'
 
+import { preserveModelSettingsAtom } from '@/helpers/atoms/AppConfig.atom'
 import { inActiveEngineProviderAtom } from '@/helpers/atoms/Extension.atom'
 import {
   configuredModelsAtom,
@@ -89,6 +90,7 @@ const ModelDropdown = ({
   const featuredModel = configuredModels.filter((x) =>
     x.metadata.tags.includes('Featured')
   )
+  const preserveModelSettings = useAtomValue(preserveModelSettingsAtom)
 
   useClickOutside(() => !filterOptionsOpen && setOpen(false), null, [
     dropdownOptions,
@@ -161,14 +163,25 @@ const ModelDropdown = ({
       if (activeThread) {
         // Default setting ctx_len for the model for a better onboarding experience
         // TODO: When Cortex support hardware instructions, we should remove this
+        const defaultContextLength = preserveModelSettings
+          ? model?.metadata?.default_ctx_len
+          : 2048
+        const defaultMaxTokens = preserveModelSettings
+          ? model?.metadata?.default_max_tokens
+          : 2048
         const overriddenSettings =
           model?.settings.ctx_len && model.settings.ctx_len > 2048
-            ? { ctx_len: 2048 }
+            ? { ctx_len: defaultContextLength }
+            : {}
+        const overriddenParameters =
+          model?.parameters.max_tokens && model.parameters.max_tokens
+            ? { max_tokens: defaultMaxTokens }
             : {}
 
         const modelParams = {
           ...model?.parameters,
           ...model?.settings,
+          ...overriddenParameters,
           ...overriddenSettings,
         }
 

web/helpers/atoms/AppConfig.atom.ts

@@ -7,6 +7,7 @@ const VULKAN_ENABLED = 'vulkanEnabled'
 const IGNORE_SSL = 'ignoreSSLFeature'
 const HTTPS_PROXY_FEATURE = 'httpsProxyFeature'
 const QUICK_ASK_ENABLED = 'quickAskEnabled'
+const PRESERVE_MODEL_SETTINGS = 'preserveModelSettings'
 
 export const janDataFolderPathAtom = atom('')
 
@@ -23,3 +24,9 @@ export const vulkanEnabledAtom = atomWithStorage(VULKAN_ENABLED, false)
 export const quickAskEnabledAtom = atomWithStorage(QUICK_ASK_ENABLED, false)
 
 export const hostAtom = atom('http://localhost:1337/')
+
+// This feature is to allow user to cache model settings on thread creation
+export const preserveModelSettingsAtom = atomWithStorage(
+  PRESERVE_MODEL_SETTINGS,
+  false
+)

web/helpers/atoms/Model.atom.ts

@@ -34,6 +34,17 @@ export const removeDownloadingModelAtom = atom(
 
 export const downloadedModelsAtom = atom<Model[]>([])
 
+export const updateDownloadedModelAtom = atom(
+  null,
+  (get, set, updatedModel: Model) => {
+    const models: Model[] = get(downloadedModelsAtom).map((c) =>
+      c.id === updatedModel.id ? updatedModel : c
+    )
+
+    set(downloadedModelsAtom, models)
+  }
+)
+
 export const removeDownloadedModelAtom = atom(
   null,
   (get, set, modelId: string) => {

web/hooks/useCreateNewThread.ts

@@ -10,7 +10,7 @@ import {
   Model,
   AssistantTool,
 } from '@janhq/core'
-import { atom, useAtomValue, useSetAtom } from 'jotai'
+import { atom, useAtom, useAtomValue, useSetAtom } from 'jotai'
 
 import { copyOverInstructionEnabledAtom } from '@/containers/CopyInstruction'
 import { fileUploadAtom } from '@/containers/Providers/Jotai'
@@ -24,7 +24,10 @@ import useSetActiveThread from './useSetActiveThread'
 
 import { extensionManager } from '@/extension'
 
-import { experimentalFeatureEnabledAtom } from '@/helpers/atoms/AppConfig.atom'
+import {
+  experimentalFeatureEnabledAtom,
+  preserveModelSettingsAtom,
+} from '@/helpers/atoms/AppConfig.atom'
 import { selectedModelAtom } from '@/helpers/atoms/Model.atom'
 import {
   threadsAtom,
@@ -62,6 +65,7 @@ export const useCreateNewThread = () => {
   const copyOverInstructionEnabled = useAtomValue(
     copyOverInstructionEnabledAtom
   )
+  const preserveModelSettings = useAtomValue(preserveModelSettingsAtom)
   const activeThread = useAtomValue(activeThreadAtom)
 
   const experimentalEnabled = useAtomValue(experimentalFeatureEnabledAtom)
@@ -99,15 +103,20 @@ export const useCreateNewThread = () => {
       enabled: true,
       settings: assistant.tools && assistant.tools[0].settings,
     }
-
+    const defaultContextLength = preserveModelSettings
+      ? model?.metadata?.default_ctx_len
+      : 2048
+    const defaultMaxTokens = preserveModelSettings
+      ? model?.metadata?.default_max_tokens
+      : 2048
     const overriddenSettings =
       defaultModel?.settings.ctx_len && defaultModel.settings.ctx_len > 2048
-        ? { ctx_len: 2048 }
+        ? { ctx_len: defaultContextLength }
         : {}
 
     const overriddenParameters =
       defaultModel?.parameters.max_tokens && defaultModel.parameters.max_tokens
-        ? { max_tokens: 2048 }
+        ? { max_tokens: defaultMaxTokens }
         : {}
 
     const createdAt = Date.now()

web/hooks/useUpdateModelParameters.ts

@@ -4,16 +4,22 @@ import {
   ConversationalExtension,
   ExtensionTypeEnum,
   InferenceEngine,
+  Model,
+  ModelExtension,
   Thread,
   ThreadAssistantInfo,
 } from '@janhq/core'
 
-import { useAtomValue, useSetAtom } from 'jotai'
+import { useAtom, useAtomValue, useSetAtom } from 'jotai'
 
 import { toRuntimeParams, toSettingParams } from '@/utils/modelParam'
 
 import { extensionManager } from '@/extension'
-import { selectedModelAtom } from '@/helpers/atoms/Model.atom'
+import { preserveModelSettingsAtom } from '@/helpers/atoms/AppConfig.atom'
+import {
+  selectedModelAtom,
+  updateDownloadedModelAtom,
+} from '@/helpers/atoms/Model.atom'
 import {
   ModelParams,
   getActiveThreadModelParamsAtom,
@@ -28,8 +34,10 @@ export type UpdateModelParameter = {
 
 export default function useUpdateModelParameters() {
   const activeModelParams = useAtomValue(getActiveThreadModelParamsAtom)
-  const selectedModel = useAtomValue(selectedModelAtom)
+  const [selectedModel, setSelectedModel] = useAtom(selectedModelAtom)
   const setThreadModelParams = useSetAtom(setThreadModelParamsAtom)
+  const updateDownloadedModel = useSetAtom(updateDownloadedModelAtom)
+  const preserveModelFeatureEnabled = useAtomValue(preserveModelSettingsAtom)
 
   const updateModelParameter = useCallback(
     async (thread: Thread, settings: UpdateModelParameter) => {
@@ -40,12 +48,11 @@ export default function useUpdateModelParameters() {
 
       // update the state
       setThreadModelParams(thread.id, updatedModelParams)
+      const runtimeParams = toRuntimeParams(updatedModelParams)
+      const settingParams = toSettingParams(updatedModelParams)
 
       const assistants = thread.assistants.map(
         (assistant: ThreadAssistantInfo) => {
-          const runtimeParams = toRuntimeParams(updatedModelParams)
-          const settingParams = toSettingParams(updatedModelParams)
-
           assistant.model.parameters = runtimeParams
           assistant.model.settings = settingParams
           if (selectedModel) {
@@ -65,6 +72,33 @@ export default function useUpdateModelParameters() {
       await extensionManager
         .get<ConversationalExtension>(ExtensionTypeEnum.Conversational)
         ?.saveThread(updatedThread)
+
+      // Persists default settings to model file
+      // Do not overwrite ctx_len and max_tokens
+      if (preserveModelFeatureEnabled && selectedModel) {
+        const updatedModel = {
+          ...selectedModel,
+          parameters: {
+            ...runtimeParams,
+            max_tokens: selectedModel.parameters.max_tokens,
+          },
+          settings: {
+            ...settingParams,
+            ctx_len: selectedModel.settings.ctx_len,
+          },
+          metadata: {
+            ...selectedModel.metadata,
+            default_ctx_len: settingParams.ctx_len,
+            default_max_tokens: runtimeParams.max_tokens,
+          },
+        } as Model
+
+        await extensionManager
+          .get<ModelExtension>(ExtensionTypeEnum.Model)
+          ?.saveModel(updatedModel)
+        setSelectedModel(updatedModel)
+        updateDownloadedModel(updatedModel)
+      }
     },
     [activeModelParams, selectedModel, setThreadModelParams]
   )

web/screens/Settings/Advanced/index.tsx

@@ -35,6 +35,7 @@ import {
   proxyEnabledAtom,
   vulkanEnabledAtom,
   quickAskEnabledAtom,
+  preserveModelSettingsAtom,
 } from '@/helpers/atoms/AppConfig.atom'
 
 type GPU = {
@@ -64,6 +65,9 @@ const Advanced = () => {
   const [proxyEnabled, setProxyEnabled] = useAtom(proxyEnabledAtom)
   const quickAskEnabled = useAtomValue(quickAskEnabledAtom)
 
+  const [preserveModelSettings, setPreserveModelSettings] = useAtom(
+    preserveModelSettingsAtom
+  )
   const [proxy, setProxy] = useAtom(proxyAtom)
   const [ignoreSSL, setIgnoreSSL] = useAtom(ignoreSslAtom)
 
@@ -385,6 +389,27 @@ const Advanced = () => {
           </div>
         )}
 
+        {experimentalEnabled && (
+          <div className="flex w-full flex-col items-start justify-between gap-4 border-b border-[hsla(var(--app-border))] py-4 first:pt-0 last:border-none sm:flex-row">
+            <div className="flex-shrink-0 space-y-1">
+              <div className="flex gap-x-2">
+                <h6 className="font-semibold capitalize">
+                  Preserve Model Settings
+                </h6>
+              </div>
+              <p className="font-medium leading-relaxed text-[hsla(var(--text-secondary))]">
+                Save model settings changes directly to the model file so that
+                new threads will reuse the previous settings.
+              </p>
+            </div>
+
+            <Switch
+              checked={preserveModelSettings}
+              onChange={(e) => setPreserveModelSettings(e.target.checked)}
+            />
+          </div>
+        )}
+
         <DataFolder />
 
         {/* Proxy */}