fix: unload model while loading cause unknown error (#2649)

* fix: unload model while loading cause unknown error

* chore: mask placeholder

extensions/inference-nitro-extension/src/node/index.ts

@@ -38,6 +38,8 @@ const NITRO_HTTP_VALIDATE_MODEL_URL = `${NITRO_HTTP_SERVER_URL}/inferences/llama
 // The URL for the Nitro subprocess to kill itself
 const NITRO_HTTP_KILL_URL = `${NITRO_HTTP_SERVER_URL}/processmanager/destroy`
 
+const NITRO_PORT_FREE_CHECK_INTERVAL = 100
+
 // The supported model format
 // TODO: Should be an array to support more models
 const SUPPORTED_MODEL_FORMAT = '.gguf'
@@ -150,19 +152,9 @@ async function loadModel(
 async function runNitroAndLoadModel(systemInfo?: SystemInformation) {
   // Gather system information for CPU physical cores and memory
   return killSubprocess()
-    .then(() => tcpPortUsed.waitUntilFree(PORT, 300, 5000))
-    .then(() => {
-      /**
-       * There is a problem with Windows process manager
-       * Should wait for awhile to make sure the port is free and subprocess is killed
-       * The tested threshold is 500ms
-       **/
-      if (process.platform === 'win32') {
-        return new Promise((resolve) => setTimeout(resolve, 500))
-      } else {
-        return Promise.resolve()
-      }
-    })
+    .then(() =>
+      tcpPortUsed.waitUntilFree(PORT, NITRO_PORT_FREE_CHECK_INTERVAL, 5000)
+    )
     .then(() => spawnNitroProcess(systemInfo))
     .then(() => loadLLMModel(currentSettings))
     .then(validateModelStatus)
@@ -235,7 +227,7 @@ function loadLLMModel(settings: any): Promise<Response> {
     },
     body: JSON.stringify(settings),
     retries: 3,
-    retryDelay: 500,
+    retryDelay: 300,
   })
     .then((res) => {
       log(
@@ -266,7 +258,7 @@ async function validateModelStatus(): Promise<void> {
       'Content-Type': 'application/json',
     },
     retries: 5,
-    retryDelay: 500,
+    retryDelay: 300,
   }).then(async (res: Response) => {
     log(
       `[NITRO]::Debug: Validate model state with response ${JSON.stringify(
@@ -311,7 +303,9 @@ async function killSubprocess(): Promise<void> {
       signal: controller.signal,
     })
       .catch(() => {}) // Do nothing with this attempt
-      .then(() => tcpPortUsed.waitUntilFree(PORT, 300, 5000))
+      .then(() =>
+        tcpPortUsed.waitUntilFree(PORT, NITRO_PORT_FREE_CHECK_INTERVAL, 5000)
+      )
       .then(() => log(`[NITRO]::Debug: Nitro process is terminated`))
       .catch((err) => {
         log(
@@ -330,7 +324,7 @@ async function killSubprocess(): Promise<void> {
           return killRequest()
         } else {
           return tcpPortUsed
-            .waitUntilFree(PORT, 300, 5000)
+            .waitUntilFree(PORT, NITRO_PORT_FREE_CHECK_INTERVAL, 5000)
             .then(() => resolve())
             .then(() => log(`[NITRO]::Debug: Nitro process is terminated`))
             .catch(() => {
@@ -391,10 +385,12 @@ function spawnNitroProcess(systemInfo?: SystemInformation): Promise<any> {
       reject(`child process exited with code ${code}`)
     })
 
-    tcpPortUsed.waitUntilUsed(PORT, 300, 30000).then(() => {
-      log(`[NITRO]::Debug: Nitro is ready`)
-      resolve()
-    })
+    tcpPortUsed
+      .waitUntilUsed(PORT, NITRO_PORT_FREE_CHECK_INTERVAL, 30000)
+      .then(() => {
+        log(`[NITRO]::Debug: Nitro is ready`)
+        resolve()
+      })
   })
 }
 

extensions/inference-triton-trtllm-extension/resources/settings.json

@@ -15,7 +15,7 @@
     "description": "The Triton LLM API uses API keys for authentication.",
     "controllerType": "input",
     "controllerProps": {
-      "placeholder": "API Key",
+      "placeholder": "xxxxxxxxxxxxxxxxxxxx",
       "value": "",
       "type": "password"
     }

web/hooks/useActiveModel.ts

@@ -25,6 +25,8 @@ export const stateModelAtom = atom<ModelState>({
   model: undefined,
 })
 
+export let loadModelController: AbortController | undefined
+
 export function useActiveModel() {
   const [activeModel, setActiveModel] = useAtom(activeModelAtom)
   const activeThread = useAtomValue(activeThreadAtom)
@@ -46,6 +48,7 @@ export function useActiveModel() {
       console.debug(`Model ${modelId} is already initialized. Ignore..`)
       return Promise.resolve()
     }
+    loadModelController = new AbortController()
 
     let model = downloadedModelsRef?.current.find((e) => e.id === modelId)
 
@@ -104,6 +107,9 @@ export function useActiveModel() {
         })
       })
       .catch((error) => {
+        if (loadModelController?.signal.aborted)
+          return Promise.reject(new Error('aborted'))
+
         setStateModel(() => ({
           state: 'start',
           loading: false,
@@ -131,12 +137,13 @@ export function useActiveModel() {
 
       setStateModel({ state: 'stop', loading: true, model: stoppingModel })
       const engine = EngineManager.instance().get(stoppingModel.engine)
-      await engine
+      return engine
         ?.unloadModel(stoppingModel)
         .catch()
         .then(() => {
           setActiveModel(undefined)
           setStateModel({ state: 'start', loading: false, model: undefined })
+          loadModelController?.abort()
         })
     },
     [activeModel, setActiveModel, setStateModel, stateModel]

web/screens/Chat/Sidebar/index.tsx

@@ -164,39 +164,33 @@ const Sidebar: React.FC = () => {
         </CardSidebar>
 
         <CardSidebar title="Model" isShow={true}>
-          <div className="px-2 pt-4">
+          <div className="flex flex-col gap-4 px-2 py-4">
             <DropdownListSidebar />
 
             {modelSettings.length > 0 && (
-              <div className="mt-6">
-                <CardSidebar title="Inference Parameters" asChild>
-                  <div className="px-2 py-4">
-                    <ModelSetting componentProps={modelSettings} />
-                  </div>
-                </CardSidebar>
-              </div>
+              <CardSidebar title="Inference Parameters" asChild>
+                <div className="px-2 py-4">
+                  <ModelSetting componentProps={modelSettings} />
+                </div>
+              </CardSidebar>
             )}
 
             {promptTemplateSettings.length > 0 && (
-              <div className="mt-4">
-                <CardSidebar title="Model Parameters" asChild>
-                  <div className="px-2 py-4">
-                    <PromptTemplateSetting
-                      componentData={promptTemplateSettings}
-                    />
-                  </div>
-                </CardSidebar>
-              </div>
+              <CardSidebar title="Model Parameters" asChild>
+                <div className="px-2 py-4">
+                  <PromptTemplateSetting
+                    componentData={promptTemplateSettings}
+                  />
+                </div>
+              </CardSidebar>
             )}
 
             {engineSettings.length > 0 && (
-              <div className="my-4">
-                <CardSidebar title="Engine Parameters" asChild>
-                  <div className="px-2 py-4">
-                    <EngineSetting componentData={engineSettings} />
-                  </div>
-                </CardSidebar>
-              </div>
+              <CardSidebar title="Engine Parameters" asChild>
+                <div className="px-2 py-4">
+                  <EngineSetting componentData={engineSettings} />
+                </div>
+              </CardSidebar>
             )}
           </div>
         </CardSidebar>