Hotfix/fix demo auth and transformer issue

client/src/components/DemoInterface.tsx

@@ -19,7 +19,10 @@ interface AggregatedResult {
 }
 
 const DemoInterface: React.FC<DemoInterfaceProps> = ({ models, tokenInfo, apiBasePath }) => {
-  const curatedFileList = getCuratedFileList(tokenInfo.tapisHost);
+  const curatedFileList = useMemo(
+    () => getCuratedFileList(tokenInfo.tapisHost),
+    [tokenInfo.tapisHost]
+  );
 
   // only supporting clip models on first pass
   const clipModels = useMemo(

client/src/hooks/useToken.ts

@@ -22,7 +22,7 @@ const getHostFromIss = (iss: string): string => {
   }
 };
 
-const fetchTokenFromPortal = async (): Promise<string | null> => {
+const fetchTokenFromPortalUsingCookie = async (): Promise<string | null> => {
   if (!isInIframe()) {
     return null;
   }
@@ -95,7 +95,7 @@ const validateTokenAndGetHost = async (token: string, fallbackHost: string): Pro
       isValid: true,
     };
   } catch (error) {
-    console.error('Token validation error:', error);
+    console.error('Token missing or validation error:', error);
     return {
       token: '',
       tapisHost: fallbackHost,
@@ -106,30 +106,14 @@ const validateTokenAndGetHost = async (token: string, fallbackHost: string): Pro
 
 const getToken = async (fallbackHost: string): Promise<TokenInfo> => {
   // First, try to get token from parent portal (if in iframe)
-  const tokenFromCorePortal = await fetchTokenFromPortal();
+  const tokenFromCorePortal = await fetchTokenFromPortalUsingCookie();
 
   if (tokenFromCorePortal) {
     // Validate the portal token
-    const result = await validateTokenAndGetHost(tokenFromCorePortal, fallbackHost);
-    if (result.isValid) {
-      // Store in sessionStorage for subsequent requests
-      sessionStorage.setItem('access_token', tokenFromCorePortal);
-
-      // Extract expiry from JWT
-      try {
-        const decoded = jwtDecode<TapisJwtPayload>(tokenFromCorePortal);
-        sessionStorage.setItem('expires_at', (decoded.exp * 1000).toString());
-      } catch {
-        console.warn('Failed to decode JWT for expiry');
-        // Fallback to 1 hour
-        sessionStorage.setItem('expires_at', (Date.now() + 3600000).toString());
-      }
-
-      return result;
-    }
+    return validateTokenAndGetHost(tokenFromCorePortal, fallbackHost);
   }
 
-  // Fall back to sessionStorage (for direct access or cached token)
+  // Use sessionStorage (direct access, not in iframe scenario)
   const token = sessionStorage.getItem('access_token');
   const expiresAt = sessionStorage.getItem('expires_at');
 
@@ -144,13 +128,23 @@ const getToken = async (fallbackHost: string): Promise<TokenInfo> => {
   return validateTokenAndGetHost(token, fallbackHost);
 };
 
+/**
+ * Manages the user's Tapis auth token via React Query.
+ *
+ * staleTime (5 min) ensures the backend middleware gets regular
+ * opportunities to refresh tokens before expiry (backend has 10
+ * min threshold).
+ *
+ * Retry disabled since auth failures require login, not retries.
+ */
 export const useToken = () => {
   const config = useConfig();
 
   return useQuery({
     queryKey: ['token'],
     queryFn: () => getToken(config.host),
     retry: false,
-    refetchOnWindowFocus: false,
+    refetchOnWindowFocus: true,
+    staleTime: 5 * 60 * 1000 /* 5min */,
   });
 };

imageinf/inference/clip_base.py

@@ -72,9 +72,10 @@ def _precompute_text_features(self):
         with torch.no_grad():
             ti = self.processor(text=all_texts, return_tensors="pt", padding=True)
             ti = {k: v.to(self.device) for k, v in ti.items()}
-            emb = self.model.get_text_features(
+            text_out = self.model.text_model(
                 input_ids=ti["input_ids"], attention_mask=ti["attention_mask"]
             )
+            emb = self.model.text_projection(text_out.pooler_output)
             emb = F.normalize(emb, dim=-1)
             self.text_pairs = emb.reshape(len(self.labels), 2, -1)
 
@@ -99,9 +100,8 @@ def classify_image(
         inputs = {k: v.to(self.device) for k, v in inputs.items()}
 
         with torch.no_grad():
-            img_feat = self.model.get_image_features(
-                pixel_values=inputs["pixel_values"]
-            )
+            vision_out = self.model.vision_model(pixel_values=inputs["pixel_values"])
+            img_feat = self.model.visual_projection(vision_out.pooler_output)
             img_feat = F.normalize(img_feat, dim=-1)
 
             sims2 = torch.einsum("bd,lcd->blc", img_feat, self.text_pairs)