Fix ZeroGPU handling for gr.Server

.changeset/soft-meals-listen.md

@@ -0,0 +1,8 @@
+---
+"@gradio/client": patch
+"@self/app": patch
+"@self/spa": patch
+"gradio": patch
+---
+
+fix:Fix ZeroGPU handling for `gr.Server`

client/js/src/client.ts

@@ -31,6 +31,7 @@ import {
 	parse_and_set_cookies
 } from "./helpers/init_helpers";
 import { check_and_wake_space, check_space_status } from "./helpers/spaces";
+import { initialize_zerogpu_handshake } from "./helpers/zerogpu";
 import { open_stream, readable_stream, close_stream } from "./utils/stream";
 import {
 	API_INFO_ERROR_MSG,
@@ -220,6 +221,8 @@ export class Client {
 	}
 
 	private async init(): Promise<void> {
+		initialize_zerogpu_handshake();
+
 		if (this.options.auth) {
 			await this.resolve_cookies();
 		}

client/js/src/helpers/zerogpu.ts

@@ -0,0 +1,40 @@
+const ZEROGPU_HEADERS_MESSAGE = "supports-zerogpu-headers";
+
+let zerogpu_handshake_initialized = false;
+
+function supports_browser_handshake(): boolean {
+	return (
+		typeof window !== "undefined" &&
+		typeof document !== "undefined" &&
+		typeof window.addEventListener === "function"
+	);
+}
+
+export function get_zerogpu_origin(hostname: string): string | null {
+	if (hostname.includes(".dev.")) {
+		return `https://moon-${hostname.split(".")[1]}.dev.spaces.huggingface.tech`;
+	}
+	if (hostname.endsWith(".hf.space")) {
+		return "https://huggingface.co";
+	}
+	return null;
+}
+
+export function initialize_zerogpu_handshake(): void {
+	if (!supports_browser_handshake() || zerogpu_handshake_initialized) {
+		return;
+	}
+
+	window.addEventListener("message", (event) => {
+		if (event.data === ZEROGPU_HEADERS_MESSAGE) {
+			window.supports_zerogpu_headers = true;
+		}
+	});
+
+	zerogpu_handshake_initialized = true;
+
+	const origin = get_zerogpu_origin(window.location.hostname);
+	if (origin && window.parent !== window) {
+		window.parent.postMessage(ZEROGPU_HEADERS_MESSAGE, origin);
+	}
+}

client/js/src/utils/submit.ts

@@ -12,6 +12,7 @@ import type {
 } from "../types";
 
 import { skip_queue, post_message, handle_payload } from "../helpers/data";
+import { get_zerogpu_origin } from "../helpers/zerogpu";
 import {
 	handle_message,
 	map_data_to_params,
@@ -412,15 +413,13 @@ export function submit(
 					hostname = window?.location?.hostname;
 				}
 
-				let hfhubdev = "dev.spaces.huggingface.tech";
-				const origin = hostname.includes(".dev.")
-					? `https://moon-${hostname.split(".")[1]}.${hfhubdev}`
-					: `https://huggingface.co`;
+				const origin = get_zerogpu_origin(hostname);
 
 				const is_zerogpu_iframe =
 					typeof window !== "undefined" &&
 					typeof document !== "undefined" &&
 					window.parent != window &&
+					!!origin &&
 					window.supports_zerogpu_headers;
 				const zerogpu_auth_promise = is_zerogpu_iframe
 					? post_message<Map<string, string>>("zerogpu-headers", origin)

gradio/queueing.py

@@ -968,13 +968,19 @@ async def process_events(
                     success = False
                     error = err or old_err
                     output = error_payload(error, app.get_blocks().show_error)
+                used_cache = output.get("used_cache") if success else None
+                used_cache = (
+                    cast(Literal["full", "partial"], used_cache)
+                    if used_cache in ("full", "partial")
+                    else None
+                )
                 for event in awake_events:
                     self.send_message(
                         event,
                         ProcessCompletedMessage(
                             output=output,
                             success=success,
-                            used_cache=output.get("used_cache") if success else None,
+                            used_cache=used_cache,
                             cache_duration=output.get("duration"),  # type: ignore[arg-type]
                             avg_time=output.get("average_duration"),  # type: ignore[arg-type]
                         ),
@@ -988,12 +994,18 @@ async def process_events(
                             e
                         ]
                     success = response is not None
+                    used_cache = output.get("used_cache") if success else None
+                    used_cache = (
+                        cast(Literal["full", "partial"], used_cache)
+                        if used_cache in ("full", "partial")
+                        else None
+                    )
                     self.send_message(
                         event,
                         ProcessCompletedMessage(
                             output=output,
                             success=success,
-                            used_cache=output.get("used_cache") if success else None,
+                            used_cache=used_cache,
                             cache_duration=output.get("duration"),  # type: ignore[arg-type]
                             avg_time=output.get("average_duration"),  # type: ignore[arg-type]
                         ),

guides/09_gradio-clients-and-lite/08_server-mode.md

@@ -119,6 +119,8 @@ python run.py
 
 Then open `http://localhost:7860` in your browser. The custom HTML page uses the `@gradio/client` JavaScript library to call the Gradio API endpoints. Meanwhile, the same endpoints are available as MCP tools and through the REST API at `/gradio_api/call/add` and `/gradio_api/call/multiply`.
 
+Note: if your `Server` app uses ZeroGPU, you _must_ call Gradio API endpoints through `@gradio/client` from the browser. The JavaScript client forwards the Hugging Face iframe auth headers needed for ZeroGPU quota handling.
+
 ## Concurrency and Streaming
 
 `app.api()` supports all of the same concurrency and streaming options as `gr.api()`:

js/app/src/routes/[...catchall]/+page.svelte

@@ -293,22 +293,6 @@
 
 		await add_custom_html_head(config.head);
 
-		const supports_zerogpu_headers = "supports-zerogpu-headers";
-		window.addEventListener("message", (event) => {
-			if (event.data === supports_zerogpu_headers) {
-				window.supports_zerogpu_headers = true;
-			}
-		});
-		const hostname = window.location.hostname;
-		const is_hf_host =
-			hostname.includes(".dev.") || hostname.endsWith(".hf.space");
-		if (is_hf_host) {
-			const origin = hostname.includes(".dev.")
-				? `https://moon-${hostname.split(".")[1]}.dev.spaces.huggingface.tech`
-				: `https://huggingface.co`;
-			window.parent.postMessage(supports_zerogpu_headers, origin);
-		}
-
 		if (config.js) {
 			try {
 				const script = document.createElement("script");

js/spa/src/Index.svelte

@@ -356,22 +356,6 @@
 		css_ready = true;
 		window.__is_colab__ = config.is_colab;
 
-		const supports_zerogpu_headers = "supports-zerogpu-headers";
-		window.addEventListener("message", (event) => {
-			if (event.data === supports_zerogpu_headers) {
-				window.supports_zerogpu_headers = true;
-			}
-		});
-		const hostname = window.location.hostname;
-		const is_hf_host =
-			hostname.includes(".dev.") || hostname.endsWith(".hf.space");
-		if (is_hf_host) {
-			const origin = hostname.includes(".dev.")
-				? `https://moon-${hostname.split(".")[1]}.dev.spaces.huggingface.tech`
-				: `https://huggingface.co`;
-			window.parent.postMessage(supports_zerogpu_headers, origin);
-		}
-
 		dispatch("loaded");
 
 		pages = config.pages;