test(webapp): cover decentralized browser training across browsers by ale18V · Pull Request #1239 · epfml/disco

ale18V · 2026-10-01T14:44:14Z

What & why

Adds end-to-end webapp coverage for decentralized browser-to-browser training, as requested in #1196 and the review of #1230. Four real browser participants (two Chrome and two Firefox) train the default MNIST task through the webapp against a live Disco server over WebRTC. The run exercises Chrome–Chrome, Firefox–Firefox, and Chrome–Firefox model exchange.

Technical plan

Keep the test with the other webapp training suites. Start the standard built Disco server and a Vite webapp, then run four Cypress browser processes on a shared Xvfb display in CI. Assert that all participants complete 20 epochs and 10 model-sharing rounds.

Deviations from plan

None. This replaces the Node participant prototype in #1232 with browser-to-browser collaboration.

Todo

None.

ale18V · 2026-10-10T15:45:16Z

I've also refactored the support folder to have a dedicated support/training folder

JulienVig

Thanks for the great work! I mostly have nitpicking comments but I'm requesting changes because the CI is currently flaky. I just merged #1259 which should fix the error. You can request the review again once you rebased and see that the CI passes consistently

JulienVig · 2026-10-12T10:17:07Z

+const browsers = process.env.DISCO_E2E_BROWSERS?.split(",") ?? [
+  "chrome",
+  "chrome",
+  "firefox",


Suggested change

"firefox",

"edge",

Could you swap one of the firefoxes by edge? I realized that in a hospital use case, most people would probably use the default microsoft browser. Since it's chromium based it shouldn't change much but let's include since it doesn't cost anything

JulienVig · 2026-10-12T10:36:17Z

+import { spawn, type ChildProcessWithoutNullStreams } from "node:child_process";
+import { createInterface } from "node:readline";
+
+const browsers = process.env.DISCO_E2E_BROWSERS?.split(",") ?? [


Is DISCO_E2E_BROWSERS ever used? If not let's remove it and we can add it back when we actually need it. Otherwise can you document it?

Also this split is error prone if whitespaces are used for ex: "chrome, firefox", I think it should fail loudly if the env contains whitespace

JulienVig · 2026-10-12T10:40:33Z

+    process.exitCode = 1;
+  });
+
+const results = browsers.map((browser, index) => {


The 4 browsers are started sequentially but MNIST by default only waits for 3 peers to start aggregation, this can leave the last browser hanging. This is maybe the reason why the CI failed for your other PR.
You could overwrite the min number of participants for MNIST to 4

JulienVig · 2026-10-12T10:46:18Z

+  for (const child of children) {
+    if (child.pid === undefined || child.exitCode !== null) continue;
+    try {
+      process.kill(-child.pid, "SIGTERM");


Suggested change

process.kill(-child.pid, "SIGTERM");

process.kill(-child.pid, "SIGTERM"); // minus sign means kill both cypress and its browser

JulienVig · 2026-10-12T10:55:50Z

+    cy.contains("button", "Start training").click();
+    assertNoErrorToast();
+
+    const expectedParticipants = Cypress.env("numberOfParticipants") ?? 4;


Cypress.env is deprecated. Use cy.env or Cypress.expose instead

JulienVig · 2026-10-12T11:00:26Z

+    cy.contains("button", "group").click();
+    cy.task<string[]>("readdir", "../datasets/CIFAR10").then((files) =>
+      cy
+        .contains("h4", "Group label: 0")


Suggested change

.contains("h4", "Group label: 0")

.contains("h4", "Group label: 0") // dummy labels for now

JulienVig · 2026-10-12T11:02:43Z

+    cy.visit("/");
+    cy.contains("a", "Start training").click();
+    cy.get(".driver-popover-close-btn").click();
+    cy.get("#mnist", { timeout: 30_000 })
+      .contains("button", "participate")
+      .click();


I think you could extend goToTaskOverview to accept an optional task ID and go to task if specified

JulienVig · 2026-10-12T13:45:24Z

+    cy.contains("h6", "Collaborative model sharing")
+      .next({ timeout: 270_000 })
+      .should("have.text", "10");
+    cy.contains("Training successfully completed", { timeout: 270_000 });


Suggested change

cy.contains("h6", "Collaborative model sharing")

.next({ timeout: 270_000 })

.should("have.text", "10");

cy.contains("Training successfully completed", { timeout: 270_000 });

cy.contains("h6", "Collaborative model sharing")

.next()

.should("have.text", "10");

cy.contains("Training successfully completed");

I think that once we reached epoch 20 there's no need to add such long timeouts

JulienVig · 2026-10-12T13:46:32Z

+    const expectedParticipants = Cypress.env("numberOfParticipants") ?? 4;
+
+    cy.contains("h6", "number of participants")
+      .next({ timeout: 270_000 })


Do we actually need to wait that long before the 3 participants are connected?

JulienVig · 2026-10-12T13:49:41Z

+    child.once("exit", (code, signal) => {
+      const passed = code === 0;
+      if (!passed) {
+        console.error(`[${label}] exited with ${code ?? signal}`);
+        stopChildren();
+      }


Could you save and print which browser failed first? (For example by saving the label in variable that you only set if undefined). This would be helpful for understanding what went wrong

ale18V mentioned this pull request Oct 2, 2026

test(webapp): cover decentralized browser training #1232

Closed

ale18V force-pushed the test/browser-browser-cypress-ci branch from 55697bc to e489b7f Compare October 2, 2026 09:51

ale18V force-pushed the test/browser-browser-cypress-ci branch from e489b7f to 0e054da Compare October 2, 2026 12:33

ale18V mentioned this pull request Oct 2, 2026

Test federated webapp training #1230

Merged

ale18V force-pushed the test/browser-browser-cypress-ci branch from 0e054da to e76d17c Compare October 3, 2026 11:17

ale18V force-pushed the test/browser-browser-cypress-ci branch from e76d17c to a7b402f Compare October 3, 2026 11:20

This was referenced Oct 3, 2026

Document WebGPU feasibility and browser training benchmarks #1240

Draft

Decentralized GPT weight exchange exceeds WebRTC message size between Chrome and Firefox #1246

Open

ale18V force-pushed the test/browser-browser-cypress-ci branch 2 times, most recently from ee165f3 to f335779 Compare October 5, 2026 21:57

ale18V changed the base branch from test/decentralized-webapp-ci to main October 5, 2026 21:57

ale18V force-pushed the test/browser-browser-cypress-ci branch from f335779 to f6cc2e0 Compare October 6, 2026 07:12

ale18V changed the title ~~test(webapp): probe Chrome–Firefox decentralized training~~ test(webapp): cover decentralized browser training across browsers Oct 6, 2026

ale18V force-pushed the test/browser-browser-cypress-ci branch 2 times, most recently from 0112c7a to 6da8cc1 Compare October 10, 2026 13:59

ale18V marked this pull request as ready for review October 10, 2026 14:01

ale18V requested a review from JulienVig October 10, 2026 14:13

ale18V added this pull request to stack #1257 October 11, 2026 10:39

ale18V added 4 commits October 12, 2026 14:25

test(webapp): cover decentralized browser training across browsers

a7e3a65

test(webapp): start standard server for browser collaboration

7921bd4

refactor(webapp): group training test support by mode

1cf6856

style(webapp): format moved federated peer

bdc2a99

ale18V force-pushed the test/browser-browser-cypress-ci branch from be145af to bdc2a99 Compare October 12, 2026 12:25

JulienVig requested changes Oct 12, 2026

View reviewed changes

test(webapp): include Edge in decentralized browser coverage

dad10c6

ale18V added 2 commits October 13, 2026 16:47

test(webapp): fix decentralized E2E browser matrix

8c862b8

test(webapp): scope decentralized training timeouts

50a5eff

Navigation Menu

Repository navigation

test(webapp): cover decentralized browser training across browsers - #1239

test(webapp): cover decentralized browser training across browsers#1239
ale18V wants to merge 7 commits into
mainfrom
test/browser-browser-cypress-ci

ale18V commented Oct 1, 2026 •

edited

Loading

Uh oh!

ale18V commented Oct 10, 2026

Uh oh!

JulienVig left a comment

Uh oh!

JulienVig Oct 12, 2026

Uh oh!

JulienVig Oct 12, 2026

Uh oh!

JulienVig Oct 12, 2026

Uh oh!

JulienVig Oct 12, 2026

Uh oh!

JulienVig Oct 12, 2026

Uh oh!

JulienVig Oct 12, 2026

Uh oh!

JulienVig Oct 12, 2026

Uh oh!

JulienVig Oct 12, 2026

Uh oh!

JulienVig Oct 12, 2026

Uh oh!

JulienVig Oct 12, 2026

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

	process.kill(-child.pid, "SIGTERM");
	process.kill(-child.pid, "SIGTERM"); // minus sign means kill both cypress and its browser

	.contains("h4", "Group label: 0")
	.contains("h4", "Group label: 0") // dummy labels for now

Conversation

ale18V commented Oct 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What & why

Technical plan

Deviations from plan

Todo

Uh oh!

ale18V commented Oct 10, 2026

Uh oh!

JulienVig left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

ale18V commented Oct 1, 2026 •

edited

Loading