Skip to content

Benchmark: async GPU decode via next_token_async (flare 0.2.15) #316

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
sauravpanda merged 1 commit into from
Apr 23, 2026
Merged

Benchmark: async GPU decode via next_token_async (flare 0.2.15) #316

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
20 changes: 12 additions & 8 deletions examples/benchmark/index.html
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -597,7 +597,7 @@ <h2>Comparison Charts</h2>
if (!flareLib) {
log('Loading @sauravpanda/flare WASM from CDN...', 'info');
T('fetching flare_web.js from CDN');
const CDN = 'https://cdn.jsdelivr.net/npm/@sauravpanda/flare@0.2.13/pkg';
const CDN = 'https://cdn.jsdelivr.net/npm/@sauravpanda/flare@0.2.15/pkg';
const wasmUrl = `${CDN}/flare_web_bg.wasm`;

const jsResp = await fetch(`${CDN}/flare_web.js`, { cache: 'no-cache' });
Expand Down Expand Up @@ -724,12 +724,12 @@ <h2>Comparison Charts</h2>
T('FlareEngine.load returned');
bytes = null; // drop source buffer before warmup / GPU upload

// GPU prefill is currently deadlocked on Chrome main-thread wasm due
// to wgpu's sync readback pattern (map_async callback can't fire
// while we're mid-sync-WASM-call). Force CPU path until the async
// readback refactor lands upstream.
const USE_GPU = new URL(location.href).searchParams.get('gpu') === '1';
T(`init_gpu: ${USE_GPU ? 'ENABLED via ?gpu=1' : 'SKIPPED (CPU only)'}`);
// GPU decode now works via `next_token_async` (flare-web 0.2.14).
// Prefill still runs sync → uses the CPU backend's Q8_0 SIMD path;
// decode reads back through the async readback so the WebGPU
// `map_async` callback can fire. `?gpu=0` opts out for debugging.
const USE_GPU = new URL(location.href).searchParams.get('gpu') !== '0';
T(`init_gpu: ${USE_GPU ? 'ENABLED (async decode)' : 'SKIPPED via ?gpu=0'}`);
try {
const gpuOk = USE_GPU ? await flareEngine.init_gpu() : false;
T(`init_gpu complete, gpuOk=${gpuOk}`);
Expand Down Expand Up @@ -792,8 +792,12 @@ <h2>Comparison Charts</h2>
let output = '';

tInf('entering decode loop');
// Prefer the async variant when available (flare-web >= 0.2.14) so
// WebGPU's map_async callback can fire between tokens. Sync
// next_token is only safe on CPU backend.
const hasAsync = typeof flareEngine.next_token_async === 'function';
while (!flareEngine.stream_done) {
const id = flareEngine.next_token();
const id = hasAsync ? await flareEngine.next_token_async() : flareEngine.next_token();
if (id === undefined) break;
tokenCount++;
output += flareEngine.decode_token_chunk(id);
Expand Down
Loading