Skip to content

Stabilize WebGPU text generation for Gemma-3 ONNX #289

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

Open
TanmayThakur2209 wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Stabilize WebGPU text generation for Gemma-3 ONNX #289

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
18636dc
Stabilize WebGPU text generation for Gemma3 ONNX
TanmayThakur2209 Jan 17, 2026
89c6187
Refactor WebGPU max token warning to avoid magic number
TanmayThakur2209 Jan 17, 2026
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
34 changes: 23 additions & 11 deletions Demos/Gemma3-on-Web/src/worker.js
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -28,7 +28,19 @@ class TextGenerationPipeline {

const stopping_criteria = new InterruptableStoppingCriteria();

let past_key_values_cache = null;
/**
* WebGPU-safe generation defaults.
* Conservative values chosen to avoid instability
* with quantized ONNX models in browser environments.
*/
const SAFE_WEBGPU_GENERATION_CONFIG = {
max_new_tokens: 256,
do_sample: true,
temperature: 0.4,
top_p: 0.85,
repetition_penalty: 1.1,
use_cache: false, // 🔴 Critical for stability
};
/**
* Generate text based on the input messages
*/
Expand Down Expand Up @@ -69,21 +81,22 @@ async function generate(messages) {

// Tell the main thread we are starting
self.postMessage({ status: "start" });

const MAX_TOKENS_WARNING_THRESHOLD = 256;

const { past_key_values, sequences } = await model.generate({
...inputs,
past_key_values: past_key_values_cache,

// Sampling
do_sample: false,
temperature: 0.3,
if (SAFE_WEBGPU_GENERATION_CONFIG.max_new_tokens > MAX_TOKENS_WARNING_THRESHOLD) {
console.warn(
"[Gemma WebGPU] Large max_new_tokens may cause unstable output. Consider reducing to ≤256."
);
}

max_new_tokens: 512,
const { sequences } = await model.generate({
...inputs,
...SAFE_WEBGPU_GENERATION_CONFIG,
streamer,
stopping_criteria,
return_dict_in_generate: true,
});
past_key_values_cache = past_key_values;

const decoded = tokenizer.batch_decode(sequences, {
skip_special_tokens: true,
Expand Down Expand Up @@ -163,7 +176,6 @@ self.addEventListener("message", async (e) => {
break;

case "reset":
past_key_values_cache = null;
stopping_criteria.reset();
break;
}
Expand Down