Xenova's picture
Xenova HF staff
New features
377005e verified
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8" />
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<title>Transformers.js | WebGPU Benchmark</title>
<script type="module" crossorigin src="/assets/index-D9dUKbnn.js"></script>
<link rel="stylesheet" crossorigin href="/assets/index-BZcCtzqA.css">
</head>
<body>
<h1>
<a href="http://github.com/xenova/transformers.js" target="_blank">🤗 Transformers.js</a> WebGPU Benchmark
</h1>
<p>
This benchmark measures the execution time of BERT-based embedding models
using the WASM and WebGPU execution providers across different batch sizes.
</p>
<div id="chart-container">
<canvas id="chart"></canvas>
</div>
<div>
<button id="start" disabled>Start Benchmark</button>
<button id="stop" disabled>Stop Benchmark</button>
</div>
<label id="status"></label>
<details open>
<summary>Options</summary>
<div>
<input class="tests" type="checkbox" value="WASM (int8)" data-color="33,150,243" data-device="wasm"
data-dtype="int8"> WASM (int8)<br />
<input class="tests" type="checkbox" value="WASM (fp16)" data-color="63,81,181" data-device="wasm"
data-dtype="fp16"> WASM (fp16)<br />
<input class="tests" type="checkbox" value="WASM (fp32)" data-color="46,204,113" data-device="wasm"
data-dtype="fp32" checked> WASM (fp32)<br />
<!-- <input class="tests" type="checkbox" value="WebGPU (int8)" data-color="233,30,99" data-device="webgpu"
data-dtype="int8"> WebGPU (int8)<br /> -->
<input class="tests" type="checkbox" value="WebGPU (fp16)" data-color="255,193,7" data-device="webgpu"
data-dtype="fp16"> WebGPU (fp16)<br />
<input class="tests" type="checkbox" value="WebGPU (fp32)" data-color="0,150,136" data-device="webgpu"
data-dtype="fp32" checked> WebGPU (fp32)<br />
</div>
<hr />
<div>
<label>Model ID</label>
<input id="model-id" value="Xenova/all-MiniLM-L6-v2" />
</div>
<div>
<label>Batch sizes</label>
<input id="batch-sizes" value="1, 2, 4, 8, 16, 32" />
</div>
<div>
<label>Sequence length</label>
<input id="sequence-length" type="number" min="1" max="512" value="512" />
</div>
<hr />
<div>
<input id="x-scale" type="checkbox" /> Log scale (x) <br />
<input id="y-scale" type="checkbox" /> Log scale (y) <br />
</div>
</details>
</body>
</html>