ort-web-benchmark/text_encoder_2_npu_test.html at main · Honry/ort-web-benchmark · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
<!DOCTYPE html>
<html lang="en">
  <head>
    <meta charset="UTF-8" />
    <meta http-equiv="X-UA-Compatible" content="IE=edge" />
    <meta name="viewport" content="width=device-width, initial-scale=1.0" />
    <title>Test Text Encoder 2 on NPU</title>
  </head>
  <style>
    body {
      font-family: sans-serif;
      padding: 20px;
    }

    h1 {
      color: #425066;
      font-size: 31px;
      margin-top: 0;
    }

    .loading-stats {
      color: #aaa;
      font-size: 12px;
      margin-top: -12px;
    }

    .hide {
      display: none;
    }

    .content {
      margin-top: 30px;
    }

    div {
      margin-top: 20px;
    }
  </style>

  <body>
    <h1>Test Text Encoder 2 on NPU</h1>

    <!-- Loading status -->
    <div class="loading-stats">Choose options then click 'Run'...</div>
    <div>
      DeviceType:
      <select id="deviceType">
        <option value="npu" selected>WebNN NPU</option>
        <option value="gpu">WebNN GPU</option>
        <option value="cpu">WebNN CPU</option>
      </select>
    </div>
    <div>
      <input type="button" value="Run" id="run" />
    </div>
    <div id="status" style="font: 1em sans-serif"></div>
    <script src="https://cdn.jsdelivr.net/npm/onnxruntime-web@1.24.0-dev.20251104-75d35474d5/dist/ort.webgpu.min.js"></script>
    <script type="module">
      import {
        AutoTokenizer,
        env,
      } from "https://cdn.jsdelivr.net/npm/@xenova/transformers/dist/transformers.js";
      env.localModelPath = "models/";
      env.allowRemoteModels = true;
      env.allowLocalModels = true;
      function log(i) {
        console.log(i);
        document.getElementById("status").innerText +=
          `\n[${performance.now().toFixed(3)}] ` + i;
      }

      ort.env.wasm.numThreads = 4;
      ort.env.wasm.simd = true;
      ort.env.wasm.proxy = false;

      async function run() {
        runBtn.disabled = true;
        let feed = {};
        const deviceType = document.getElementById("deviceType").value;
        const modelUrl =
          "https://huggingface.co/webnn/sdxl-turbo/resolve/main/onnx/text_encoder_2_model_qdq_q4f16.onnx";
        const modelBuffer = await getModelOPFS(
          "text_encoder_2_model_qdq_q4f16",
          modelUrl,
          false
        );
        log("entering run ...");
        try {
          log("creating session ...");
          const options = {
            executionProviders: [
              {
                name: "webnn",
                deviceType: deviceType,
              },
            ],
            logSeverityLevel: 0,
            freeDimensionOverrides: {
              batch_size: 1,
              sequence_length: 77,
            },
          };
          console.log("sessionOptions: ", options);
          const sess_start = performance.now();
          const sess = await ort.InferenceSession.create(modelBuffer, options);
          log(
            "session created in " +
              (performance.now() - sess_start).toFixed(2) +
              " ms"
          );
          log("running ...");

          // inputInfo = {
          //     input_ids: { dataType: "int64", dims: [1, 77] },
          // };
          // outputInfo = {
          //     "hidden_states.31": { dataType: "float16", dims: [1, 77, 1280] },
          //     text_embeds: { dataType: "float16", dims: [1, 1280] },
          // };

          const tokenizer2 = await AutoTokenizer.from_pretrained("tokenizer_2");
          const prompt =
            "An artistic baby raccoon DJ in a vintage suit, adjusting knobs on a futuristic mixer. The scene is a dim nightclub with vibrant lights and soft bokeh. Highly detailed, cinematic style.";
          const { input_ids: inputIds2 } = await tokenizer2(prompt, {
            padding: "max_length",
            maxLength: 77,
            truncation: true,
            return_tensor: false,
          });

          const inputIds2Data = new BigInt64Array(
            inputIds2.map((x) => BigInt(x))
          );
          feed = {
            input_ids: new ort.Tensor("int64", inputIds2Data, [1, 77]),
          };
          console.log("inputs: ", feed);

          const outputs = await sess.run(feed);

          console.log("outputs: ", outputs);

          const hs_result = `output hidden_states.31 first 10 values: [${outputs[
            "hidden_states.31"
          ].data.slice(0, 10)}]`;
          log(hs_result);
          const te_result = `output text_embeds first 10 values: [${outputs[
            "text_embeds"
          ].data.slice(0, 10)}]`;
          log(te_result);

          await sess.release();
          runBtn.disabled = false;
        } catch (e) {
          log(e);
        }
      }
      // Get model via Origin Private File System
      async function getModelOPFS(name, url, updateModel) {
        const root = await navigator.storage.getDirectory();
        let fileHandle;

        async function updateFile() {
          const response = await fetch(url);
          const buffer = await readResponse(response);
          fileHandle = await root.getFileHandle(name, { create: true });
          const writable = await fileHandle.createWritable();
          await writable.write(buffer);
          await writable.close();
          return buffer;
        }

        if (updateModel) {
          return await updateFile();
        }

        try {
          fileHandle = await root.getFileHandle(name);
          const blob = await fileHandle.getFile();
          return await blob.arrayBuffer();
        } catch (e) {
          return await updateFile();
        }
      }

      async function readResponse(response) {
        const contentLength = response.headers.get("Content-Length");
        let total = parseInt(contentLength ?? "0");
        let buffer = new Uint8Array(total);
        let loaded = 0;

        const reader = response.body.getReader();
        async function read() {
          const { done, value } = await reader.read();
          if (done) return;

          let newLoaded = loaded + value.length;
          if (newLoaded > total) {
            total = newLoaded;
            let newBuffer = new Uint8Array(total);
            newBuffer.set(buffer);
            buffer = newBuffer;
          }
          buffer.set(value, loaded);
          loaded = newLoaded;
          return read();
        }

        await read();
        return buffer;
      }
      const runBtn = document.getElementById("run");
      runBtn.onclick = async () => {
        await run();
      };
    </script>
  </body>
</html>