Spaces:
Running
Running
Update index.html
Browse files- index.html +78 -15
index.html
CHANGED
|
@@ -3,7 +3,7 @@
|
|
| 3 |
<head>
|
| 4 |
<meta charset="UTF-8">
|
| 5 |
<meta name="viewport" content="width=device-width, initial-scale=1.0">
|
| 6 |
-
<title>
|
| 7 |
<style>
|
| 8 |
body {
|
| 9 |
font-family: Arial, sans-serif;
|
|
@@ -22,11 +22,14 @@
|
|
| 22 |
border-right: 2px solid #ee4a4f;
|
| 23 |
}
|
| 24 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 25 |
.output {
|
| 26 |
width: 80%;
|
| 27 |
padding: 20px;
|
| 28 |
overflow-y: auto;
|
| 29 |
-
font-size : 14px;
|
| 30 |
}
|
| 31 |
|
| 32 |
table {
|
|
@@ -99,12 +102,21 @@
|
|
| 99 |
button:hover {
|
| 100 |
background-color: #e60000;
|
| 101 |
}
|
| 102 |
-
|
| 103 |
-
th, td {
|
| 104 |
-
font-size: 14px;
|
| 105 |
-
}
|
| 106 |
</style>
|
| 107 |
<script>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 108 |
function calculateParameters() {
|
| 109 |
const hiddenSize = parseInt(document.getElementById('hidden_size').value);
|
| 110 |
const num_kv_heads = parseInt(document.getElementById('num_key_value_heads').value);
|
|
@@ -112,7 +124,10 @@
|
|
| 112 |
const intermediateSize = parseInt(document.getElementById('intermediate_size').value);
|
| 113 |
const vocabSize = parseInt(document.getElementById('vocab_size').value);
|
| 114 |
const numHiddenLayers = parseInt(document.getElementById('num_hidden_layers').value);
|
|
|
|
|
|
|
| 115 |
const includeBias = document.getElementById('include_bias').value === 'yes';
|
|
|
|
| 116 |
|
| 117 |
const calcTotal = (input, output) => input * output;
|
| 118 |
const calcKVdim = (hiddenSize, num_heads,num_kv_heads) => hiddenSize * num_kv_heads / num_heads;
|
|
@@ -138,25 +153,49 @@
|
|
| 138 |
const luB = includeBias ? intermediateSize : 0;
|
| 139 |
const projW = calcTotal(intermediateSize, hiddenSize);
|
| 140 |
const projB = includeBias ? hiddenSize : 0;
|
|
|
|
|
|
|
|
|
|
|
|
|
| 141 |
|
| 142 |
-
|
| 143 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 144 |
// Embedding
|
| 145 |
const embeddingTotal = calcTotal(vocabSize, hiddenSize);
|
| 146 |
|
| 147 |
// 1 layer parameters
|
| 148 |
const oneLayerParams = attentionTotal + feedForwardTotal;
|
| 149 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 150 |
// Full layers parameters
|
| 151 |
const fullLayersParams = oneLayerParams * numHiddenLayers;
|
| 152 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 153 |
// Full size (includes embedding)
|
| 154 |
const fullSize = fullLayersParams + embeddingTotal;
|
| 155 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 156 |
// Display results
|
| 157 |
const outputDiv = document.getElementById('output');
|
| 158 |
outputDiv.innerHTML = `
|
| 159 |
-
<h1>
|
| 160 |
<div class="section">
|
| 161 |
<h3>Attention</h3>
|
| 162 |
<table>
|
|
@@ -198,13 +237,13 @@
|
|
| 198 |
<th>Total (input*output)</th>
|
| 199 |
</tr>
|
| 200 |
<tr>
|
| 201 |
-
<td>Swish
|
| 202 |
<td>${hiddenSize.toLocaleString()}</td>
|
| 203 |
<td>${intermediateSize.toLocaleString()}</td>
|
| 204 |
<td>${switchW.toLocaleString()}</td>
|
| 205 |
</tr>
|
| 206 |
${includeBias ? `<tr>
|
| 207 |
-
<td>Swish
|
| 208 |
<td>-</td>
|
| 209 |
<td>${intermediateSize.toLocaleString()}</td>
|
| 210 |
<td>${switchB.toLocaleString()}</td>
|
|
@@ -237,11 +276,20 @@
|
|
| 237 |
<th colspan="3">Total Feed Forward parameters</th>
|
| 238 |
<td>${feedForwardTotal.toLocaleString()}</td>
|
| 239 |
</tr>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 240 |
</table>
|
| 241 |
</div>
|
| 242 |
<div class="highlight">
|
| 243 |
-
<strong>1 Layer Parameters (
|
| 244 |
-
<strong>Full Layers Parameters (1 layer parameters * num layers):</strong> ${fullLayersParams.toLocaleString()}<br>
|
|
|
|
|
|
|
|
|
|
|
|
|
| 245 |
</div>
|
| 246 |
<div class="section">
|
| 247 |
<h3>Embedding</h3>
|
|
@@ -261,7 +309,8 @@
|
|
| 261 |
</table>
|
| 262 |
</div>
|
| 263 |
<div class="highlight">
|
| 264 |
-
<strong>Complete Model Parmeters (embedding size + full layers size):</strong> ${fullSize.toLocaleString()}
|
|
|
|
| 265 |
</div>
|
| 266 |
`;
|
| 267 |
}
|
|
@@ -270,7 +319,7 @@
|
|
| 270 |
<body>
|
| 271 |
|
| 272 |
<div class="console">
|
| 273 |
-
<h3>Input
|
| 274 |
<label for="hidden_size">Hidden size:</label><br>
|
| 275 |
<input type="number" id="hidden_size" value="896"><br><br>
|
| 276 |
|
|
@@ -295,12 +344,26 @@
|
|
| 295 |
<option value="yes">Yes</option>
|
| 296 |
</select><br><br>
|
| 297 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 298 |
<button onclick="calculateParameters()">Calculate</button>
|
| 299 |
</div>
|
| 300 |
|
| 301 |
<div class="output" id="output">
|
| 302 |
<h1>Transformer total number of parameters Calculator</h1>
|
| 303 |
-
<h3>Enter model hyperparameters in the console and press calculate (curently working for classic transformer/LLM architecture with GQA and GLU
|
| 304 |
</div>
|
| 305 |
|
| 306 |
</body>
|
|
|
|
| 3 |
<head>
|
| 4 |
<meta charset="UTF-8">
|
| 5 |
<meta name="viewport" content="width=device-width, initial-scale=1.0">
|
| 6 |
+
<title>Model Parameter Calculator</title>
|
| 7 |
<style>
|
| 8 |
body {
|
| 9 |
font-family: Arial, sans-serif;
|
|
|
|
| 22 |
border-right: 2px solid #ee4a4f;
|
| 23 |
}
|
| 24 |
|
| 25 |
+
#additionalFieldsMOE {
|
| 26 |
+
display: none;
|
| 27 |
+
}
|
| 28 |
+
|
| 29 |
.output {
|
| 30 |
width: 80%;
|
| 31 |
padding: 20px;
|
| 32 |
overflow-y: auto;
|
|
|
|
| 33 |
}
|
| 34 |
|
| 35 |
table {
|
|
|
|
| 102 |
button:hover {
|
| 103 |
background-color: #e60000;
|
| 104 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
| 105 |
</style>
|
| 106 |
<script>
|
| 107 |
+
function handleSelectChange() {
|
| 108 |
+
// Récupère la valeur sélectionnée
|
| 109 |
+
const isMoe = document.getElementById('is_moe').value;
|
| 110 |
+
const additionalFieldsMOE = document.getElementById('additionalFieldsMOE');
|
| 111 |
+
|
| 112 |
+
// Affiche ou cache les champs supplémentaires
|
| 113 |
+
if (isMoe === 'yes') {
|
| 114 |
+
additionalFieldsMOE.style.display = 'block';
|
| 115 |
+
} else {
|
| 116 |
+
additionalFieldsMOE.style.display = 'none';
|
| 117 |
+
}
|
| 118 |
+
}
|
| 119 |
+
|
| 120 |
function calculateParameters() {
|
| 121 |
const hiddenSize = parseInt(document.getElementById('hidden_size').value);
|
| 122 |
const num_kv_heads = parseInt(document.getElementById('num_key_value_heads').value);
|
|
|
|
| 124 |
const intermediateSize = parseInt(document.getElementById('intermediate_size').value);
|
| 125 |
const vocabSize = parseInt(document.getElementById('vocab_size').value);
|
| 126 |
const numHiddenLayers = parseInt(document.getElementById('num_hidden_layers').value);
|
| 127 |
+
const active_expert_number = parseInt(document.getElementById('active_expert_number').value);
|
| 128 |
+
const expert_number = parseInt(document.getElementById('expert_number').value);
|
| 129 |
const includeBias = document.getElementById('include_bias').value === 'yes';
|
| 130 |
+
const isMoe = document.getElementById('is_moe').value === 'yes';
|
| 131 |
|
| 132 |
const calcTotal = (input, output) => input * output;
|
| 133 |
const calcKVdim = (hiddenSize, num_heads,num_kv_heads) => hiddenSize * num_kv_heads / num_heads;
|
|
|
|
| 153 |
const luB = includeBias ? intermediateSize : 0;
|
| 154 |
const projW = calcTotal(intermediateSize, hiddenSize);
|
| 155 |
const projB = includeBias ? hiddenSize : 0;
|
| 156 |
+
let feedForwardTotal;
|
| 157 |
+
|
| 158 |
+
feedForwardTotal = switchW + switchB + luW + luB + projW + projB;
|
| 159 |
+
const feedForwardTotalCache = feedForwardTotal
|
| 160 |
|
| 161 |
+
let feedForwardTotalActive;
|
| 162 |
|
| 163 |
+
if (isMoe) {
|
| 164 |
+
feedForwardTotalActive = feedForwardTotalCache * active_expert_number;
|
| 165 |
+
feedForwardTotal = feedForwardTotalCache * expert_number;
|
| 166 |
+
console.log(feedForwardTotal)
|
| 167 |
+
}
|
| 168 |
// Embedding
|
| 169 |
const embeddingTotal = calcTotal(vocabSize, hiddenSize);
|
| 170 |
|
| 171 |
// 1 layer parameters
|
| 172 |
const oneLayerParams = attentionTotal + feedForwardTotal;
|
| 173 |
|
| 174 |
+
let oneLayerParamsActive;
|
| 175 |
+
if (isMoe) {
|
| 176 |
+
oneLayerParamsActive = attentionTotal + feedForwardTotalActive;
|
| 177 |
+
}
|
| 178 |
+
|
| 179 |
// Full layers parameters
|
| 180 |
const fullLayersParams = oneLayerParams * numHiddenLayers;
|
| 181 |
|
| 182 |
+
let fullLayersParamsMOEActive;
|
| 183 |
+
if (isMoe) {
|
| 184 |
+
fullLayersParamsMOEActive = oneLayerParamsActive * numHiddenLayers;
|
| 185 |
+
}
|
| 186 |
+
|
| 187 |
// Full size (includes embedding)
|
| 188 |
const fullSize = fullLayersParams + embeddingTotal;
|
| 189 |
|
| 190 |
+
let fullSizeActive;
|
| 191 |
+
if (isMoe) {
|
| 192 |
+
fullSizeActive = fullLayersParamsMOEActive + embeddingTotal;
|
| 193 |
+
}
|
| 194 |
+
|
| 195 |
// Display results
|
| 196 |
const outputDiv = document.getElementById('output');
|
| 197 |
outputDiv.innerHTML = `
|
| 198 |
+
<h1>Model Parameter Calculator</h1>
|
| 199 |
<div class="section">
|
| 200 |
<h3>Attention</h3>
|
| 201 |
<table>
|
|
|
|
| 237 |
<th>Total (input*output)</th>
|
| 238 |
</tr>
|
| 239 |
<tr>
|
| 240 |
+
<td>Swish - W</td>
|
| 241 |
<td>${hiddenSize.toLocaleString()}</td>
|
| 242 |
<td>${intermediateSize.toLocaleString()}</td>
|
| 243 |
<td>${switchW.toLocaleString()}</td>
|
| 244 |
</tr>
|
| 245 |
${includeBias ? `<tr>
|
| 246 |
+
<td>Swish - b</td>
|
| 247 |
<td>-</td>
|
| 248 |
<td>${intermediateSize.toLocaleString()}</td>
|
| 249 |
<td>${switchB.toLocaleString()}</td>
|
|
|
|
| 276 |
<th colspan="3">Total Feed Forward parameters</th>
|
| 277 |
<td>${feedForwardTotal.toLocaleString()}</td>
|
| 278 |
</tr>
|
| 279 |
+
${isMoe ?`
|
| 280 |
+
<tr>
|
| 281 |
+
<th colspan="3">Total active Feed Forward parameters</th>
|
| 282 |
+
<td>${feedForwardTotalActive.toLocaleString()}</td>
|
| 283 |
+
</tr>` : ''}
|
| 284 |
</table>
|
| 285 |
</div>
|
| 286 |
<div class="highlight">
|
| 287 |
+
<strong>1 Layer Parameters (Attention + FFN):</strong> ${oneLayerParams.toLocaleString()}<br>
|
| 288 |
+
<strong>Full Layers Parameters (1 layer parameters * num layers):</strong> ${fullLayersParams.toLocaleString()}<br><br>
|
| 289 |
+
${isMoe ? `<tr>
|
| 290 |
+
<strong>1 Layer Parameters Active(Attention + FFN):</strong> ${oneLayerParamsActive.toLocaleString()}<br>
|
| 291 |
+
<strong>Full Layers Parameters Active(1 layer active parameters * num layers):</strong> ${fullLayersParamsMOEActive.toLocaleString()}<br>
|
| 292 |
+
</tr>` : ''}
|
| 293 |
</div>
|
| 294 |
<div class="section">
|
| 295 |
<h3>Embedding</h3>
|
|
|
|
| 309 |
</table>
|
| 310 |
</div>
|
| 311 |
<div class="highlight">
|
| 312 |
+
<strong>Complete Model Parmeters (embedding size + full layers size):</strong> ${fullSize.toLocaleString()}<br><br>
|
| 313 |
+
${isMoe ? `<strong>Complete Model Parmeters Active:</strong> ${fullSizeActive.toLocaleString()}` : ''}
|
| 314 |
</div>
|
| 315 |
`;
|
| 316 |
}
|
|
|
|
| 319 |
<body>
|
| 320 |
|
| 321 |
<div class="console">
|
| 322 |
+
<h3>Input Parameters</h3>
|
| 323 |
<label for="hidden_size">Hidden size:</label><br>
|
| 324 |
<input type="number" id="hidden_size" value="896"><br><br>
|
| 325 |
|
|
|
|
| 344 |
<option value="yes">Yes</option>
|
| 345 |
</select><br><br>
|
| 346 |
|
| 347 |
+
<label for="is_moe">Is MOE ?</label><br>
|
| 348 |
+
<select id="is_moe" onchange="handleSelectChange()">
|
| 349 |
+
<option value="no">No</option>
|
| 350 |
+
<option value="yes">Yes</option>
|
| 351 |
+
</select><br><br>
|
| 352 |
+
|
| 353 |
+
<div id="additionalFieldsMOE">
|
| 354 |
+
<label for="expert_number">Total expert number :</label><br>
|
| 355 |
+
<input type="number" id="expert_number" name="expert_number"><br><br>
|
| 356 |
+
|
| 357 |
+
<label for="active_expert_number">Total active experts (shared + specifics):</label><br>
|
| 358 |
+
<input type="number" id="active_expert_number" name="active_expert_number"><br><br>
|
| 359 |
+
</div>
|
| 360 |
+
|
| 361 |
<button onclick="calculateParameters()">Calculate</button>
|
| 362 |
</div>
|
| 363 |
|
| 364 |
<div class="output" id="output">
|
| 365 |
<h1>Transformer total number of parameters Calculator</h1>
|
| 366 |
+
<h3>Enter model hyperparameters in the console and press calculate (curently working for classic transformer/LLM architecture with GQA and GLU)</h3>
|
| 367 |
</div>
|
| 368 |
|
| 369 |
</body>
|