modules.qwen2_moe.modeling_qwen2_moe_flax
FlaxQwen2MoeAttention
Bases: BaseJAXAttentionModule
Source code in src/python/easydel/modules/qwen2_moe/modeling_qwen2_moe_flax.py
163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 |
|
__call__(hidden_states, freq_cis, attention_mask, position_ids, causal_mask, segment_ids=None, deterministic=True, init_cache=False, output_attentions=False, fcm_mask=None)
The call function is the main function of a JAX module. It defines how the module behaves when called with inputs. The call function can be thought of as a "forward pass" through the model, and it should return all outputs that are needed for training or inference.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
self |
Access variables that belong to the class |
required | |
hidden_states |
Array
|
chex.Array: Pass the hidden states of the previous layer |
required |
freq_cis |
Tuple[Array, Array]
|
Tuple[chex.Array, chex.Array],: Pass in the frequency coefficients for each position |
required |
attention_mask |
Array
|
chex.Array: Mask out certain tokens in the input sequence |
required |
position_ids |
Array
|
chex.Array: Determine the position of each token in a sequence |
required |
causal_mask |
Array
|
chex.Array: Mask out the future tokens in the decoder |
required |
deterministic |
bool
|
bool: Determine whether to use dropout or not |
True
|
init_cache |
bool
|
bool: Initialize the cache |
False
|
output_attentions |
bool
|
bool: Determine whether to return the attention weights or not |
False
|
fcm_mask |
Mask out the attention weights between the input and output tokens |
None
|
|
|
Determine if the attention is causal or not |
required |
Returns:
Type | Description |
---|---|
A tuple of two arrays |
Source code in src/python/easydel/modules/qwen2_moe/modeling_qwen2_moe_flax.py
305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 |
|
apply_rotary(batch_size, sequence_length, query, key, value, freq_cis, position_ids)
The apply_rotary function is a modified version of the apply_attention function in the BertModel class. The main difference is that it takes in an additional argument, freq_cis, which are used to calculate the rotary attention weights. The other differences are minor and mostly related to reshaping tensors.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
self |
Access variables that belong to the class |
required | |
batch_size |
Reshape the query, key and value tensors |
required | |
sequence_length |
Reshape the query, key and value tensors |
required | |
query |
Calculate the attention weights |
required | |
key |
Calculate the attention |
required | |
value |
Compute the attention weights |
required | |
freq_cis |
Calculate the frequency of each word in the vocabulary |
required | |
position_ids |
Identify the position of each token in the sequence |
required |
Returns:
Type | Description |
---|---|
A tuple of 3 tensors: query, key and value |
Source code in src/python/easydel/modules/qwen2_moe/modeling_qwen2_moe_flax.py
276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 |
|
FlaxQwen2MoeBlock
Bases: Module
Source code in src/python/easydel/modules/qwen2_moe/modeling_qwen2_moe_flax.py
596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 |
|
__call__(hidden_states, freq_cis, attention_mask, position_ids, causal_mask, deterministic=True, init_cache=False, output_attentions=False, output_hidden_states=False, output_router_logits=None, return_dict=True, segment_ids=None, fcm_mask=None)
The call function is the main function of a TransformerEncoderLayer. It takes in hidden states, frequency-domain inputs, and masks as input. It then applies self-attention to the hidden states using those inputs and returns an output tensor with shape (batch_size, sequence_length, model_dim).
Parameters:
Name | Type | Description | Default |
---|---|---|---|
self |
Refer to the class instance itself |
required | |
hidden_states |
Array
|
chex.Array: Pass in the hidden state of the previous layer |
required |
freq_cis |
Tuple[Array, Array]
|
Tuple[chex.Array, chex.Array],: Pass in the frequency information |
required |
attention_mask |
Array
|
chex.Array: Mask out the attention weights for padding tokens |
required |
position_ids |
Array
|
chex.Array: Determine the position of each token in the sequence |
required |
causal_mask |
Array
|
chex.Array: Mask the attention weights |
required |
deterministic |
bool
|
bool: Control whether the dropout is applied or not |
True
|
init_cache |
bool
|
bool: Initialize the cache in the attention layer |
False
|
output_attentions |
Optional[bool]
|
bool: Return the attention weights |
False
|
fcm_mask |
Optional[ndarray]
|
Optional[jnp.ndarray]: Mask the self-attention |
None
|
|
Control the dropout in the self attention layer |
required |
Returns:
Type | Description |
---|---|
A tuple of two items |
Source code in src/python/easydel/modules/qwen2_moe/modeling_qwen2_moe_flax.py
647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 |
|
FlaxQwen2MoeBlockCollection
Bases: Module
Source code in src/python/easydel/modules/qwen2_moe/modeling_qwen2_moe_flax.py
927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 |
|
__call__(hidden_states, freq_cis, attention_mask, position_ids, causal_mask, deterministic=True, init_cache=False, output_attentions=False, output_hidden_states=False, output_router_logits=None, return_dict=True)
The call function is the main function of a JAX nn.Module. It defines how the module behaves when called as a function, and it's what you'll use to call your model in training loops or inference scripts. The call method should take all inputs that are necessary for computing outputs from the module, and return all outputs that are computed by this module.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
self |
Represent the instance of the class |
required | |
hidden_states |
Array
|
chex.Array: Pass the input tensor to the encoder |
required |
freq_cis |
Tuple[Array, Array]
|
Tuple[chex.Array, chex.Array],: Pass in the frequency of each token |
required |
attention_mask |
Array
|
chex.Array: Mask out certain tokens in the input sequence |
required |
position_ids |
Array
|
chex.Array: Specify the position of each token in a sequence |
required |
causal_mask |
Array
|
chex.Array: Mask the attention weights |
required |
deterministic |
bool
|
bool: Determine whether the model is in training or evaluation mode |
True
|
init_cache |
bool
|
bool: Initialize the cache for each layer |
False
|
output_attentions |
Optional[bool]
|
bool: Determine whether to output the attention weights |
False
|
output_hidden_states |
Optional[bool]
|
bool: Determine whether to return the hidden states of each layer |
False
|
return_dict |
bool
|
bool: Return a dictionary of the outputs |
True
|
|
Determine whether to use the forgetful causal mask |
required |
Returns:
Type | Description |
---|---|
A tuple of 3 values |
Source code in src/python/easydel/modules/qwen2_moe/modeling_qwen2_moe_flax.py
947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 |
|
FlaxQwen2MoeForCausalLM
Bases: FlaxQwen2MoePreTrainedModel
Source code in src/python/easydel/modules/qwen2_moe/modeling_qwen2_moe_flax.py
1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 |
|
prepare_inputs_for_generation(input_ids, max_length, attention_mask=None)
The prepare_inputs_for_generation function is used to prepare the inputs for a generation task.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
self |
Access variables that belong to the class |
required | |
input_ids |
Pass in the input tokens |
required | |
max_length |
Set the length of the sequence to be generated |
required | |
attention_mask |
Optional[Array]
|
Optional[chex.Array]: Mask the attention weights |
None
|
Returns:
Type | Description |
---|---|
A dictionary of the past_key_values, attention_mask and position ids |
Source code in src/python/easydel/modules/qwen2_moe/modeling_qwen2_moe_flax.py
1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 |
|
FlaxQwen2MoeForCausalLMModule
Bases: Module
Source code in src/python/easydel/modules/qwen2_moe/modeling_qwen2_moe_flax.py
1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 |
|
__call__(input_ids, attention_mask=None, position_ids=None, deterministic=True, init_cache=False, output_attentions=None, output_hidden_states=None, output_router_logits=None, return_dict=True, extra_embedding=None)
The call function is the main function of a Flax module. It takes in inputs and returns outputs.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
self |
Refer to the object itself |
required | |
input_ids |
Array
|
chex.Array: Pass the input token ids to the model |
required |
attention_mask |
Array
|
chex.Array: Mask out the padding tokens |
None
|
position_ids |
Array
|
chex.Array: Specify the position of each token in the input sequence |
None
|
deterministic |
bool
|
bool: Control whether the model is trained or not |
True
|
init_cache |
bool
|
bool: Initialize the cache for the decoder |
False
|
output_attentions |
Optional[bool]
|
bool: Return the attention weights |
None
|
output_hidden_states |
Optional[bool]
|
bool: Determine whether to return the hidden states |
None
|
return_dict |
bool
|
bool: Return a dictionary of the outputs or not |
True
|
extra_embedding |
Optional[Union[ndarray, None]]
|
Optional[Union[jnp.ndarray: Pass in the embedding of the word that we want to predict |
None
|
None]] |
Pass in the extra embedding |
required |
Returns:
Type | Description |
---|---|
The logits and the hidden states |
Source code in src/python/easydel/modules/qwen2_moe/modeling_qwen2_moe_flax.py
1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 |
|
FlaxQwen2MoeForSequenceClassificationModule
Bases: Module
Source code in src/python/easydel/modules/qwen2_moe/modeling_qwen2_moe_flax.py
1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 |
|
__call__(input_ids, attention_mask=None, position_ids=None, deterministic=True, init_cache=False, output_attentions=False, output_hidden_states=False, return_dict=True, extra_embedding=None)
The call function is the main function of a Flax module. It takes in all the inputs to the model and returns all outputs from it. The call function can be called directly on an instance of a class, or by using parentheses after an instance: >>> my_model = MyModel() # instantiate your model class >>> output = my_model(input) # call your model with input data as arguments to call
Parameters:
Name | Type | Description | Default |
---|---|---|---|
self |
Refer to the class instance |
required | |
input_ids |
Array
|
chex.Array: Pass the input to the model |
required |
attention_mask |
Array
|
chex.Array: Specify which tokens are masked |
None
|
position_ids |
Array
|
chex.Array: Specify the position of each token in the sequence |
None
|
deterministic |
bool
|
bool: Control whether the model is run in deterministic or stochastic mode |
True
|
init_cache |
bool
|
bool: Initialize the cache for the transformer |
False
|
output_attentions |
bool
|
bool: Return the attention weights |
False
|
output_hidden_states |
bool
|
bool: Return the hidden states of all layers |
False
|
return_dict |
bool
|
bool: Return a dictionary of outputs |
True
|
extra_embedding |
Optional[Union[ndarray, None]]
|
Optional[Union[jnp.ndarray: Pass in the embedding of a new word |
None
|
None]] |
Pass the extra embedding to the model |
required |
Returns:
Type | Description |
---|---|
A tuple of logits and hidden_states |
Source code in src/python/easydel/modules/qwen2_moe/modeling_qwen2_moe_flax.py
1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 |
|
setup()
The setup function is called once at the beginning of training. It initializes the model and optimizer, and sets up any other state that needs to be initialized.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
self |
Access variables that belong to the class |
required |
Returns:
Type | Description |
---|---|
A tuple of the model and the classifier |
Source code in src/python/easydel/modules/qwen2_moe/modeling_qwen2_moe_flax.py
1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 |
|
FlaxQwen2MoeMLP
Bases: Module
Source code in src/python/easydel/modules/qwen2_moe/modeling_qwen2_moe_flax.py
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 |
|
__call__(x, deterministic=True)
The call function is the main function of a class. It is called when an instance of the class (an object) is invoked as a function, i.e., obj(arguments). The call method enables instances of a class to be called like standard Python functions.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
self |
Represent the instance of the class |
required | |
x |
ndarray
|
jnp.ndarray: Pass in the input to the layer |
required |
deterministic |
bool
|
bool: Determine whether to use dropout |
True
|
Returns:
Type | Description |
---|---|
ndarray
|
A tensor that is the result of applying a dropout function to x |
Source code in src/python/easydel/modules/qwen2_moe/modeling_qwen2_moe_flax.py
147 148 149 150 151 152 153 154 155 156 157 158 159 160 |
|
FlaxQwen2MoePreTrainedModel
Bases: EasyDeLFlaxPretrainedModel
Source code in src/python/easydel/modules/qwen2_moe/modeling_qwen2_moe_flax.py
717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 |
|
__call__(input_ids, attention_mask=None, position_ids=None, params=None, past_key_values=None, dropout_rng=None, train=False, output_attentions=None, output_hidden_states=None, output_router_logits=None, return_dict=None, extra_embedding=None, add_params_field=False, **kwargs)
The call function is the main function of a JAX module. It takes in inputs and returns outputs, but it also has some other important features: - It can take in mutable state (e.g., past_key_values) that will be updated during the call and returned at the end. - It can take in random number generators (rngs) that are used to generate random numbers for dropout or sampling operations.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
self |
Represent the instance of the class |
required | |
input_ids |
Array
|
chex.Array: Pass in the input tokens |
required |
attention_mask |
Array
|
chex.Array: Mask out certain tokens in the input |
None
|
position_ids |
Array
|
chex.Array: Create the positional embeddings |
None
|
params |
dict
|
dict: Pass in the parameters of the model |
None
|
past_key_values |
dict
|
dict: Pass in the past key values from a previous call to call |
None
|
dropout_rng |
PRNGKey
|
jax.random.PRNGKey: Make sure that the dropout is applied in a random way |
None
|
train |
bool
|
bool: Determine whether to use dropout or not |
False
|
output_attentions |
Optional[bool]
|
Optional[bool]: Determine whether to return the attention weights |
None
|
output_hidden_states |
Optional[bool]
|
Optional[bool]: Return the hidden states of all layers |
None
|
return_dict |
Optional[bool]
|
Optional[bool]: Determine whether to return a dictionary or not |
None
|
extra_embedding |
Optional[Union[ndarray, None]]
|
Optional[Union[jnp.ndarray,None]]: Pass in the embedding for the input_ids |
None
|
add_params_field |
bool
|
bool: Add the params field to the inputs dictionary |
False
|
Returns:
Type | Description |
---|---|
A tuple of the following: |
Source code in src/python/easydel/modules/qwen2_moe/modeling_qwen2_moe_flax.py
820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 |
|
__init__(config, input_shape=(1, 1), seed=0, dtype=jnp.float32, _do_init=True, **kwargs)
The init function is called when the class is instantiated. It sets up the instance of the class, and defines what happens when it's created. The init function can take arguments, but self is always required (it refers to the instance of the object).
Parameters:
Name | Type | Description | Default |
---|---|---|---|
self |
Refer to the object itself |
required | |
config |
Qwen2MoeConfig
|
Qwen2MoeConfig: Pass the configuration to the module |
required |
input_shape |
Tuple
|
Tuple: Specify the shape of the input to the model |
(1, 1)
|
seed |
int
|
int: Set the seed for random number generation |
0
|
dtype |
dtype
|
jnp.dtype: Specify the data type of the input |
float32
|
_do_init |
bool
|
bool: Control whether the module is initialized or not |
True
|
kwargs |
Pass in any additional parameters that the module_class might need |
{}
|
|
|
Specify the number of layers in the network |
required |
Returns:
Type | Description |
---|---|
The super() of the class |
Source code in src/python/easydel/modules/qwen2_moe/modeling_qwen2_moe_flax.py
722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 |
|
init_cache(batch_size, max_length)
The init_cache function is used to initialize the cache for a given batch size and sequence length. The cache is a dictionary that contains all the intermediate states from each layer in the model. This allows us to run inference on multiple batches without having to re-run forward passes through every layer in the model, which would be very slow.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
self |
Access the module |
required | |
batch_size |
Define the batch size of the input tensors |
required | |
max_length |
Set the length of the input sequence |
required |
Returns:
Type | Description |
---|---|
A dictionary with the following keys: |
Source code in src/python/easydel/modules/qwen2_moe/modeling_qwen2_moe_flax.py
797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 |
|
init_weights(rng, input_shape, params=None)
The init_weights function is used to initialize the weights of a model.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
self |
Access variables that belong to the class |
required | |
rng |
PRNGKey
|
jax.random.PRNGKey: Initialize the weights of the model |
required |
input_shape |
Tuple
|
Tuple: Specify the shape of the input tensor |
required |
params |
FrozenDict
|
FrozenDict: Pass in the parameters of a pre-trained model |
None
|
Returns:
Type | Description |
---|---|
FrozenDict
|
A frozendict of parameters |
Source code in src/python/easydel/modules/qwen2_moe/modeling_qwen2_moe_flax.py
751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 |
|
FlaxQwen2MoeSparseMoeBlock
Bases: Module
This implementation is strictly equivalent to standard MoE with full capacity (no dropped tokens). It's faster since it formulates MoE operations in terms of block-sparse operations to accomodate imbalanced assignments of tokens to experts, whereas standard MoE either (1) drop tokens at the cost of reduced performance or (2) set capacity factor to number of experts and thus waste computation and memory on padding.
Source code in src/python/easydel/modules/qwen2_moe/modeling_qwen2_moe_flax.py
502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 |
|