Spaces:
Paused
Paused
File size: 69,192 Bytes
a080fe0 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 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 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 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 1282 1283 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 1339 1340 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 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 |
# Copyright 2020-2025 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import pathlib
import tempfile
import unittest
import numpy as np
import torch
from datasets import Dataset, Image, Sequence, load_dataset
from parameterized import parameterized
from transformers import (
AutoModelForCausalLM,
AutoProcessor,
AutoTokenizer,
LlavaForConditionalGeneration,
is_vision_available,
)
from transformers.testing_utils import require_flash_attn, require_peft, require_vision
from transformers.utils import is_peft_available
from trl import SFTConfig, SFTTrainer
from trl.trainer.sft_trainer import DataCollatorForLanguageModeling
if is_peft_available():
from peft import LoraConfig, PeftModel, get_peft_model
if is_vision_available():
from PIL import Image as PILImage
def formatting_prompts_func(example):
text = f"### Question: {example['question']}\n ### Answer: {example['answer']}"
return text
def formatting_func_for_pretokenized(example):
return example["input_ids"]
class TestDataCollatorForLanguageModeling(unittest.TestCase):
def test_basic_padding(self):
"""Test basic padding functionality without completion masks."""
self.collator = DataCollatorForLanguageModeling(pad_token_id=0)
examples = [{"input_ids": [1, 2, 3]}, {"input_ids": [4, 5]}]
result = self.collator(examples)
torch.testing.assert_close(result["input_ids"], torch.tensor([[1, 2, 3], [4, 5, 0]]))
torch.testing.assert_close(result["attention_mask"], torch.tensor([[1, 1, 1], [1, 1, 0]]))
torch.testing.assert_close(result["position_ids"], torch.tensor([[0, 1, 2], [0, 1, 0]]))
torch.testing.assert_close(result["labels"], torch.tensor([[1, 2, 3], [4, 5, -100]]))
def test_completion_mask(self):
"""Test completion mask functionality."""
self.collator = DataCollatorForLanguageModeling(pad_token_id=0)
examples = [
{"input_ids": [1, 2, 3], "completion_mask": [0, 1, 1]},
{"input_ids": [4, 5], "completion_mask": [0, 1]},
]
result = self.collator(examples)
torch.testing.assert_close(result["input_ids"], torch.tensor([[1, 2, 3], [4, 5, 0]]))
torch.testing.assert_close(result["attention_mask"], torch.tensor([[1, 1, 1], [1, 1, 0]]))
torch.testing.assert_close(result["position_ids"], torch.tensor([[0, 1, 2], [0, 1, 0]]))
torch.testing.assert_close(result["labels"], torch.tensor([[-100, 2, 3], [-100, 5, -100]]))
def test_completion_only_loss_disabled(self):
"""Test behavior when completion_only_loss is disabled."""
collator = DataCollatorForLanguageModeling(pad_token_id=0, completion_only_loss=False)
examples = [
{"input_ids": [1, 2, 3], "completion_mask": [0, 1, 1]},
{"input_ids": [4, 5], "completion_mask": [0, 1]},
]
result = collator(examples)
# Labels should not be masked when completion_only_loss=False
torch.testing.assert_close(result["input_ids"], torch.tensor([[1, 2, 3], [4, 5, 0]]))
torch.testing.assert_close(result["attention_mask"], torch.tensor([[1, 1, 1], [1, 1, 0]]))
torch.testing.assert_close(result["position_ids"], torch.tensor([[0, 1, 2], [0, 1, 0]]))
torch.testing.assert_close(result["labels"], torch.tensor([[1, 2, 3], [4, 5, -100]]))
def test_padding_free_mode(self):
"""Test padding-free mode where sequences are concatenated."""
collator = DataCollatorForLanguageModeling(pad_token_id=0, padding_free=True)
examples = [{"input_ids": [1, 2, 3]}, {"input_ids": [4, 5]}]
result = collator(examples)
torch.testing.assert_close(result["input_ids"], torch.tensor([[1, 2, 3, 4, 5]]))
torch.testing.assert_close(result["attention_mask"], torch.tensor([[1, 1, 1, 1, 1]]))
torch.testing.assert_close(result["position_ids"], torch.tensor([[0, 1, 2, 0, 1]]))
torch.testing.assert_close(result["labels"], torch.tensor([[1, 2, 3, 4, 5]]))
def test_padding_free_with_completion_mask(self):
"""Test padding-free mode with completion masks."""
collator = DataCollatorForLanguageModeling(pad_token_id=0, padding_free=True)
examples = [
{"input_ids": [1, 2, 3], "completion_mask": [0, 1, 1]},
{"input_ids": [4, 5], "completion_mask": [1, 1]},
]
result = collator(examples)
torch.testing.assert_close(result["input_ids"], torch.tensor([[1, 2, 3, 4, 5]]))
torch.testing.assert_close(result["attention_mask"], torch.tensor([[1, 1, 1, 1, 1]]))
torch.testing.assert_close(result["position_ids"], torch.tensor([[0, 1, 2, 0, 1]]))
torch.testing.assert_close(result["labels"], torch.tensor([[-100, 2, 3, 4, 5]]))
def test_packing_drops_attention_mask_for_flash_attention(self):
"""Test that when using packing with position_ids, attention_mask is dropped with fa2."""
collator = DataCollatorForLanguageModeling(pad_token_id=0, padding_free=True, return_position_ids=True)
# Simulate packed sequences with position_ids that restart (typical of BFD packing)
examples = [
{
"input_ids": [1, 2, 3, 4, 5, 6, 7, 8], # Packed: [1,2,3] + [4,5] + [6,7,8]
"seq_lengths": [3, 2, 3],
}
]
result = collator(examples)
# Verify that attention_mask is NOT present - this allows FlashAttention to use position_ids
self.assertNotIn("attention_mask", result, "attention_mask should be dropped for packing with position_ids")
# Verify essential keys are present
self.assertIn("input_ids", result)
self.assertIn("position_ids", result)
self.assertIn("labels", result)
# Verify the data is correctly processed
torch.testing.assert_close(result["input_ids"], torch.tensor([[1, 2, 3, 4, 5, 6, 7, 8]]))
torch.testing.assert_close(result["position_ids"], torch.tensor([[0, 1, 2, 0, 1, 0, 1, 2]]))
torch.testing.assert_close(result["labels"], torch.tensor([[1, 2, 3, 4, 5, 6, 7, 8]]))
def test_padding_free_without_position_ids_keeps_attention_mask(self):
"""
Test that padding_free mode without explicit position_ids still creates attention_mask.
"""
collator = DataCollatorForLanguageModeling(pad_token_id=0, padding_free=True, return_position_ids=True)
# Examples without position_ids (not packed)
examples = [{"input_ids": [1, 2, 3, 4, 5]}]
result = collator(examples)
# Should still have attention_mask since no packed position_ids
self.assertIn("attention_mask", result, "attention_mask should be present when no packed position_ids")
self.assertIn("position_ids", result)
self.assertIn("input_ids", result)
torch.testing.assert_close(result["input_ids"], torch.tensor([[1, 2, 3, 4, 5]]))
torch.testing.assert_close(result["attention_mask"], torch.tensor([[1, 1, 1, 1, 1]]))
torch.testing.assert_close(result["position_ids"], torch.tensor([[0, 1, 2, 3, 4]]))
def test_pad_to_multiple_of(self):
"""Test padding to multiple of specified value."""
collator = DataCollatorForLanguageModeling(pad_token_id=0, pad_to_multiple_of=4)
examples = [{"input_ids": [1, 2, 3]}, {"input_ids": [4, 5]}]
result = collator(examples)
torch.testing.assert_close(result["input_ids"], torch.tensor([[1, 2, 3, 0], [4, 5, 0, 0]]))
torch.testing.assert_close(result["attention_mask"], torch.tensor([[1, 1, 1, 0], [1, 1, 0, 0]]))
torch.testing.assert_close(result["position_ids"], torch.tensor([[0, 1, 2, 0], [0, 1, 0, 0]]))
torch.testing.assert_close(result["labels"], torch.tensor([[1, 2, 3, -100], [4, 5, -100, -100]]))
def test_custom_position_ids(self):
"""Test handling of custom position IDs in examples."""
self.collator = DataCollatorForLanguageModeling(pad_token_id=0)
examples = [{"input_ids": [1, 2, 3], "seq_lengths": [1, 2]}, {"input_ids": [4, 5], "seq_lengths": [2]}]
result = self.collator(examples)
torch.testing.assert_close(result["input_ids"], torch.tensor([[1, 2, 3], [4, 5, 0]]))
torch.testing.assert_close(result["attention_mask"], torch.tensor([[1, 1, 1], [1, 1, 0]]))
torch.testing.assert_close(result["position_ids"], torch.tensor([[0, 0, 1], [0, 1, 0]]))
torch.testing.assert_close(result["labels"], torch.tensor([[1, 2, 3], [4, 5, -100]]))
def test_single_example(self):
"""Test collator with a single example."""
self.collator = DataCollatorForLanguageModeling(pad_token_id=0)
examples = [{"input_ids": [1, 2, 3, 4]}]
result = self.collator(examples)
torch.testing.assert_close(result["input_ids"], torch.tensor([[1, 2, 3, 4]]))
torch.testing.assert_close(result["attention_mask"], torch.tensor([[1, 1, 1, 1]]))
torch.testing.assert_close(result["position_ids"], torch.tensor([[0, 1, 2, 3]]))
torch.testing.assert_close(result["labels"], torch.tensor([[1, 2, 3, 4]]))
def test_different_pad_token_id(self):
"""Test with different pad token ID."""
collator = DataCollatorForLanguageModeling(pad_token_id=999)
examples = [{"input_ids": [1, 2, 3]}, {"input_ids": [4, 5]}]
result = collator(examples)
torch.testing.assert_close(result["input_ids"], torch.tensor([[1, 2, 3], [4, 5, 999]]))
torch.testing.assert_close(result["attention_mask"], torch.tensor([[1, 1, 1], [1, 1, 0]]))
torch.testing.assert_close(result["position_ids"], torch.tensor([[0, 1, 2], [0, 1, 0]]))
torch.testing.assert_close(result["labels"], torch.tensor([[1, 2, 3], [4, 5, -100]]))
def test_assistant_masks(self):
"""Test handling of assistant masks in examples."""
self.collator = DataCollatorForLanguageModeling(pad_token_id=0)
examples = [
{"input_ids": [1, 2, 3], "assistant_masks": [0, 1, 1]},
{"input_ids": [4, 5], "assistant_masks": [0, 1]},
]
result = self.collator(examples)
torch.testing.assert_close(result["input_ids"], torch.tensor([[1, 2, 3], [4, 5, 0]]))
torch.testing.assert_close(result["attention_mask"], torch.tensor([[1, 1, 1], [1, 1, 0]]))
torch.testing.assert_close(result["position_ids"], torch.tensor([[0, 1, 2], [0, 1, 0]]))
torch.testing.assert_close(result["labels"], torch.tensor([[-100, 2, 3], [-100, 5, -100]]))
class SFTTrainerTester(unittest.TestCase):
r""" """
def setUp(self):
self.model_id = "trl-internal-testing/tiny-Qwen2ForCausalLM-2.5"
self.model = AutoModelForCausalLM.from_pretrained(self.model_id)
self.tokenizer = AutoTokenizer.from_pretrained(self.model_id)
self.dummy_dataset = Dataset.from_dict(
{
"question": [
"Does llamas know how to code?",
"Does llamas know how to fly?",
"Does llamas know how to talk?",
"Does llamas know how to code?",
"Does llamas know how to fly?",
"Does llamas know how to talk?",
"Does llamas know how to swim?",
],
"answer": [
"Yes, llamas are very good at coding.",
"No, llamas can't fly.",
"Yes, llamas are very good at talking.",
"Yes, llamas are very good at coding.",
"No, llamas can't fly.",
"Yes, llamas are very good at talking.",
"No, llamas can't swim.",
],
"text": [
"### Question: Does llamas know how to code?\n ### Answer: Yes, llamas are very good at coding.",
"### Question: Does llamas know how to fly?\n ### Answer: No, llamas can't fly.",
"### Question: Does llamas know how to talk?\n ### Answer: Yes, llamas are very good at talking.",
"### Question: Does llamas know how to code?\n ### Answer: Yes, llamas are very good at coding.",
"### Question: Does llamas know how to fly?\n ### Answer: No, llamas can't fly.",
"### Question: Does llamas know how to talk?\n ### Answer: Yes, llamas are very good at talking.",
"### Question: Does llamas know how to swim?\n ### Answer: No, llamas can't swim.",
],
}
)
self.conversational_lm_dataset = load_dataset("trl-internal-testing/zen", "conversational_language_modeling")
self.standard_prompt_completion_dataset = load_dataset(
"trl-internal-testing/zen", "standard_prompt_completion"
)
if is_vision_available():
self.dummy_vsft_instruction_dataset = Dataset.from_dict(
{
"messages": [
[
{
"role": "user",
"content": [{"type": "text", "text": "What is in this image?"}, {"type": "image"}],
},
{
"role": "assistant",
"content": [{"type": "text", "text": "It is random noise."}],
},
{
"role": "user",
"content": [{"type": "text", "text": "Oh ye, you are right, what is 1+1"}],
},
{
"role": "assistant",
"content": [{"type": "text", "text": "2"}],
},
],
[
{
"role": "user",
"content": [{"type": "text", "text": "What is in this image?"}, {"type": "image"}],
},
{
"role": "assistant",
"content": [{"type": "text", "text": "It is random noise."}],
},
],
],
"images": [
[PILImage.fromarray((np.random.rand(40, 50, 3) * 255).astype("uint8")).convert("RGBA")],
[PILImage.fromarray((np.random.rand(50, 60, 3) * 255).astype("uint8")).convert("RGBA")],
],
}
)
self.dummy_vsft_instruction_dataset.cast_column("images", Sequence(Image()))
self.dummy_vsft_instruction_dataset = self.dummy_vsft_instruction_dataset.cast_column(
"images", Sequence(Image())
)
def test_uncorrect_data(self):
with tempfile.TemporaryDirectory() as tmp_dir:
# Shoud work as SFTTrainer natively supports conversational lm dataset
training_args = SFTConfig(
output_dir=tmp_dir,
per_device_train_batch_size=2,
max_length=32, # make sure there is at least 1 packed sequence
packing=True,
report_to="none",
)
_ = SFTTrainer(
model="trl-internal-testing/tiny-Qwen2ForCausalLM-2.5",
args=training_args,
train_dataset=self.conversational_lm_dataset["train"],
)
# Same, but without packing
training_args = SFTConfig(
output_dir=tmp_dir,
per_device_train_batch_size=2,
packing=False,
report_to="none",
)
_ = SFTTrainer(
model="trl-internal-testing/tiny-Qwen2ForCausalLM-2.5",
args=training_args,
train_dataset=self.conversational_lm_dataset["train"],
)
# Same, but with packing with `max_length`
training_args = SFTConfig(
output_dir=tmp_dir,
per_device_train_batch_size=2,
max_length=16, # make sure there is at least 1 packed sequence
packing=True,
report_to="none",
)
_ = SFTTrainer(
model="trl-internal-testing/tiny-Qwen2ForCausalLM-2.5",
args=training_args,
train_dataset=self.standard_prompt_completion_dataset["train"],
)
# Same but with prompt-completion dataset
training_args = SFTConfig(
output_dir=tmp_dir,
per_device_train_batch_size=2,
packing=False,
report_to="none",
)
_ = SFTTrainer(
model="trl-internal-testing/tiny-Qwen2ForCausalLM-2.5",
args=training_args,
train_dataset=self.standard_prompt_completion_dataset["train"],
)
# Should work as dummy dataset are supported with a formatting function
training_args = SFTConfig(
output_dir=tmp_dir,
per_device_train_batch_size=2,
max_length=32, # make sure there is at least 1 packed sequence
packing=True,
report_to="none",
)
_ = SFTTrainer(
model="trl-internal-testing/tiny-Qwen2ForCausalLM-2.5",
args=training_args,
train_dataset=self.dummy_dataset,
formatting_func=formatting_prompts_func,
)
def test_with_model_(self):
with tempfile.TemporaryDirectory() as tmp_dir:
training_args = SFTConfig(
output_dir=tmp_dir,
per_device_train_batch_size=2,
max_length=16,
packing=True,
report_to="none",
)
trainer = SFTTrainer(
model=self.model,
args=training_args,
train_dataset=self.dummy_dataset,
)
trainer.train()
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
# with formatting_func + packed
with tempfile.TemporaryDirectory() as tmp_dir:
training_args = SFTConfig(
output_dir=tmp_dir,
per_device_train_batch_size=2,
max_length=16,
packing=True,
report_to="none",
)
trainer = SFTTrainer(
model=self.model,
args=training_args,
train_dataset=self.dummy_dataset,
formatting_func=formatting_prompts_func,
)
trainer.train()
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
with tempfile.TemporaryDirectory() as tmp_dir:
training_args = SFTConfig(
output_dir=tmp_dir,
per_device_train_batch_size=2,
max_length=16,
report_to="none",
)
trainer = SFTTrainer(
model=self.model,
args=training_args,
train_dataset=self.dummy_dataset,
)
trainer.train()
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
def test_only_train_packing(self):
with tempfile.TemporaryDirectory() as tmp_dir:
training_args = SFTConfig(
output_dir=tmp_dir,
per_device_train_batch_size=2,
gradient_checkpointing=True,
packing=True,
max_length=128, # make sure there is at least 1 packed sequence
eval_packing=False,
report_to="none",
)
trainer = SFTTrainer(
model="trl-internal-testing/tiny-Qwen2ForCausalLM-2.5",
args=training_args,
train_dataset=self.conversational_lm_dataset["train"],
eval_dataset=self.conversational_lm_dataset["test"],
)
self.assertEqual(len(trainer.train_dataset["input_ids"]), 7) # w/ this dataset, we end up with 46 seqs
self.assertEqual(len(trainer.eval_dataset["input_ids"]), len(self.conversational_lm_dataset["test"]))
def test_eval_packing(self):
with tempfile.TemporaryDirectory() as tmp_dir:
training_args = SFTConfig(
output_dir=tmp_dir,
per_device_train_batch_size=2,
max_length=128, # make sure there is at least 1 packed sequence
packing=True,
report_to="none",
)
trainer = SFTTrainer(
model="trl-internal-testing/tiny-Qwen2ForCausalLM-2.5",
args=training_args,
train_dataset=self.conversational_lm_dataset["train"],
eval_dataset=self.conversational_lm_dataset["test"],
)
self.assertEqual(len(trainer.train_dataset["input_ids"]), 7) # w/ this dataset, we end up with 46 seqs
self.assertEqual(len(trainer.eval_dataset["input_ids"]), 1) # w/ this dataset, we end up with 6 seqs
def test_no_packing(self):
with tempfile.TemporaryDirectory() as tmp_dir:
training_args = SFTConfig(
output_dir=tmp_dir,
per_device_train_batch_size=2,
max_length=128, # make sure there is at least 1 packed sequence
packing=False,
report_to="none",
)
trainer = SFTTrainer(
model="trl-internal-testing/tiny-Qwen2ForCausalLM-2.5",
args=training_args,
train_dataset=self.conversational_lm_dataset["train"],
eval_dataset=self.conversational_lm_dataset["test"],
)
self.assertEqual(len(trainer.train_dataset["input_ids"]), len(self.conversational_lm_dataset["train"]))
self.assertEqual(len(trainer.eval_dataset["input_ids"]), len(self.conversational_lm_dataset["test"]))
@require_vision
def test_skip_prepare_dataset(self):
with tempfile.TemporaryDirectory() as tmp_dir:
training_args = SFTConfig(
output_dir=tmp_dir,
per_device_train_batch_size=2,
remove_unused_columns=False,
dataset_kwargs={"skip_prepare_dataset": True},
report_to="none",
)
trainer = SFTTrainer(
model="trl-internal-testing/tiny-Qwen2ForCausalLM-2.5",
args=training_args,
train_dataset=self.dummy_vsft_instruction_dataset,
)
self.assertEqual(trainer.train_dataset.features, self.dummy_vsft_instruction_dataset.features)
def test_skip_prepare_dataset_with_no_packing(self):
with tempfile.TemporaryDirectory() as tmp_dir:
training_args = SFTConfig(
output_dir=tmp_dir,
per_device_train_batch_size=2,
remove_unused_columns=False,
packing=False,
dataset_kwargs={"skip_prepare_dataset": True},
report_to="none",
)
trainer = SFTTrainer(
model="trl-internal-testing/tiny-Qwen2ForCausalLM-2.5",
args=training_args,
train_dataset=self.dummy_dataset,
)
self.assertEqual(trainer.train_dataset.features, self.dummy_dataset.features)
@require_vision
def test_llava(self):
with tempfile.TemporaryDirectory() as tmp_dir:
training_args = SFTConfig(
output_dir=tmp_dir,
remove_unused_columns=False,
dataset_kwargs={"skip_prepare_dataset": True},
report_to="none",
)
tiny_llava = LlavaForConditionalGeneration.from_pretrained(
"trl-internal-testing/tiny-LlavaForConditionalGeneration"
)
processor = AutoProcessor.from_pretrained("trl-internal-testing/tiny-LlavaForConditionalGeneration")
processor.chat_template = """{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}A chat between a curious
user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's
questions. {% for message in messages %}{% if message['role'] == 'user' %}USER: {% else %}ASSISTANT: {% endif %}{% for
item in message['content'] %}{% if item['type'] == 'text' %}{{ item['text'] }}{% elif item['type'] == 'image'
%}<image>{% endif %}{% endfor %}{% if message['role'] == 'user' %} {% else %}{{eos_token}}{% endif %}{% endfor %}{% if
add_generation_prompt %}ASSISTANT: {% endif %}"""
def collate_fn(examples):
# Get the texts and images, and apply the chat template
texts = [processor.apply_chat_template(example["messages"], tokenize=False) for example in examples]
images = [example["images"][0] for example in examples]
# Tokenize the texts and process the images
batch = processor(images=images, text=texts, return_tensors="pt", padding=True)
# The labels are the input_ids, and we mask the padding tokens in the loss computation
labels = batch["input_ids"].clone()
labels[labels == processor.tokenizer.pad_token_id] = -100
batch["labels"] = labels
return batch
trainer = SFTTrainer(
model=tiny_llava,
args=training_args,
data_collator=collate_fn,
train_dataset=self.dummy_vsft_instruction_dataset,
)
trainer.train()
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
# This new tester aims to replace the first one at some point
class SFTTrainerTester2(unittest.TestCase):
@parameterized.expand(
[
("trl-internal-testing/tiny-Qwen2ForCausalLM-2.5",),
("trl-internal-testing/tiny-Qwen3MoeForCausalLM",),
("trl-internal-testing/tiny-GptOssForCausalLM",),
]
)
def test_train(self, model_id):
# Get the dataset
dataset = load_dataset("trl-internal-testing/zen", "standard_language_modeling", split="train")
with tempfile.TemporaryDirectory() as tmp_dir:
# Initialize the trainer
training_args = SFTConfig(output_dir=tmp_dir, report_to="none")
trainer = SFTTrainer(model=model_id, args=training_args, train_dataset=dataset)
# Save the initial parameters to compare them later
previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
# Train the model
trainer.train()
# Check that the training loss is not None
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
# Check the params have changed
for n, param in previous_trainable_params.items():
new_param = trainer.model.get_parameter(n)
self.assertFalse(torch.allclose(param, new_param), f"Parameter {n} has not changed")
# Special case for harmony
def test_train_gpt_oss(self):
# Get the dataset
dataset = load_dataset("trl-internal-testing/harmony", "language_modeling", split="train")
with tempfile.TemporaryDirectory() as tmp_dir:
# Initialize the trainer
training_args = SFTConfig(output_dir=tmp_dir, report_to="none")
trainer = SFTTrainer(
model="trl-internal-testing/tiny-GptOssForCausalLM", args=training_args, train_dataset=dataset
)
# Save the initial parameters to compare them later
previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
# Train the model
trainer.train()
# Check that the training loss is not None
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
# Check the params have changed
for n, param in previous_trainable_params.items():
new_param = trainer.model.get_parameter(n)
self.assertFalse(torch.allclose(param, new_param), f"Parameter {n} has not changed")
def test_train_model(self):
# Instantiate the model
model = AutoModelForCausalLM.from_pretrained("trl-internal-testing/tiny-Qwen2ForCausalLM-2.5")
# Get the dataset
dataset = load_dataset("trl-internal-testing/zen", "standard_language_modeling", split="train")
with tempfile.TemporaryDirectory() as tmp_dir:
# Initialize the trainer
training_args = SFTConfig(output_dir=tmp_dir, report_to="none")
trainer = SFTTrainer(model=model, args=training_args, train_dataset=dataset)
# Save the initial parameters to compare them later
previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
# Train the model
trainer.train()
# Check that the training loss is not None
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
# Check the params have changed
for n, param in previous_trainable_params.items():
new_param = trainer.model.get_parameter(n)
self.assertFalse(torch.allclose(param, new_param), f"Parameter {n} has not changed")
def test_train_model_torch_dtype(self):
# Get the dataset
dataset = load_dataset("trl-internal-testing/zen", "standard_language_modeling", split="train")
with tempfile.TemporaryDirectory() as tmp_dir:
# Initialize the trainer
training_args = SFTConfig(
output_dir=tmp_dir, model_init_kwargs={"torch_dtype": torch.float16}, report_to="none"
)
trainer = SFTTrainer(
model="trl-internal-testing/tiny-Qwen2ForCausalLM-2.5", args=training_args, train_dataset=dataset
)
# Save the initial parameters to compare them later
previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
# Train the model
trainer.train()
# Check that the training loss is not None
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
# Check the params have changed
for n, param in previous_trainable_params.items():
new_param = trainer.model.get_parameter(n)
# Check the torch dtype
self.assertEqual(new_param.dtype, torch.float16)
self.assertFalse(torch.allclose(param, new_param), f"Parameter {n} has not changed")
@require_peft
def test_train_dense_with_peft_config(self):
# Get the base model parameter names
model_id = "trl-internal-testing/tiny-Qwen2ForCausalLM-2.5"
model = AutoModelForCausalLM.from_pretrained(model_id)
base_param_names = [f"base_model.model.{n}" for n, _ in model.named_parameters()]
# Get the dataset
dataset = load_dataset("trl-internal-testing/zen", "standard_language_modeling", split="train")
with tempfile.TemporaryDirectory() as tmp_dir:
# Initialize the trainer
training_args = SFTConfig(output_dir=tmp_dir, report_to="none")
trainer = SFTTrainer(
model=model_id,
args=training_args,
train_dataset=dataset,
peft_config=LoraConfig(),
)
# Save the initial parameters to compare them later
previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
# Train the model
trainer.train()
# Check that the training loss is not None
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
# Check the peft params have changed and the base model params have not changed
for n, param in previous_trainable_params.items():
new_param = trainer.model.get_parameter(n)
if n in base_param_names: # We expect the base model parameters to be the same
self.assertTrue(torch.allclose(param, new_param), f"Parameter {n} has changed")
elif (
"base_layer" not in n
): # We expect the peft parameters to be different (except for the base layer)
self.assertFalse(torch.allclose(param, new_param), f"Parameter {n} has not changed")
@require_peft
def test_train_moe_with_peft_config(self):
# Get the base model parameter names
model_id = "trl-internal-testing/tiny-GptOssForCausalLM"
model = AutoModelForCausalLM.from_pretrained(model_id)
base_param_names = [f"base_model.model.{n}" for n, _ in model.named_parameters()]
# Get the dataset
dataset = load_dataset("trl-internal-testing/zen", "standard_language_modeling", split="train")
with tempfile.TemporaryDirectory() as tmp_dir:
# Initialize the trainer
training_args = SFTConfig(output_dir=tmp_dir, report_to="none")
trainer = SFTTrainer(
model=model_id,
args=training_args,
train_dataset=dataset,
peft_config=LoraConfig(target_parameters=["mlp.experts.down_proj", "mlp.experts.gate_up_proj"]),
)
# Save the initial parameters to compare them later
previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
# Train the model
trainer.train()
# Check that the training loss is not None
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
# Check the peft params have changed and the base model params have not changed
for n, param in previous_trainable_params.items():
new_param = trainer.model.get_parameter(n)
if n in base_param_names: # We expect the base model parameters to be the same
self.assertTrue(torch.allclose(param, new_param), f"Parameter {n} has changed")
elif (
"base_layer" not in n
): # We expect the peft parameters to be different (except for the base layer)
self.assertFalse(torch.allclose(param, new_param), f"Parameter {n} has not changed")
@require_peft
def test_train_peft_model(self):
# Get the base model
model_id = "trl-internal-testing/tiny-Qwen2ForCausalLM-2.5"
model = AutoModelForCausalLM.from_pretrained(model_id)
# Get the base model parameter names
base_param_names = [f"base_model.model.{n}" for n, _ in model.named_parameters()]
# Turn the model into a peft model
lora_config = LoraConfig()
model = get_peft_model(model, lora_config)
# Get the dataset
dataset = load_dataset("trl-internal-testing/zen", "standard_language_modeling", split="train")
with tempfile.TemporaryDirectory() as tmp_dir:
# Initialize the trainer
training_args = SFTConfig(output_dir=tmp_dir, report_to="none")
trainer = SFTTrainer(model=model, args=training_args, train_dataset=dataset)
# Save the initial parameters to compare them later
previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
# Train the model
trainer.train()
# Check that the training loss is not None
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
# Check the peft params have changed and the base model params have not changed
for n, param in previous_trainable_params.items():
new_param = trainer.model.get_parameter(n)
if n in base_param_names: # We expect the base model parameters to be the same
self.assertTrue(torch.allclose(param, new_param), f"Parameter {n} has changed")
elif (
"base_layer" not in n
): # We expect the peft parameters to be different (except for the base layer)
self.assertFalse(torch.allclose(param, new_param), f"Parameter {n} has not changed")
@require_peft
def test_train_dense_with_peft_config_and_gradient_checkpointing(self):
# Get the base model parameter names
model_id = "trl-internal-testing/tiny-Qwen2ForCausalLM-2.5"
model = AutoModelForCausalLM.from_pretrained(model_id)
base_param_names = [f"base_model.model.{n}" for n, _ in model.named_parameters()]
# Get the dataset
dataset = load_dataset("trl-internal-testing/zen", "standard_language_modeling", split="train")
with tempfile.TemporaryDirectory() as tmp_dir:
# Initialize the trainer
training_args = SFTConfig(output_dir=tmp_dir, gradient_checkpointing=True, report_to="none")
trainer = SFTTrainer(
model=model_id,
args=training_args,
train_dataset=dataset,
peft_config=LoraConfig(),
)
# Save the initial parameters to compare them later
previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
# Train the model
trainer.train()
# Check that the training loss is not None
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
# Check the peft params have changed and the base model params have not changed
for n, param in previous_trainable_params.items():
new_param = trainer.model.get_parameter(n)
if n in base_param_names: # We expect the base model parameters to be the same
self.assertTrue(torch.allclose(param, new_param), f"Parameter {n} has changed")
elif (
"base_layer" not in n
): # We expect the peft parameters to be different (except for the base layer)
self.assertFalse(torch.allclose(param, new_param), f"Parameter {n} has not changed")
@require_peft
def test_train_moe_with_peft_config_and_gradient_checkpointing(self):
# Get the base model parameter names
model_id = "trl-internal-testing/tiny-GptOssForCausalLM"
model = AutoModelForCausalLM.from_pretrained(model_id)
base_param_names = [f"base_model.model.{n}" for n, _ in model.named_parameters()]
# Get the dataset
dataset = load_dataset("trl-internal-testing/zen", "standard_language_modeling", split="train")
with tempfile.TemporaryDirectory() as tmp_dir:
# Initialize the trainer
training_args = SFTConfig(output_dir=tmp_dir, gradient_checkpointing=True, report_to="none")
trainer = SFTTrainer(
model=model_id,
args=training_args,
train_dataset=dataset,
peft_config=LoraConfig(target_parameters=["mlp.experts.down_proj", "mlp.experts.gate_up_proj"]),
)
# Save the initial parameters to compare them later
previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
# Train the model
trainer.train()
# Check that the training loss is not None
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
# Check the peft params have changed and the base model params have not changed
for n, param in previous_trainable_params.items():
new_param = trainer.model.get_parameter(n)
if n in base_param_names: # We expect the base model parameters to be the same
self.assertTrue(torch.allclose(param, new_param), f"Parameter {n} has changed")
elif (
"base_layer" not in n
): # We expect the peft parameters to be different (except for the base layer)
self.assertFalse(torch.allclose(param, new_param), f"Parameter {n} has not changed")
@require_peft
def test_train_with_peft_model_and_gradient_checkpointing(self):
# Get the base model parameter names
model_id = "trl-internal-testing/tiny-Qwen2ForCausalLM-2.5"
model = AutoModelForCausalLM.from_pretrained(model_id)
base_param_names = [f"base_model.model.{n}" for n, _ in model.named_parameters()]
model = get_peft_model(model, LoraConfig())
# Get the dataset
dataset = load_dataset("trl-internal-testing/zen", "standard_language_modeling", split="train")
with tempfile.TemporaryDirectory() as tmp_dir:
# Initialize the trainer
training_args = SFTConfig(output_dir=tmp_dir, gradient_checkpointing=True, report_to="none")
trainer = SFTTrainer(model=model, args=training_args, train_dataset=dataset)
# Verify model is a PeftModel
self.assertIsInstance(trainer.model, PeftModel)
# Save the initial parameters to compare them later
previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
# Train the model
trainer.train()
# Check that the training loss is not None
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
# Check the peft params have changed and the base model params have not changed
for n, param in previous_trainable_params.items():
new_param = trainer.model.get_parameter(n)
if n in base_param_names: # We expect the base model parameters to be the same
self.assertTrue(torch.allclose(param, new_param), f"Parameter {n} has changed")
elif (
"base_layer" not in n
): # We expect the peft parameters to be different (except for the base layer)
self.assertFalse(torch.allclose(param, new_param), f"Parameter {n} has not changed")
def test_train_with_non_chatml_conversational_data(self):
# Get the dataset
dataset = load_dataset("trl-internal-testing/zen", "conversational_language_modeling", split="train")
# Rename role/content to from/value to ensure SFT works with non-chatML conversational data
def rename_fields(example: list[dict]):
return {"conversations": [{"from": m["role"], "value": m["content"]} for m in example["messages"]]}
dataset = dataset.map(rename_fields, remove_columns="messages")
with tempfile.TemporaryDirectory() as tmp_dir:
# Initialize the trainer
training_args = SFTConfig(output_dir=tmp_dir, report_to="none")
trainer = SFTTrainer(
model="trl-internal-testing/tiny-Qwen2ForCausalLM-2.5", args=training_args, train_dataset=dataset
)
# Save the initial parameters to compare them later
previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
# Train the model
trainer.train()
# Check that the training loss is not None
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
# Check the params have changed
for n, param in previous_trainable_params.items():
new_param = trainer.model.get_parameter(n)
self.assertFalse(torch.allclose(param, new_param), f"Parameter {n} has not changed")
def test_train_with_pretokenized_data(self):
# Get the dataset
model_id = "trl-internal-testing/tiny-Qwen2ForCausalLM-2.5"
tokenizer = AutoTokenizer.from_pretrained(model_id)
dataset = load_dataset("trl-internal-testing/zen", "standard_language_modeling", split="train")
def tokenize_example(example):
return tokenizer(example["text"])
# Apply tokenization
tokenized_dataset = dataset.map(tokenize_example, remove_columns=["text"])
with tempfile.TemporaryDirectory() as tmp_dir:
# Initialize the trainer
training_args = SFTConfig(output_dir=tmp_dir, report_to="none")
trainer = SFTTrainer(model=model_id, args=training_args, train_dataset=tokenized_dataset)
# Save the initial parameters to compare them later
previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
# Train the model
trainer.train()
# Check that the training loss is not None
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
# Check the params have changed
for n, param in previous_trainable_params.items():
new_param = trainer.model.get_parameter(n)
self.assertFalse(torch.allclose(param, new_param), f"Parameter {n} has not changed")
def test_train_with_iterable_dataset(self):
# Get the dataset
dataset = load_dataset("trl-internal-testing/zen", "standard_language_modeling", split="train", streaming=True)
with tempfile.TemporaryDirectory() as tmp_dir:
# Initialize the trainer
training_args = SFTConfig(output_dir=tmp_dir, max_steps=3, report_to="none")
trainer = SFTTrainer(
model="trl-internal-testing/tiny-Qwen2ForCausalLM-2.5", args=training_args, train_dataset=dataset
)
# Save the initial parameters to compare them later
previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
# Train the model
trainer.train()
# Check that the training loss is not None
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
# Check the params have changed
for n, param in previous_trainable_params.items():
new_param = trainer.model.get_parameter(n)
self.assertFalse(torch.allclose(param, new_param), f"Parameter {n} has not changed")
@require_flash_attn
def test_train_padding_free(self):
# Get the dataset
dataset = load_dataset("trl-internal-testing/zen", "standard_language_modeling", split="train")
with tempfile.TemporaryDirectory() as tmp_dir:
# Initialize the trainer
training_args = SFTConfig(
output_dir=tmp_dir,
padding_free=True,
model_init_kwargs={"attn_implementation": "flash_attention_2"},
bf16=True, # flash_attention_2 only supports bf16 and fp16
report_to="none",
)
trainer = SFTTrainer(
model="trl-internal-testing/tiny-Qwen2ForCausalLM-2.5", args=training_args, train_dataset=dataset
)
# Save the initial parameters to compare them later
previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
# Train the model
trainer.train()
# Check that the training loss is not None
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
# Check the params have changed
for n, param in previous_trainable_params.items():
new_param = trainer.model.get_parameter(n)
self.assertFalse(torch.allclose(param, new_param), f"Parameter {n} has not changed")
@parameterized.expand([("bfd",), ("wrapped",)])
def test_train_packing(self, packing_strategy):
# Get the dataset
dataset = load_dataset("trl-internal-testing/zen", "standard_language_modeling", split="train")
with tempfile.TemporaryDirectory() as tmp_dir:
# Initialize the trainer
training_args = SFTConfig(
output_dir=tmp_dir, packing=True, packing_strategy=packing_strategy, max_length=10, report_to="none"
)
trainer = SFTTrainer(
model="trl-internal-testing/tiny-Qwen2ForCausalLM-2.5", args=training_args, train_dataset=dataset
)
# Save the initial parameters to compare them later
previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
# Train the model
trainer.train()
# Check that the training loss is not None
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
# Check the params have changed
for n, param in previous_trainable_params.items():
new_param = trainer.model.get_parameter(n)
self.assertFalse(torch.allclose(param, new_param), f"Parameter {n} has not changed")
def test_train_with_chat_template_kwargs(self):
# Get the dataset
dataset = load_dataset("trl-internal-testing/zen", "standard_language_modeling", split="train")
with tempfile.TemporaryDirectory() as tmp_dir:
# Initialize the trainer
training_args = SFTConfig(output_dir=tmp_dir, report_to="none")
tokenizer = AutoTokenizer.from_pretrained("trl-internal-testing/tiny-Qwen2ForCausalLM-2.5")
# The following template is a simplified version of the Qwen chat template, where an additional argument
# `role_capital` is used to control the capitalization of roles.
tokenizer.chat_template = '{%- if messages[0]["role"] == "system" -%} {{ "<|im_start|>" + ("SYSTEM" if role_capital else "system") + "\\n" + messages[0]["content"] + "<|im_end|>\\n" }}{%- else -%} {{ "<|im_start|>" + ("SYSTEM" if role_capital else "system") + "\\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>\\n" }}{%- endif -%}{%- for message in messages -%} {%- if (message.role == "user") or (message.role == "system" and not loop.first) or (message.role == "assistant" and not message.tool_calls) -%} {{ "<|im_start|>" + (message.role.upper() if role_capital else message.role) + "\\n" + message.content + "<|im_end|>\\n" }} {%- elif message.role == "assistant" -%} {{ "<|im_start|>" + ("ASSISTANT" if role_capital else "assistant") }} {%- if message.content -%} {{ "\\n" + message.content }} {%- endif -%} {{ "<|im_end|>\\n" }} {%- elif message.role == "tool" -%} {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != "tool") -%} {{ "<|im_start|>" + ("USER" if role_capital else "user") }} {%- endif -%} {{ "\\n<tool_response>\\n" + message.content + "\\n</tool_response>" }} {%- if loop.last or (messages[loop.index0 + 1].role != "tool") -%} {{ "<|im_end|>\\n" }} {%- endif -%} {%- endif -%}{%- endfor -%}{%- if add_generation_prompt -%} {{ "<|im_start|>" + ("ASSISTANT" if role_capital else "assistant") + "\\n" }}{%- endif -%}'
dataset.add_column("chat_template_kwargs", [{"role_capital": bool(i % 2)} for i in range(len(dataset))])
trainer = SFTTrainer(
model="trl-internal-testing/tiny-Qwen2ForCausalLM-2.5", args=training_args, train_dataset=dataset
)
# Save the initial parameters to compare them later
previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
# Train the model
trainer.train()
# Check that the training loss is not None
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
# Check the params have changed
for n, param in previous_trainable_params.items():
new_param = trainer.model.get_parameter(n)
self.assertFalse(torch.allclose(param, new_param), f"Parameter {n} has not changed")
def test_train_assistant_only(self):
# Get the dataset
dataset = load_dataset("trl-internal-testing/zen", "conversational_language_modeling", split="train")
with tempfile.TemporaryDirectory() as tmp_dir:
# Initialize the trainer
training_args = SFTConfig(output_dir=tmp_dir, assistant_only_loss=True, report_to="none")
trainer = SFTTrainer(
model="trl-internal-testing/tiny-Qwen3ForCausalLM", args=training_args, train_dataset=dataset
)
# Save the initial parameters to compare them later
previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
# Train the model
trainer.train()
# Check that the training loss is not None
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
# Check the params have changed
for n, param in previous_trainable_params.items():
new_param = trainer.model.get_parameter(n)
self.assertFalse(torch.allclose(param, new_param), f"Parameter {n} has not changed")
def test_train_completion_only(self):
# Get the dataset
dataset = load_dataset("trl-internal-testing/zen", "conversational_prompt_completion", split="train")
with tempfile.TemporaryDirectory() as tmp_dir:
# Initialize the trainer
training_args = SFTConfig(output_dir=tmp_dir, completion_only_loss=True, report_to="none")
trainer = SFTTrainer(
model="trl-internal-testing/tiny-Qwen3ForCausalLM", args=training_args, train_dataset=dataset
)
# Save the initial parameters to compare them later
previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
# Train the model
trainer.train()
# Check that the training loss is not None
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
# Check the params have changed
for n, param in previous_trainable_params.items():
new_param = trainer.model.get_parameter(n)
self.assertFalse(torch.allclose(param, new_param), f"Parameter {n} has not changed")
def test_train_completion_only_harmony(self):
# Get the dataset
dataset = load_dataset("trl-internal-testing/harmony", "prompt_completion", split="train")
with tempfile.TemporaryDirectory() as tmp_dir:
# Initialize the trainer
training_args = SFTConfig(output_dir=tmp_dir, completion_only_loss=True, report_to="none")
trainer = SFTTrainer(
model="trl-internal-testing/tiny-GptOssForCausalLM", args=training_args, train_dataset=dataset
)
# Save the initial parameters to compare them later
previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
# Train the model
trainer.train()
# Check that the training loss is not None
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
# Check the params have changed
for n, param in previous_trainable_params.items():
new_param = trainer.model.get_parameter(n)
self.assertFalse(torch.allclose(param, new_param), f"Parameter {n} has not changed")
def test_train_assistant_only_and_completion_only(self):
# Get the dataset
dataset = load_dataset("trl-internal-testing/zen", "conversational_prompt_completion", split="train")
# To test this case, we need to add user messages in the completion (they'll be masked in the loss)
def add_to_completion(example):
example["completion"].append(example["prompt"][0])
example["completion"].append(example["completion"][0])
return example
dataset = dataset.map(add_to_completion)
with tempfile.TemporaryDirectory() as tmp_dir:
# Initialize the trainer
training_args = SFTConfig(
output_dir=tmp_dir, assistant_only_loss=True, completion_only_loss=True, report_to="none"
)
trainer = SFTTrainer(
model="trl-internal-testing/tiny-Qwen3ForCausalLM", args=training_args, train_dataset=dataset
)
# Save the initial parameters to compare them later
previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
# Train the model
trainer.train()
# Check that the training loss is not None
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
# Check the params have changed
for n, param in previous_trainable_params.items():
new_param = trainer.model.get_parameter(n)
self.assertFalse(torch.allclose(param, new_param), f"Parameter {n} has not changed")
def test_train_assistant_only_iterable_dataset(self):
# Get the dataset
dataset = load_dataset(
"trl-internal-testing/zen", "conversational_language_modeling", split="train", streaming=True
)
with tempfile.TemporaryDirectory() as tmp_dir:
# Initialize the trainer
training_args = SFTConfig(output_dir=tmp_dir, assistant_only_loss=True, max_steps=3, report_to="none")
trainer = SFTTrainer(
model="trl-internal-testing/tiny-Qwen3ForCausalLM", args=training_args, train_dataset=dataset
)
# Save the initial parameters to compare them later
previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
# Train the model
trainer.train()
# Check that the training loss is not None
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
# Check the params have changed
for n, param in previous_trainable_params.items():
new_param = trainer.model.get_parameter(n)
self.assertFalse(torch.allclose(param, new_param), f"Parameter {n} has not changed")
def test_train_with_set_chat_template_from_model(self):
# Get the dataset
dataset = load_dataset("trl-internal-testing/zen", "conversational_language_modeling", split="train")
with tempfile.TemporaryDirectory() as tmp_dir:
# Initialize the trainer
training_args = SFTConfig(output_dir=tmp_dir, chat_template_path="Qwen/Qwen3-4B", report_to="none")
# trl-internal-testing/tiny-GPTNeoXForCausalLM doesn't have a chat template set by default
trainer = SFTTrainer(
model="trl-internal-testing/tiny-GPTNeoXForCausalLM", args=training_args, train_dataset=dataset
)
# Save the initial parameters to compare them later
previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
# Train the model
trainer.train()
# Check that the training loss is not None
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
# Check the params have changed
for n, param in previous_trainable_params.items():
new_param = trainer.model.get_parameter(n)
self.assertFalse(torch.allclose(param, new_param), f"Parameter {n} has not changed")
def test_train_with_set_chat_template_from_path(self):
# Get the dataset
dataset = load_dataset("trl-internal-testing/zen", "conversational_language_modeling", split="train")
with tempfile.TemporaryDirectory() as tmp_dir:
# Initialize the trainer
training_args = SFTConfig(
output_dir=tmp_dir,
chat_template_path=str(pathlib.Path(__file__).parent / "data" / "template.jinja"),
report_to="none",
)
# trl-internal-testing/tiny-GPTNeoXForCausalLM doesn't have a chat template set by default
trainer = SFTTrainer(
model="trl-internal-testing/tiny-GPTNeoXForCausalLM", args=training_args, train_dataset=dataset
)
# Save the initial parameters to compare them later
previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
# Train the model
trainer.train()
# Check that the training loss is not None
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
# Check the params have changed
for n, param in previous_trainable_params.items():
new_param = trainer.model.get_parameter(n)
self.assertFalse(torch.allclose(param, new_param), f"Parameter {n} has not changed")
# Check that the template saved in the output directory is the same as the one used for training
template_path = pathlib.Path(tmp_dir) / "checkpoint-9" / "chat_template.jinja"
self.assertTrue(template_path.exists(), f"Chat template not found at {template_path}")
with open(template_path) as f:
template_content = f.read()
with open(training_args.chat_template_path) as f:
original_template_content = f.read()
self.assertEqual(
template_content, original_template_content, "Chat template content does not match the original"
)
def test_train_toolcall_data(self):
# Get the dataset
dataset = load_dataset("trl-internal-testing/toolcall", split="train")
with tempfile.TemporaryDirectory() as tmp_dir:
# Initialize the trainer
training_args = SFTConfig(output_dir=tmp_dir, report_to="none")
trainer = SFTTrainer(
model="trl-internal-testing/tiny-Qwen2ForCausalLM-2.5", args=training_args, train_dataset=dataset
)
# Save the initial parameters to compare them later
previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
# Train the model
trainer.train()
# Check that the training loss is not None
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
# Check the params have changed
for n, param in previous_trainable_params.items():
new_param = trainer.model.get_parameter(n)
self.assertFalse(torch.allclose(param, new_param), f"Parameter {n} has not changed")
def test_train_with_eval(self):
# Get the dataset
dataset = load_dataset("trl-internal-testing/zen", "standard_language_modeling")
with tempfile.TemporaryDirectory() as tmp_dir:
# Initialize the trainer
training_args = SFTConfig(output_dir=tmp_dir, eval_strategy="steps", eval_steps=3, report_to="none")
trainer = SFTTrainer(
model="trl-internal-testing/tiny-Qwen2ForCausalLM-2.5",
args=training_args,
train_dataset=dataset["train"],
eval_dataset=dataset["test"],
)
# Train the model
trainer.train()
# Check that the eval loss is not None
self.assertIsNotNone(trainer.state.log_history[0]["eval_loss"])
def test_train_with_multiple_eval_dataset(self):
# Get the dataset
dataset = load_dataset("trl-internal-testing/zen", "standard_language_modeling")
with tempfile.TemporaryDirectory() as tmp_dir:
# Initialize the trainer
training_args = SFTConfig(output_dir=tmp_dir, eval_strategy="steps", eval_steps=3, report_to="none")
trainer = SFTTrainer(
model="trl-internal-testing/tiny-Qwen2ForCausalLM-2.5",
args=training_args,
train_dataset=dataset["train"],
eval_dataset={"data1": dataset["test"], "data2": dataset["test"]},
)
# Train the model
trainer.train()
# Check that the eval losses are not None
self.assertIsNotNone(trainer.state.log_history[-3]["eval_data1_loss"])
self.assertIsNotNone(trainer.state.log_history[-2]["eval_data2_loss"])
def test_train_with_gradient_checkpointing(self):
# Get the dataset
dataset = load_dataset("trl-internal-testing/zen", "standard_language_modeling", split="train")
with tempfile.TemporaryDirectory() as tmp_dir:
# Initialize the trainer
training_args = SFTConfig(output_dir=tmp_dir, gradient_checkpointing=True, report_to="none")
trainer = SFTTrainer(
model="trl-internal-testing/tiny-Qwen2ForCausalLM-2.5", args=training_args, train_dataset=dataset
)
# Save the initial parameters to compare them later
previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
# Train the model
trainer.train()
# Check that the training loss is not None
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
# Check the params have changed
for n, param in previous_trainable_params.items():
new_param = trainer.model.get_parameter(n)
self.assertFalse(torch.allclose(param, new_param), f"Parameter {n} has not changed")
def test_tag_added(self):
# Get the dataset
dataset = load_dataset("trl-internal-testing/zen", "standard_language_modeling", split="train")
# Initialize the trainer
trainer = SFTTrainer(
model="trl-internal-testing/tiny-Qwen2ForCausalLM-2.5",
train_dataset=dataset,
)
for tag in ["sft", "trl"]:
self.assertIn(tag, trainer.model.model_tags)
@require_peft
def test_tag_added_peft(self):
# Get the dataset
dataset = load_dataset("trl-internal-testing/zen", "standard_language_modeling", split="train")
# Initialize the trainer
trainer = SFTTrainer(
model="trl-internal-testing/tiny-Qwen2ForCausalLM-2.5",
train_dataset=dataset,
peft_config=LoraConfig(),
)
for tag in ["sft", "trl"]:
self.assertIn(tag, trainer.model.model_tags)
def test_train_with_torch_dtype(self):
# Get the dataset
dataset = load_dataset("trl-internal-testing/zen", "standard_language_modeling", split="train")
with tempfile.TemporaryDirectory() as tmp_dir:
# Initialize the trainer
training_args = SFTConfig(
output_dir=tmp_dir, model_init_kwargs={"torch_dtype": torch.float16}, report_to="none"
)
trainer = SFTTrainer(
model="trl-internal-testing/tiny-Qwen2ForCausalLM-2.5", args=training_args, train_dataset=dataset
)
# Save the initial parameters to compare them later
previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
# Train the model
trainer.train()
# Check that the training loss is not None
self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
# Check the params have changed
for n, param in previous_trainable_params.items():
new_param = trainer.model.get_parameter(n)
self.assertFalse(torch.allclose(param, new_param), f"Parameter {n} has not changed")
|