File size: 16,239 Bytes
2f5127c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# 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 json
import os
import tempfile
import unittest

from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig, Trainer, TrainingArguments
from transformers.testing_utils import require_peft, require_wandb
from transformers.trainer_utils import get_last_checkpoint
from transformers.utils import is_peft_available

from tests.testing_utils import require_comet, require_mergekit
from trl import BasePairwiseJudge, DPOConfig, DPOTrainer, LogCompletionsCallback, MergeModelCallback, WinRateCallback
from trl.mergekit_utils import MergeConfig


if is_peft_available():
    from peft import LoraConfig


class HalfPairwiseJudge(BasePairwiseJudge):
    """Naive pairwise judge that always returns [1, 0] for two prompts"""

    def judge(self, prompts, completions, shuffle_order=True, return_scores=False):
        # just check that the batch size is 2
        assert len(prompts) == 2
        if return_scores:
            return [0.3, 0.9]
        return [1, 0]


class TrainerWithRefModel(Trainer):
    # This is a dummy class to test the callback. Compared to the Trainer class, it only has an additional
    # ref_model attribute
    def __init__(self, model, ref_model, args, train_dataset, eval_dataset, processing_class):
        super().__init__(
            model=model,
            args=args,
            train_dataset=train_dataset,
            eval_dataset=eval_dataset,
            processing_class=processing_class,
        )
        self.ref_model = ref_model


class WinRateCallbackTester(unittest.TestCase):
    def setUp(self):
        self.model = AutoModelForCausalLM.from_pretrained("trl-internal-testing/tiny-Qwen2ForCausalLM-2.5")
        self.ref_model = AutoModelForCausalLM.from_pretrained("trl-internal-testing/tiny-Qwen2ForCausalLM-2.5")
        self.tokenizer = AutoTokenizer.from_pretrained("trl-internal-testing/tiny-Qwen2ForCausalLM-2.5")
        self.tokenizer.pad_token = self.tokenizer.eos_token
        dataset = load_dataset("trl-internal-testing/zen", "standard_prompt_only")
        dataset["train"] = dataset["train"].select(range(8))
        self.expected_winrates = [
            {"eval_win_rate": 0.5, "epoch": 0.0, "step": 0},
            {"eval_win_rate": 0.5, "epoch": 0.5, "step": 2},
            {"eval_win_rate": 0.5, "epoch": 1.0, "step": 4},
            {"eval_win_rate": 0.5, "epoch": 1.5, "step": 6},
            {"eval_win_rate": 0.5, "epoch": 2.0, "step": 8},
            {"eval_win_rate": 0.5, "epoch": 2.5, "step": 10},
            {"eval_win_rate": 0.5, "epoch": 3.0, "step": 12},
        ]

        def tokenize_function(examples):
            out = self.tokenizer(examples["prompt"], padding="max_length", max_length=16, truncation=True)
            out["labels"] = out["input_ids"].copy()
            return out

        self.dataset = dataset.map(tokenize_function, batched=True)

        self.generation_config = GenerationConfig(max_length=32)
        self.judge = HalfPairwiseJudge()

    def test_basic(self):
        with tempfile.TemporaryDirectory() as tmp_dir:
            training_args = TrainingArguments(
                output_dir=tmp_dir,
                eval_strategy="steps",
                eval_steps=2,  # evaluate every 2 steps
                per_device_train_batch_size=2,  # 8 samples in total so 4 batches of 2 per epoch
                per_device_eval_batch_size=2,
                report_to="none",
            )
            trainer = TrainerWithRefModel(
                model=self.model,
                ref_model=self.ref_model,
                args=training_args,
                train_dataset=self.dataset["train"],
                eval_dataset=self.dataset["test"],
                processing_class=self.tokenizer,
            )
            win_rate_callback = WinRateCallback(
                judge=self.judge, trainer=trainer, generation_config=self.generation_config
            )
            trainer.add_callback(win_rate_callback)
            trainer.train()
            winrate_history = [h for h in trainer.state.log_history if "eval_win_rate" in h]
            self.assertListEqual(winrate_history, self.expected_winrates)

    def test_without_ref_model(self):
        # Same as before, but without the ref_model attribute. It should use the model attribute instead
        with tempfile.TemporaryDirectory() as tmp_dir:
            training_args = TrainingArguments(
                output_dir=tmp_dir,
                eval_strategy="steps",
                eval_steps=2,  # evaluate every 2 steps
                per_device_train_batch_size=2,  # 8 samples in total so 4 batches of 2 per epoch
                per_device_eval_batch_size=2,
                report_to="none",
            )
            trainer = Trainer(
                model=self.model,
                args=training_args,
                train_dataset=self.dataset["train"],
                eval_dataset=self.dataset["test"],
                processing_class=self.tokenizer,
            )
            win_rate_callback = WinRateCallback(
                judge=self.judge, trainer=trainer, generation_config=self.generation_config
            )
            trainer.add_callback(win_rate_callback)
            trainer.train()
            winrate_history = [h for h in trainer.state.log_history if "eval_win_rate" in h]
            self.assertListEqual(winrate_history, self.expected_winrates)

    def test_soft_judge(self):
        """Test that the soft judge functionality works correctly"""
        with tempfile.TemporaryDirectory() as tmp_dir:
            training_args = TrainingArguments(
                output_dir=tmp_dir,
                eval_strategy="steps",
                eval_steps=2,  # evaluate every 2 steps
                per_device_train_batch_size=2,  # 8 samples in total so 4 batches of 2 per epoch
                per_device_eval_batch_size=2,
                report_to="none",
            )
            trainer = TrainerWithRefModel(
                model=self.model,
                ref_model=self.ref_model,
                args=training_args,
                train_dataset=self.dataset["train"],
                eval_dataset=self.dataset["test"],
                processing_class=self.tokenizer,
            )
            win_rate_callback = WinRateCallback(
                judge=self.judge, trainer=trainer, generation_config=self.generation_config, use_soft_judge=True
            )
            trainer.add_callback(win_rate_callback)
            trainer.train()

            # Expected values based on judge returning [0.3, 0.9] for each pair
            expected_soft_winrates = [
                {"eval_avg_win_prob": 0.4, "eval_win_rate": 0.5, "epoch": 0.0, "step": 0},
                {"eval_avg_win_prob": 0.4, "eval_win_rate": 0.5, "epoch": 0.5, "step": 2},
                {"eval_avg_win_prob": 0.4, "eval_win_rate": 0.5, "epoch": 1.0, "step": 4},
                {"eval_avg_win_prob": 0.4, "eval_win_rate": 0.5, "epoch": 1.5, "step": 6},
                {"eval_avg_win_prob": 0.4, "eval_win_rate": 0.5, "epoch": 2.0, "step": 8},
                {"eval_avg_win_prob": 0.4, "eval_win_rate": 0.5, "epoch": 2.5, "step": 10},
                {"eval_avg_win_prob": 0.4, "eval_win_rate": 0.5, "epoch": 3.0, "step": 12},
            ]

            winrate_history = [
                {k: h[k] for k in ["eval_avg_win_prob", "eval_win_rate", "epoch", "step"]}
                for h in trainer.state.log_history
                if "eval_avg_win_prob" in h
            ]
            self.assertListEqual(winrate_history, expected_soft_winrates)

    @require_peft
    def test_lora(self):
        with tempfile.TemporaryDirectory() as tmp_dir:
            peft_config = LoraConfig(
                r=16,
                lora_alpha=32,
                lora_dropout=0.05,
                bias="none",
                task_type="CAUSAL_LM",
            )
            self.model.add_adapter(peft_config)
            training_args = TrainingArguments(
                output_dir=tmp_dir,
                eval_strategy="steps",
                eval_steps=2,  # evaluate every 2 steps
                per_device_train_batch_size=2,  # 8 samples in total so 4 batches of 2 per epoch
                per_device_eval_batch_size=2,
                report_to="none",
            )
            trainer = Trainer(
                model=self.model,
                args=training_args,
                train_dataset=self.dataset["train"],
                eval_dataset=self.dataset["test"],
                processing_class=self.tokenizer,
            )
            win_rate_callback = WinRateCallback(
                judge=self.judge, trainer=trainer, generation_config=self.generation_config
            )
            trainer.add_callback(win_rate_callback)
            trainer.train()
            winrate_history = [h for h in trainer.state.log_history if "eval_win_rate" in h]
            self.assertListEqual(winrate_history, self.expected_winrates)


class LogCompletionsCallbackTester(unittest.TestCase):
    def setUp(self):
        self.model = AutoModelForCausalLM.from_pretrained("trl-internal-testing/tiny-Qwen2ForCausalLM-2.5")
        self.tokenizer = AutoTokenizer.from_pretrained("trl-internal-testing/tiny-Qwen2ForCausalLM-2.5")
        self.tokenizer.pad_token = self.tokenizer.eos_token
        dataset = load_dataset("trl-internal-testing/zen", "standard_prompt_only")
        dataset["train"] = dataset["train"].select(range(8))

        def tokenize_function(examples):
            out = self.tokenizer(examples["prompt"], padding="max_length", max_length=16, truncation=True)
            out["labels"] = out["input_ids"].copy()
            return out

        self.dataset = dataset.map(tokenize_function, batched=True)

        self.generation_config = GenerationConfig(max_length=32)

    @require_wandb
    def test_basic_wandb(self):
        import wandb

        with tempfile.TemporaryDirectory() as tmp_dir:
            training_args = TrainingArguments(
                output_dir=tmp_dir,
                eval_strategy="steps",
                eval_steps=2,  # evaluate every 2 steps
                per_device_train_batch_size=2,  # 8 samples in total so 4 batches of 2 per epoch
                per_device_eval_batch_size=2,
                report_to="wandb",
            )
            trainer = Trainer(
                model=self.model,
                args=training_args,
                train_dataset=self.dataset["train"],
                eval_dataset=self.dataset["test"],
                processing_class=self.tokenizer,
            )
            completions_callback = LogCompletionsCallback(trainer, self.generation_config, num_prompts=2)
            trainer.add_callback(completions_callback)
            trainer.train()

            # Get the current run
            completions_path = wandb.run.summary.completions["path"]
            json_path = os.path.join(wandb.run.dir, completions_path)
            with open(json_path) as f:
                completions = json.load(f)

            # Check that the columns are correct
            self.assertIn("step", completions["columns"])
            self.assertIn("prompt", completions["columns"])
            self.assertIn("completion", completions["columns"])

            # Check that the prompt is in the log
            self.assertIn(self.dataset["test"][0]["prompt"], completions["data"][0])

    @require_comet
    def test_basic_comet(self):
        import comet_ml

        with tempfile.TemporaryDirectory() as tmp_dir:
            training_args = TrainingArguments(
                output_dir=tmp_dir,
                eval_strategy="steps",
                eval_steps=2,  # evaluate every 2 steps
                per_device_train_batch_size=2,  # 8 samples in total so 4 batches of 2 per epoch
                per_device_eval_batch_size=2,
                report_to="comet_ml",
            )
            trainer = Trainer(
                model=self.model,
                args=training_args,
                train_dataset=self.dataset["train"],
                eval_dataset=self.dataset["test"],
                processing_class=self.tokenizer,
            )
            completions_callback = LogCompletionsCallback(trainer, self.generation_config, num_prompts=2)
            trainer.add_callback(completions_callback)
            trainer.train()

            # close experiment to make sure all pending data are flushed
            experiment = comet_ml.get_running_experiment()
            assert experiment is not None
            experiment.end()

            # get experiment assets and check that all required tables was logged
            steps = len(self.dataset["train"]) + len(self.dataset["test"])
            tables_logged = int(steps / 2) + 1  # +1 to include zero step

            api_experiment = comet_ml.APIExperiment(previous_experiment=experiment.id)
            tables = api_experiment.get_asset_list("dataframe")
            assert tables is not None
            assert len(tables) == tables_logged
            assert all(table["fileName"] == "completions.csv" for table in tables)


@require_mergekit
class MergeModelCallbackTester(unittest.TestCase):
    def setUp(self):
        self.model = AutoModelForCausalLM.from_pretrained("trl-internal-testing/tiny-Qwen2ForCausalLM-2.5")
        self.tokenizer = AutoTokenizer.from_pretrained("trl-internal-testing/tiny-Qwen2ForCausalLM-2.5")
        self.dataset = load_dataset("trl-internal-testing/zen", "standard_preference", split="train")

    def test_callback(self):
        with tempfile.TemporaryDirectory() as tmp_dir:
            training_args = DPOConfig(
                output_dir=tmp_dir,
                num_train_epochs=1,
                report_to="none",
                save_strategy="steps",
                save_steps=1,
            )
            config = MergeConfig()
            merge_callback = MergeModelCallback(config)
            trainer = DPOTrainer(
                model=self.model,
                args=training_args,
                train_dataset=self.dataset,
                processing_class=self.tokenizer,
                callbacks=[merge_callback],
            )
            trainer.train()
            last_checkpoint = get_last_checkpoint(tmp_dir)
            merged_path = os.path.join(last_checkpoint, "merged")
            self.assertTrue(os.path.isdir(merged_path), "Merged folder does not exist in the last checkpoint.")

    def test_every_checkpoint(self):
        with tempfile.TemporaryDirectory() as tmp_dir:
            training_args = DPOConfig(
                output_dir=tmp_dir,
                num_train_epochs=1,
                report_to="none",
                save_strategy="steps",
                save_steps=1,
            )
            config = MergeConfig()
            merge_callback = MergeModelCallback(config, merge_at_every_checkpoint=True)
            trainer = DPOTrainer(
                model=self.model,
                args=training_args,
                train_dataset=self.dataset,
                processing_class=self.tokenizer,
                callbacks=[merge_callback],
            )
            trainer.train()

            checkpoints = sorted(
                [os.path.join(tmp_dir, cp) for cp in os.listdir(tmp_dir) if cp.startswith("checkpoint-")]
            )

            for checkpoint in checkpoints:
                merged_path = os.path.join(checkpoint, "merged")
                self.assertTrue(
                    os.path.isdir(merged_path), f"Merged folder does not exist in checkpoint {checkpoint}."
                )