File size: 10,723 Bytes
7157974
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
'''
This file is inspired by the code from https://github.com/ML-GSAI/SMDM
'''
import accelerate
import torch
import re
from pathlib import Path
import random
import numpy as np
import torch.nn.functional as F
from datasets import Dataset
from lm_eval.__main__ import cli_evaluate
from lm_eval.api.instance import Instance
from lm_eval.api.model import LM
from lm_eval.api.registry import register_model
from tqdm import tqdm

from transformers import AutoTokenizer, AutoModel
from generate import generate


def set_seed(seed):
    torch.manual_seed(seed)
    random.seed(seed)
    np.random.seed(seed)

    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = False


@register_model("llada_dist")
class LLaDAEvalHarness(LM):
    def __init__(
        self,
        model_path='',
        mask_id=126336,
        max_length=4096,
        batch_size=32,
        mc_num=128,
        is_check_greedy=True,
        cfg=0.,
        steps=1024,
        gen_length=1024,
        block_length=1024,
        remasking='low_confidence',
        device="cuda",
        **kwargs,
    ):
        '''
        Args:
            model_path: LLaDA-8B-Base model path.
            mask_id: The token id of [MASK] is 126336.
            max_length: the max sequence length.
            batch_size: mini batch size.
            mc_num: Monte Carlo estimation iterations
            is_check_greedy: For certain metrics like LAMBADA, the evaluation requires the model to verify whether the answer 
                             is generated through greedy sampling conditioned on the prompt (note that this differs from conditional
                             generation). We implement this verification through the suffix_greedy_prediction() function, which 
                             returns a True/False judgment used for accuracy calculation. 
                             When is_check_greedy is set to True, the lm-evaluation-harness library automatically invokes this function. 
                             However, since none of the metrics in the LLaDA paper (https://arxiv.org/abs/2502.09992) require this functionality, 
                             we recommend setting is_check_greedy to False. This configuration causes suffix_greedy_prediction() to return False 
                             by default, significantly accelerating the evaluation process.
            cfg_scale: Unsupervised classifier-free guidance scale.
        '''
        super().__init__()

        accelerator = accelerate.Accelerator()
        if accelerator.num_processes > 1:
            self.accelerator = accelerator
        else:
            self.accelerator = None
        
        model_kwargs = {}
        if self.accelerator is not None:
            model_kwargs.update({'device_map': {'': f'{self.accelerator.device}'}})

        self.model = AutoModel.from_pretrained(model_path, trust_remote_code=True, torch_dtype=torch.bfloat16, **model_kwargs)
        self.model.eval()

        self.device = torch.device(device)
        if self.accelerator is not None:
            self.model = self.accelerator.prepare(self.model)
            self.device = torch.device(f'{self.accelerator.device}')
            self._rank = self.accelerator.local_process_index
            self._world_size = self.accelerator.num_processes
        else: 
            self.model = self.model.to(device)

        self.mask_id = mask_id
        self.tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)

        self.mc_num = mc_num
        self.batch_size = int(batch_size)
        assert mc_num % self.batch_size == 0
        self.sampling_eps = 0.
        self.max_length = max_length
        self.is_check_greedy = is_check_greedy

        self.cfg = cfg
        self.steps = steps
        self.gen_length = gen_length
        self.block_length = block_length
        self.remasking = remasking    
    @property
    def rank(self):
        return self._rank
    
    @property
    def world_size(self):
        return self._world_size

    def _forward_process(self, batch, prompt_index):
        b, l = batch.shape

        target_len = (l - prompt_index.sum()).item()
        k = torch.randint(1, target_len + 1, (), device=batch.device)

        x = torch.round(torch.linspace(float(k), k + (b - 1) * (target_len / b), steps=b, device=batch.device)).long()
        x = ((x - 1) % target_len) + 1
        assert x.min() >= 1 and x.max() <= target_len

        indices = torch.arange(target_len, device=batch.device).repeat(b, 1)
        is_mask = indices < x.unsqueeze(1)

        for i in range(b):
            is_mask[i] = is_mask[i][torch.randperm(target_len)]

        is_mask = torch.cat((torch.zeros(b, prompt_index.sum(), dtype=torch.bool, device=batch.device), is_mask), dim=1)

        noisy_batch = torch.where(is_mask, self.mask_id, batch)

        return noisy_batch, (x / target_len).unsqueeze(1).repeat(1, l)

    @torch.no_grad()
    def get_logits(self, batch, prompt_index):
        if self.cfg > 0.:
            assert len(prompt_index) == batch.shape[1]
            prompt_index = prompt_index.unsqueeze(0).repeat(batch.shape[0], 1)
            un_batch = batch.clone()
            un_batch[prompt_index] = self.mask_id
            batch = torch.cat([batch, un_batch])

        logits = self.model(batch).logits

        if self.cfg > 0.:
            logits, un_logits = torch.chunk(logits, 2, dim=0)
            logits = un_logits + (self.cfg + 1) * (logits - un_logits)
        return logits[:, :batch.shape[1]]

    @torch.no_grad()
    def get_loglikelihood(self, prefix, target):
        seq = torch.concatenate([prefix, target])[None, :]
        seq = seq.repeat((self.batch_size, 1)).to(self.device)

        prompt_index = torch.arange(seq.shape[1], device=self.device) < len(prefix)

        loss_acc = []
        for _ in range(self.mc_num // self.batch_size):
            perturbed_seq, p_mask = self._forward_process(seq, prompt_index)

            mask_indices = perturbed_seq == self.mask_id

            logits = self.get_logits(perturbed_seq, prompt_index)

            loss = F.cross_entropy(logits[mask_indices], seq[mask_indices], reduction='none') / p_mask[mask_indices]
            loss = loss.sum() / self.batch_size
            loss_acc.append(loss.item())

        return - sum(loss_acc) / len(loss_acc)

    @torch.no_grad()
    def suffix_greedy_prediction(self, prefix, target):
        if not self.is_check_greedy:
            return False

        seq = torch.full((1, len(prefix) + len(target)), self.mask_id, device=self.device)
        prompt_index = torch.arange(seq.shape[1], device=self.device) < len(prefix)
        prefix, target = prefix.to(self.device), target.to(self.device)
        seq[0, :len(prefix)] = prefix

        for i in range(len(target)):
            mask_index = (seq == self.mask_id)
            logits = self.get_logits(seq, prompt_index)[mask_index]
            x0 = torch.argmax(logits, dim=-1)

            p = torch.softmax(logits.to(torch.float32), dim=-1)
            confidence = torch.gather(p, dim=-1, index=torch.unsqueeze(x0, -1)).squeeze(dim=-1)
            _, index = torch.sort(confidence, descending=True)
            x0[index[1:]] = self.mask_id
            seq[mask_index] = x0.clone()
        correct = target == seq[0, len(prefix):]
        correct = torch.all(correct)
        return correct

    def _encode_pair(self, context, continuation):
        n_spaces = len(context) - len(context.rstrip())
        if n_spaces > 0:
            continuation = context[-n_spaces:] + continuation
            context = context[:-n_spaces]

        whole_enc = self.tokenizer(context + continuation)["input_ids"]
        context_enc = self.tokenizer(context)["input_ids"]

        context_enc_len = len(context_enc)
        continuation_enc = whole_enc[context_enc_len:]

        return context_enc, continuation_enc

    def loglikelihood(self, requests):
        def _tokenize(e):
            prefix, target = self._encode_pair(e["prefix"], e["target"])
            return {
                "prefix_text": e["prefix"],
                "target_text": e["target"],
                "prefix": prefix,
                "target": target,
            }

        ds = []
        ds = [{"prefix": req.args[0], "target": req.args[1]} for req in requests]
        ds = Dataset.from_list(ds)
        ds = ds.map(_tokenize)
        ds = ds.with_format("torch")
        prompt_len = [len(x["prefix"]) + len(x["target"]) for x in ds]

        assert max(prompt_len) <= 4096

        out = []
        with torch.no_grad():
            for elem in tqdm(ds, desc="Computing likelihood..."):
                prefix = elem["prefix"]
                target = elem["target"]

                ll = self.get_loglikelihood(prefix, target)

                is_target_greedy_dec = self.suffix_greedy_prediction(prefix, target)

                out.append((ll, 1.0 if is_target_greedy_dec else 0.0))
        torch.cuda.empty_cache()
        return out

    def loglikelihood_rolling(self, requests):
        raise NotImplementedError

    def generate_until(self, requests: list[Instance]):
        def _tokenize(e):
            return {
                "question": self.tokenizer(e["question"])["input_ids"],
                "question_text": e["question"],
                "until": e["until"],
            }

        ds = [{"question": req.args[0], "until": req.args[1]['until']} for req in requests]
        ds = Dataset.from_list(ds)
        ds = ds.map(_tokenize)
        ds = ds.with_format("torch")

        out = []
        for elem in tqdm(ds, desc="Generating..."):
            prompt = elem["question"].unsqueeze(0).to(self.device)
            stop_tokens = elem["until"]
 
            generated_answer = generate(self.model, prompt, steps=self.steps, gen_length=self.gen_length, block_length=self.block_length, 
                                        temperature=0, cfg_scale=self.cfg, remasking=self.remasking, mask_id=self.mask_id)
            
            generated_answer = self.tokenizer.decode(generated_answer[0][prompt.shape[1]:], skip_special_tokens=False)
            for stop_seq in stop_tokens:
                    if stop_seq in generated_answer:
                        generated_answer = generated_answer.split(stop_seq)[0]

            # remove special tokens
            generated_answer_ids = self.tokenizer(generated_answer)["input_ids"]
            generated_answer = self.tokenizer.decode(generated_answer_ids, skip_special_tokens=True)
            out.append(generated_answer)

            self.accelerator.wait_for_everyone()

        return out


if __name__ == "__main__":
    set_seed(1234)
    cli_evaluate()