File size: 7,393 Bytes
4a28d4d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from __future__ import annotations

import unittest
from types import SimpleNamespace

import torch
from torch import nn

from metrics.phase2_critic_guided_math import _planned_processed_samples
from networks.acdir import ACDiRPolicy
from training.phase2_critic.rollouts import rollout_count_set_policy


class _TinyTokenizer:
    eos_token_id = 18
    pad_token_id = 0
    all_special_ids = [0, 18]
    chat_template = None

    def convert_tokens_to_ids(self, token):
        return 19 if token == "<|mdm_mask|>" else -1

    def __call__(self, prompts, return_tensors="pt", padding=True, **_):
        input_ids = torch.tensor([[1, 2] for _ in prompts], dtype=torch.long)
        return {"input_ids": input_ids, "attention_mask": torch.ones_like(input_ids)}

    def batch_decode(self, seqs, skip_special_tokens=True):
        return [" ".join(map(str, row.tolist())) for row in seqs]

    def decode(self, seq, skip_special_tokens=True):
        return " ".join(map(str, seq.tolist()))


class _TinyActor(nn.Module):
    def __init__(self):
        super().__init__()
        self.config = SimpleNamespace(hidden_size=8, vocab_size=20)
        self.emb = nn.Embedding(20, 8)
        with torch.no_grad():
            self.emb.weight.zero_()
            self.emb.weight[:, 0] = torch.arange(20, dtype=torch.float32)

    def get_input_embeddings(self):
        return self.emb

    def forward(
        self,
        input_ids,
        attention_mask=None,
        logits_indices=None,
        output_hidden_states=False,
        output_last_hidden_state=False,
        **_,
    ):
        hidden = self.emb(input_ids)
        positions = logits_indices.cpu()
        logits = torch.full((input_ids.shape[0], positions.shape[1], 20), -10.0)
        top = (positions % 17 + 1).long()
        logits.scatter_(2, top.unsqueeze(-1), 10.0)
        hidden_states = (hidden,) if output_last_hidden_state else None
        return SimpleNamespace(logits=logits, hidden_states=hidden_states)


class _CountSetCritic(nn.Module):
    count_support = (0, 1, 2)

    def __init__(self):
        super().__init__()
        self.bias = nn.Parameter(torch.zeros(()))

    def forward_policy(
        self,
        hidden_states,
        token_embeddings,
        time_embed,
        candidate_mask,
        **_,
    ):
        token_scores = token_embeddings[..., 0].to(hidden_states.dtype) + self.bias
        count_logits = torch.tensor(
            [[-1.0, 2.0, 0.0]],
            dtype=hidden_states.dtype,
            device=hidden_states.device,
        ).expand(hidden_states.shape[0], -1).contiguous()
        return {
            "token_scores": token_scores,
            "remask_token_logits": token_scores,
            "count_logits": count_logits,
            "count_support": self.count_support,
            "retention_prior_logits": token_scores,
            "delta_value_logits": torch.zeros_like(token_scores),
            "state_value": hidden_states.new_zeros((hidden_states.shape[0],)),
            "encoded": hidden_states,
        }


def _rollout(**overrides):
    kwargs = dict(
        actor=_TinyActor(),
        critic=_CountSetCritic(),
        tokenizer=_TinyTokenizer(),
        batch={"problems": ["a", "bb"]},
        reward_fn=None,
        device=torch.device("cpu"),
        precision_dtype=torch.float32,
        time_embed_dim=8,
        steps=4,
        gen_length=4,
        block_length=2,
        no_sample=True,
        mask_id=19,
        eos_id=18,
        compute_rewards=False,
        return_responses=True,
        lookback_blocks=1,
        remask_min_age_current=0,
        max_total_remask_per_sample=2,
        reforward_after_remask=True,
    )
    kwargs.update(overrides)
    return rollout_count_set_policy(**kwargs)


class ReleaseRegressionTest(unittest.TestCase):
    def test_clean_progress_planning_handles_padded_ranks(self):
        self.assertEqual(_planned_processed_samples(5, 2, 2, 0), 4)
        self.assertEqual(_planned_processed_samples(5, 2, 2, 1), 5)
        self.assertEqual(_planned_processed_samples(6, 4, 1, 0), 4)
        self.assertEqual(_planned_processed_samples(6, 4, 1, 1), 6)

    def test_empty_context_rows_are_made_attention_safe(self):
        class Recorder(nn.Module):
            def __init__(self):
                super().__init__()
                self.padding_mask = None

            def forward(self, encoded, src_key_padding_mask=None):
                self.padding_mask = src_key_padding_mask.detach().clone()
                return encoded

        policy = ACDiRPolicy(
            hidden_size=4,
            time_embed_dim=4,
            mlp_hidden=8,
            policy_dim=8,
            encoder_layers=1,
            encoder_heads=2,
            dropout=0.0,
        )
        recorder = Recorder()
        policy.encoder = recorder
        candidate = torch.tensor([[False, False, False], [True, False, True]])
        context = torch.tensor([[False, False, False], [False, True, False]])
        policy.forward_policy(
            torch.randn(2, 3, 4),
            torch.randn(2, 3, 4),
            torch.randn(2, 4),
            candidate,
            context_mask=context,
        )
        safe_context = ~recorder.padding_mask
        self.assertEqual(safe_context.tolist(), [[True, False, False], [True, True, True]])

    def test_deterministic_rollout_matches_preoptimization_golden(self):
        torch.manual_seed(123)
        out = _rollout()
        self.assertEqual(out["tokens"].tolist(), [[1, 2, 3, 4, 5, 6]] * 2)
        self.assertEqual(out["responses"], ["3 4 5 6"] * 2)
        self.assertEqual(out["remask_counts"].tolist(), [2, 2])
        self.assertEqual(out["unique_remask_counts"].tolist(), [2, 2])
        self.assertEqual(out["unmask_decisions"].tolist(), [4, 4])
        self.assertEqual(out["policy_decisions"].tolist(), [4, 4])
        self.assertEqual(out["count0_decisions"].tolist(), [2, 2])
        self.assertEqual(out["count1_decisions"].tolist(), [2, 2])
        self.assertEqual(out["count2_decisions"].tolist(), [0, 0])

    def test_fork_resume_trace_has_fresh_time_left(self):
        common = dict(
            no_sample=False,
            temperature=0.4,
            return_tokens=False,
            return_responses=False,
            acdir_deferred_unmask=False,
            sample_remask=False,
        )
        capture = _rollout(mti_capture_transition=True, **common)
        self.assertIsNotNone(capture["mti_transition_state"])
        repaired = _rollout(
            mti_transition_state=capture["mti_transition_state"],
            cfpg_branch="remask",
            return_mti_trace=True,
            **common,
        )
        self.assertGreaterEqual(len(repaired["mti_trace"]["repair_steps"]), 1)
        for step in repaired["mti_trace"]["repair_steps"]:
            self.assertEqual(tuple(step["time_left"].shape), (2,))

    def test_ablation_trace_has_fresh_time_left(self):
        torch.manual_seed(0)
        out = _rollout(
            critic=None,
            ablation_remask_policy="low_confidence",
            ablation_remask_probability=1.0,
            return_mti_trace=True,
        )
        self.assertGreaterEqual(len(out["mti_trace"]["repair_steps"]), 1)
        for step in out["mti_trace"]["repair_steps"]:
            self.assertEqual(tuple(step["time_left"].shape), (2,))


if __name__ == "__main__":
    unittest.main()