File size: 9,667 Bytes
5a46e5d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from __future__ import annotations

import json
from pathlib import Path

import pytest
import torch
import torch.nn.functional as F

from barunlm import BarunConfig, BarunLM
from barunlm import model as model_module

ROOT = Path(__file__).resolve().parents[1]


def tiny_config(**overrides: object) -> BarunConfig:
    values: dict[str, object] = {
        "vocab_size": 32,
        "dim": 16,
        "n_layers": 2,
        "n_heads": 2,
        "n_kv_heads": 1,
        "ffn_dim": 32,
        "max_seq_len": 16,
        "rope_fraction": 0.5,
        "local_window": 3,
        "full_attention_every": 2,
        "attention_gate": True,
        "qk_norm": True,
        "residual_select_every": 0,
        "dropout": 0.0,
        "tie_embeddings": True,
        "mtp_offset": 2,
        "mtp_loss_weight": 0.0,
    }
    values.update(overrides)
    return BarunConfig(**values)


def test_release_config_and_parameter_count() -> None:
    config = BarunConfig.from_json(ROOT / "barun_config.json")
    model = BarunLM(config)

    assert config.ffn_dim == 1_228
    assert config.mtp_loss_weight == 0.0
    assert model.parameter_counts() == {
        "total": 35_072_768,
        "embedding": 7_340_032,
        "non_embedding": 27_732_736,
    }
    assert model.lm_head.weight is model.embedding.weight


def test_default_config_matches_release_json() -> None:
    payload = json.loads((ROOT / "barun_config.json").read_text())
    assert BarunConfig() == BarunConfig(**payload)


def test_forward_shape_and_finite_logits() -> None:
    model = BarunLM(BarunConfig()).eval()
    with torch.inference_mode():
        output = model(torch.tensor([[1, 2, 3, 4]], dtype=torch.long))
    assert output.logits.shape == (1, 4, 16_384)
    assert torch.isfinite(output.logits).all()


def test_causal_loss_uses_conventional_next_token_shift() -> None:
    torch.manual_seed(1)
    model = BarunLM(tiny_config()).eval()
    input_ids = torch.tensor([[1, 2, 3, 4]], dtype=torch.long)

    output = model(input_ids, labels=input_ids)
    expected = F.cross_entropy(
        output.logits[:, :-1].reshape(-1, model.config.vocab_size),
        input_ids[:, 1:].reshape(-1),
    )

    assert output.causal_loss is not None
    torch.testing.assert_close(output.causal_loss, expected)


def test_completion_masks_and_mtp_targets_align_to_label_positions() -> None:
    torch.manual_seed(2)
    model = BarunLM(tiny_config(mtp_loss_weight=0.2)).eval()
    input_ids = torch.tensor([[1, 2, 3, 4, 5]], dtype=torch.long)
    labels = torch.tensor([[-100, -100, 3, -100, 5]], dtype=torch.long)
    captured_mtp_hidden: list[torch.Tensor] = []
    assert model.mtp_norm is not None
    hook = model.mtp_norm.register_forward_hook(
        lambda _module, _inputs, output: captured_mtp_hidden.append(output)
    )

    output = model(input_ids, labels=labels)
    hook.remove()

    expected_causal = F.cross_entropy(
        output.logits[:, :-1].reshape(-1, model.config.vocab_size),
        labels[:, 1:].reshape(-1),
        ignore_index=-100,
    )
    expected_mtp = F.cross_entropy(
        model.lm_head(captured_mtp_hidden[0]).reshape(-1, model.config.vocab_size),
        labels[:, model.config.mtp_offset :].reshape(-1),
        ignore_index=-100,
    )
    assert output.causal_loss is not None
    assert output.mtp_loss is not None
    torch.testing.assert_close(output.causal_loss, expected_causal)
    torch.testing.assert_close(output.mtp_loss, expected_mtp)
    torch.testing.assert_close(
        output.loss, expected_causal + model.config.mtp_loss_weight * expected_mtp
    )


def test_all_ignored_targets_return_finite_differentiable_zero() -> None:
    torch.manual_seed(3)
    model = BarunLM(tiny_config(mtp_loss_weight=0.2)).train()
    input_ids = torch.tensor([[1, 2, 3, 4]], dtype=torch.long)
    labels = torch.full_like(input_ids, -100)

    output = model(input_ids, labels=labels)

    assert output.loss is not None and output.loss.requires_grad
    assert output.causal_loss is not None and output.causal_loss.item() == 0.0
    assert output.mtp_loss is not None and output.mtp_loss.item() == 0.0
    assert torch.isfinite(output.loss)
    output.loss.backward()
    assert model.embedding.weight.grad is not None
    assert torch.count_nonzero(model.embedding.weight.grad) == 0


def test_single_token_sequence_has_finite_zero_causal_loss() -> None:
    model = BarunLM(tiny_config()).train()
    input_ids = torch.tensor([[1]], dtype=torch.long)

    output = model(input_ids, labels=input_ids)

    assert output.loss is not None and output.loss.item() == 0.0
    output.loss.backward()


def test_cached_local_mask_created_in_inference_mode_is_safe_for_backward() -> None:
    model_module._local_causal_mask.cache_clear()
    torch.manual_seed(4)
    model = BarunLM(tiny_config())
    input_ids = torch.tensor([[1, 2, 3, 4]], dtype=torch.long)

    model.eval()
    with torch.inference_mode():
        inference_output = model(input_ids)
    assert torch.isfinite(inference_output.logits).all()
    cached_mask = model_module._local_causal_mask(4, 3, "cpu")
    assert not cached_mask.is_inference()

    model.train()
    training_output = model(input_ids, labels=input_ids)
    assert training_output.loss is not None
    training_output.loss.backward()
    assert model.embedding.weight.grad is not None


def test_integer_padding_mask_rejects_values_other_than_zero_and_one() -> None:
    model = BarunLM(tiny_config()).eval()
    input_ids = torch.tensor([[1, 2, 3, 4], [5, 6, 7, 8]], dtype=torch.long)
    invalid = torch.tensor([[1, 1, 2, 1], [1, 1, 1, 1]], dtype=torch.long)

    with pytest.raises(ValueError, match="only 0 and 1"):
        model(input_ids, attention_mask=invalid)


def test_standard_batched_padding_mask_matches_unpadded_for_valid_tokens() -> None:
    torch.manual_seed(5)
    model = BarunLM(tiny_config()).eval()
    padded = torch.tensor([[0, 0, 5, 6], [7, 8, 9, 10]], dtype=torch.long)
    attention_mask = torch.tensor([[0, 0, 1, 1], [1, 1, 1, 1]], dtype=torch.long)

    with torch.inference_mode():
        batched = model(padded, attention_mask=attention_mask).logits
        short = model(torch.tensor([[5, 6]], dtype=torch.long)).logits
        long = model(torch.tensor([[7, 8, 9, 10]], dtype=torch.long)).logits

    torch.testing.assert_close(batched[0, 2:], short[0], atol=2e-6, rtol=2e-5)
    torch.testing.assert_close(batched[1], long[0], atol=2e-6, rtol=2e-5)


def test_additive_prefix_policy_mask_can_make_prompt_bidirectional() -> None:
    torch.manual_seed(6)
    model = BarunLM(tiny_config()).eval()
    first = torch.tensor([[1, 2, 3]], dtype=torch.long)
    changed_future = torch.tensor([[1, 9, 3]], dtype=torch.long)
    policy = torch.full((3, 3), float("-inf"))
    policy[torch.tril(torch.ones(3, 3, dtype=torch.bool))] = 0.0
    policy[0, 1] = 0.0

    with torch.inference_mode():
        causal_first = model(first).logits[:, 0]
        causal_changed = model(changed_future).logits[:, 0]
        prefix_first = model(first, attention_mask=policy).logits[:, 0]
        prefix_changed = model(changed_future, attention_mask=policy).logits[:, 0]

    torch.testing.assert_close(causal_first, causal_changed)
    assert not torch.allclose(prefix_first, prefix_changed)


def test_kv_cache_logits_match_full_forward() -> None:
    torch.manual_seed(7)
    model = BarunLM(tiny_config()).eval()
    input_ids = torch.tensor([[1, 2, 3, 4, 5, 6]], dtype=torch.long)

    with torch.inference_mode():
        expected = model(input_ids).logits
        cache = None
        pieces = []
        for position in range(input_ids.shape[1]):
            output = model(
                input_ids[:, position : position + 1],
                past_key_values=cache,
                use_cache=True,
                position_offset=position,
            )
            pieces.append(output.logits)
            cache = output.past_key_values
        actual = torch.cat(pieces, dim=1)

    torch.testing.assert_close(actual, expected, atol=2e-6, rtol=2e-5)


def test_left_padded_batched_generation_matches_individual_generation() -> None:
    torch.manual_seed(8)
    model = BarunLM(tiny_config()).eval()
    padded = torch.tensor([[0, 0, 5, 6], [7, 8, 9, 10]], dtype=torch.long)
    attention_mask = torch.tensor([[0, 0, 1, 1], [1, 1, 1, 1]], dtype=torch.long)

    batched = model.generate(padded, max_new_tokens=2, temperature=0, attention_mask=attention_mask)
    short = model.generate(torch.tensor([[5, 6]]), max_new_tokens=2, temperature=0)
    long = model.generate(torch.tensor([[7, 8, 9, 10]]), max_new_tokens=2, temperature=0)

    assert torch.equal(batched[0, -2:], short[0, -2:])
    assert torch.equal(batched[1, -2:], long[0, -2:])


def test_generation_stops_when_every_sequence_emits_eos() -> None:
    model = BarunLM(tiny_config(tie_embeddings=False)).eval()
    with torch.no_grad():
        for parameter in model.parameters():
            parameter.zero_()
    prompts = torch.tensor([[1, 2], [3, 4]], dtype=torch.long)

    output = model.generate(
        prompts,
        max_new_tokens=5,
        temperature=0,
        eos_token_id=0,
        pad_token_id=1,
    )

    assert output.shape == (2, 3)
    assert torch.equal(output[:, -1], torch.zeros(2, dtype=torch.long))


def test_batched_generation_rejects_right_padding() -> None:
    model = BarunLM(tiny_config()).eval()
    input_ids = torch.tensor([[5, 6, 0], [7, 8, 9]], dtype=torch.long)
    attention_mask = torch.tensor([[1, 1, 0], [1, 1, 1]], dtype=torch.long)

    with pytest.raises(ValueError, match="left padding"):
        model.generate(input_ids, max_new_tokens=1, attention_mask=attention_mask)