File size: 10,477 Bytes
d9d6118
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""MiniNeuralHorner development model for modular multiplication.

The learned component is one modulus-conditioned recurrent transition cell.
It predicts the next binary residue state for

    s_next = (2 * s + d * x) mod p.

A fixed Horner schedule applies that cell to reduce both operands and then
multiply the two residues. The emitted answer is a list of base-2 digits. The
SAIR evaluation harness performs the final digit decoding.
"""

from __future__ import annotations

import hashlib
import json
from pathlib import Path

import torch
from torch import nn

from modchallenge.interface.base_model import ModularMultiplicationModel

_MASK32 = (1 << 32) - 1
_CHECKPOINT_SCHEMA = "mini-neuralhorner-inference-v1"
_EXPECTED_CONFIG = {
    "bidirectional": True,
    "dmodel": 96,
    "hidden": 61,
    "num_layers": 2,
}
_EXPECTED_PARAMETERS = 126_603
_EXPECTED_WIDTH = 2_048
_EXPECTED_TENSOR_SHA256 = (
    "7d1768ae1260f750e0a80ec93d98f86a80d441e479ce29e0a8e21fd098c742a3"
)
_QUALIFIED_CHECKPOINT_SHA256 = (
    "d296b711bb6a7faaa1dd81e05478cfa75f11071c42a8c36fbf60e758ee7eb407"
)


def _to_bits_small(values: torch.Tensor, width: int) -> torch.Tensor:
    shifts = torch.arange(width - 1, -1, -1, device=values.device)
    return (values[:, None] >> shifts[None, :]) & 1


def to_bits_limbs(values: list[int], device: torch.device, width: int) -> torch.Tensor:
    """Convert nonnegative Python integers to MSB-first bits without int64 overflow."""
    limb_count = (width + 31) // 32
    columns = []
    for limb_index in range(limb_count - 1, -1, -1):
        limb = torch.tensor(
            [(value >> (32 * limb_index)) & _MASK32 for value in values],
            dtype=torch.int64,
            device=device,
        )
        columns.append(_to_bits_small(limb, 32))
    bits = torch.cat(columns, dim=1)
    excess = limb_count * 32 - width
    return bits[:, excess:] if excess else bits


def _tensor_digest(state_dict: dict[str, torch.Tensor]) -> str:
    """Hash tensor names, dtypes, shapes, and raw values in a stable order."""
    digest = hashlib.sha256()
    for name in sorted(state_dict):
        tensor = state_dict[name].detach().cpu().contiguous()
        header = json.dumps(
            {"dtype": str(tensor.dtype), "name": name, "shape": list(tensor.shape)},
            sort_keys=True,
            separators=(",", ":"),
        ).encode("utf-8")
        raw = tensor.numpy().tobytes(order="C")
        digest.update(len(header).to_bytes(8, "big"))
        digest.update(header)
        digest.update(len(raw).to_bytes(8, "big"))
        digest.update(raw)
    return digest.hexdigest()


class TransitionCell(nn.Module):
    def __init__(
        self,
        dmodel: int,
        hidden: int,
        num_layers: int,
        bidirectional: bool,
    ) -> None:
        super().__init__()
        directions = 2 if bidirectional else 1
        self.in_proj = nn.Linear(3, dmodel)
        self.d_emb = nn.Embedding(2, dmodel)
        self.gru = nn.GRU(
            dmodel,
            hidden,
            num_layers=num_layers,
            batch_first=True,
            bidirectional=bidirectional,
        )
        self.head = nn.Linear(directions * hidden, 1)

    def forward(
        self,
        features: torch.Tensor,
        control: torch.Tensor,
    ) -> torch.Tensor:
        embedded = self.in_proj(features) + self.d_emb(control)[:, None, :]
        hidden, _ = self.gru(embedded)
        return self.head(hidden).squeeze(-1)


def _bits_of(value: int) -> list[int]:
    if value <= 0:
        return [0]
    digits = []
    while value:
        digits.append(value & 1)
        value >>= 1
    digits.reverse()
    return digits


class MiniNeuralHorner(ModularMultiplicationModel):
    """SAIR interface adapter for the 126,603-parameter transition cell."""

    def __init__(self) -> None:
        self.model: TransitionCell | None = None
        self.device = torch.device("cpu")
        self.width = _EXPECTED_WIDTH
        self._sequence_width = 32

    def load(self, model_dir: str) -> None:
        checkpoint_path = Path(model_dir) / "weights.pt"
        checkpoint = torch.load(checkpoint_path, map_location="cpu", weights_only=True)
        if checkpoint.get("schema") != _CHECKPOINT_SCHEMA:
            raise ValueError("unsupported MiniNeuralHorner checkpoint schema")
        if checkpoint.get("config") != _EXPECTED_CONFIG:
            raise ValueError("checkpoint architecture does not match the packaged model")
        if checkpoint.get("L") != _EXPECTED_WIDTH:
            raise ValueError("checkpoint inference width does not match the packaged model")
        provenance = checkpoint.get("provenance", {})
        if provenance.get("qualified_checkpoint_sha256") != _QUALIFIED_CHECKPOINT_SHA256:
            raise ValueError("checkpoint provenance does not match the qualified source")

        state_dict = checkpoint.get("state_dict")
        if not isinstance(state_dict, dict):
            raise ValueError("checkpoint is missing its state_dict")
        parameter_count = sum(tensor.numel() for tensor in state_dict.values())
        if parameter_count != _EXPECTED_PARAMETERS:
            raise ValueError("checkpoint parameter count is invalid")
        if _tensor_digest(state_dict) != _EXPECTED_TENSOR_SHA256:
            raise ValueError("checkpoint tensor digest is invalid")

        if torch.cuda.is_available():
            self.device = torch.device("cuda")
            torch.backends.cudnn.benchmark = False
            torch.backends.cudnn.deterministic = True
            torch.backends.cudnn.allow_tf32 = False
            torch.backends.cuda.matmul.allow_tf32 = False
        elif torch.backends.mps.is_available():
            self.device = torch.device("mps")

        self.model = TransitionCell(**_EXPECTED_CONFIG)
        self.model.load_state_dict(state_dict, strict=True)
        self.model.to(self.device)
        self.model.eval()

    def preprocess_a(self, a: str) -> list[int]:
        return _bits_of(int(a))

    def preprocess_b(self, b: str) -> list[int]:
        return _bits_of(int(b))

    def preprocess_p(self, p: str) -> int:
        return int(p)

    @torch.no_grad()
    def predict_digits(
        self,
        a_enc: list[int],
        b_enc: list[int],
        p_enc: int,
    ) -> list[int]:
        return self.predict_digits_batch([(a_enc, b_enc, p_enc)])[0]

    @torch.no_grad()
    def predict_digits_batch(
        self,
        inputs: list[tuple[list[int], list[int], int]],
    ) -> list[list[int]]:
        if self.model is None:
            raise RuntimeError("load() must be called before inference")

        max_operand_bits = 4 * self.width
        outputs: list[list[int]] = [[0] for _ in inputs]
        valid_indices = []
        a_bit_lists = []
        b_bit_lists = []
        moduli = []
        for index, (a_enc, b_enc, p_enc) in enumerate(inputs):
            modulus = int(p_enc)
            a_bits = list(a_enc)
            b_bits = list(b_enc)
            if (
                modulus < 2
                or modulus >= (1 << self.width)
                or len(a_bits) > max_operand_bits
                or len(b_bits) > max_operand_bits
            ):
                continue
            valid_indices.append(index)
            a_bit_lists.append(a_bits)
            b_bit_lists.append(b_bits)
            moduli.append(modulus)

        if not valid_indices:
            return outputs

        maximum_modulus_bits = max(modulus.bit_length() for modulus in moduli)
        self._sequence_width = min(
            self.width,
            max(32, ((maximum_modulus_bits + 31) // 32) * 32),
        )
        modulus_bits = to_bits_limbs(
            moduli,
            self.device,
            self._sequence_width,
        ).float()
        a_residues = self._reduce(a_bit_lists, modulus_bits)
        b_residues = self._reduce(b_bit_lists, modulus_bits)
        product = self._multiply(a_residues, b_residues, modulus_bits)

        for result_index, input_index in enumerate(valid_indices):
            outputs[input_index] = [int(bit) for bit in product[result_index].long().tolist()]
        return outputs

    def max_batch_size(self) -> int:
        return 256

    def _step(
        self,
        state_bits: torch.Tensor,
        multiplicand_bits: torch.Tensor,
        modulus_bits: torch.Tensor,
        control: torch.Tensor,
    ) -> torch.Tensor:
        if self.model is None:
            raise RuntimeError("model is not loaded")
        features = torch.stack(
            [state_bits, multiplicand_bits, modulus_bits],
            dim=-1,
        )
        logits = self.model(features, control)
        return (torch.sigmoid(logits) > 0.5).float()

    def _reduce(
        self,
        bit_lists: list[list[int]],
        modulus_bits: torch.Tensor,
    ) -> torch.Tensor:
        batch_size = len(bit_lists)
        operand_width = max(len(bits) for bits in bit_lists)
        padded = torch.zeros(
            (batch_size, operand_width),
            dtype=torch.long,
            device=self.device,
        )
        for row, bits in enumerate(bit_lists):
            if bits:
                padded[row, operand_width - len(bits) :] = torch.tensor(
                    bits,
                    dtype=torch.long,
                    device=self.device,
                )
        state_bits = torch.zeros(
            (batch_size, self._sequence_width),
            device=self.device,
        )
        one_bits = to_bits_limbs(
            [1] * batch_size,
            self.device,
            self._sequence_width,
        ).float()
        for position in range(operand_width):
            state_bits = self._step(
                state_bits,
                one_bits,
                modulus_bits,
                padded[:, position],
            )
        return state_bits

    def _multiply(
        self,
        a_residue: torch.Tensor,
        b_residue: torch.Tensor,
        modulus_bits: torch.Tensor,
    ) -> torch.Tensor:
        batch_size = a_residue.shape[0]
        state_bits = torch.zeros(
            (batch_size, self._sequence_width),
            device=self.device,
        )
        b_digits = b_residue.long()
        for position in range(self._sequence_width):
            state_bits = self._step(
                state_bits,
                a_residue,
                modulus_bits,
                b_digits[:, position],
            )
        return state_bits