botp
/

File size: 18,263 Bytes
1d2de8a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""The answer contract engine: contract v3 on the reference engine's float32 cached path.

Adds to solomon/engine_reference.py (which is not modified):
  * a switchable LoRA layer, so an adapter can act on every token ('full', as the reference engine) or only on the
    question branch ('question'): off while the document is prefilled, on for the branch, and in full
    execution applied from the prefix boundary onward;
  * documents made of parts: text and page images, in any order. The vision tower runs once per document,
    the library's positions are used for the prefix, and the state carries the rope delta;
  * task blocks for every answer type, and ten answer letters;
  * state format v3.
"""
import copy
import hashlib
import json
from pathlib import Path

import numpy as np

from solomon.engine_reference import (BOOLEAN_TASK, CHOICE_TASK, CHUNK, SYSTEM, TOKEN_CAP, Engine, Ledger, boolean_block,  # noqa: F401
                            choice_block, dump, read_rows, sha, softmax)

ROOT = Path(__file__).resolve().parents[1]
MODEL = ROOT / 'runtime/qwen38-8bit'
ART = ROOT / 'artifacts/contract'
CONTRACT = 'solomon-answer-contract-v3'
STATE_FORMAT = 'solomon-answer-state-v3'
LETTERS = 'ABCDEFGHIJ'
PAGE = '<|vision_start|><|image_pad|><|vision_end|>'
IMAGE_PAD, VISION_END = '<|image_pad|>', '<|vision_end|>'
RESERVED = ['The document does not state this', 'The document gives conflicting answers']
NOT_STATED, CONFLICTING = 'not_stated', 'conflicting'

SINGLE_R = ('Task: choose the single option that the document establishes as the answer. If the document does not establish any of '
            'the other listed answers, choose the option that says it does not state this. If the document establishes two different '
            'listed answers and does not say which prevails, choose the option that says it gives conflicting answers. A replaced or '
            'withdrawn statement establishes nothing. Respond with exactly one letter. Do not explain.')
ORDERED_R = ('Task: the options form an ordered scale, lowest first, followed by two special options. Choose the single level that the '
             'document establishes. If the document does not establish any level, choose the option that says it does not state this. If '
             'it establishes two different levels and does not say which prevails, choose the option that says it gives conflicting '
             'answers. A replaced or withdrawn statement establishes nothing. Respond with exactly one letter. Do not explain.')
SINGLE_S = CHOICE_TASK
ORDERED_S = ('Task: the options form an ordered scale, lowest first. Choose the single level that the document best supports. '
             'Respond with exactly one letter. Do not explain.')
SUFFICIENCY = ('Task: classify what the document establishes about the answer to the question, using only the document. '
               'A = it establishes exactly one of the listed answers. B = it does not establish any of the listed answers. '
               'C = it establishes two or more different listed answers and does not say which prevails. '
               'A missing fact is not a negative fact. A replaced or withdrawn statement establishes nothing. '
               'Respond with exactly one letter: A, B, or C. Do not explain.')
LABEL = ('Task: decide whether the label applies, using only the document. A = the document establishes that it applies, and nothing in '
         'it establishes that it does not. B = the document establishes that it does not apply, and nothing in it establishes that it '
         'does. C = the document establishes neither. D = the document establishes both. A missing fact is not a negative fact. '
         'Evidence about another person or subject does not count. A replaced or withdrawn statement establishes nothing. '
         'Respond with exactly one letter: A, B, C, or D. Do not explain.')
OPTION = ('Task: decide whether the proposed answer is correct, using only the document. A = the document establishes this answer, and '
          'nothing in it establishes a different one. B = the document establishes a different answer, or establishes that this one is '
          'wrong, and nothing in it establishes this one. C = the document establishes neither. D = the document establishes both this '
          'answer and a different one. A missing fact is not a negative fact. A replaced or withdrawn statement establishes nothing. '
          'Respond with exactly one letter: A, B, C, or D. Do not explain.')


def _lettered(options):
    return '\n'.join(f'{LETTERS[i]}. {text}' for i, text in enumerate(options))


def listwise_block(question, options, ordered=False, reserved=True):
    """Caller options in caller order; with `reserved` the service appends the two reserved outcomes."""
    if not 2 <= len(options) <= 8 or len(set(options)) != len(options):
        raise ValueError('needs 2 to 8 distinct options')
    shown = list(options) + (RESERVED if reserved else [])
    task = (ORDERED_R if ordered else SINGLE_R) if reserved else (ORDERED_S if ordered else SINGLE_S)
    return task + '\nQuestion: ' + question + '\nOptions:\n' + _lettered(shown) + '\nAnswer (one letter):', len(shown)


def sufficiency_block(question, options):
    return SUFFICIENCY + '\nQuestion: ' + question + '\nListed answers:\n' + '\n'.join('- ' + o for o in options) + '\nAnswer (one letter):', 3


def label_block(question, label):
    return LABEL + '\nQuestion: ' + question + '\nLabel: ' + label + '\nAnswer (one letter):', 4


def option_block(question, option):
    return OPTION + '\nQuestion: ' + question + '\nProposed answer: ' + option + '\nAnswer (one letter):', 4


def four_state_block(question):
    return boolean_block(question), 4


CTL = {'start': None}   # None: adapter off. int s: adapter applied to positions >= s of the current call.


def _switch_lora_class():
    import mlx.core as mx
    import mlx.nn as nn

    class SwitchLoRA(nn.Module):
        def __init__(self, linear, lora_a, lora_b, scale):
            super().__init__()
            self.linear = linear
            self.lora_a = lora_a
            self.lora_b = lora_b
            self.scale = scale

        def __call__(self, x):
            y = self.linear(x)
            s = CTL['start']
            if s is None or s >= x.shape[1]:
                return y
            xs = x if s <= 0 else x[:, s:]
            z = (self.scale * ((xs @ self.lora_a) @ self.lora_b)).astype(y.dtype)
            if s <= 0:
                return y + z
            return mx.concatenate([y[:, :s], y[:, s:] + z], axis=1)

    return SwitchLoRA


def attach_adapter(lm, path, scale=2.0):
    """Wrap the projections named in the adapter file. Returns (wrapped count, rank)."""
    import mlx.core as mx
    weights = mx.load(str(path))
    cls = _switch_lora_class()
    names = sorted({k.rsplit('.', 1)[0] for k in weights})
    rank = None
    for name in names:
        parts = name.split('.')          # model.layers.N.owner.key
        owner = lm.model.layers[int(parts[2])]
        for p in parts[3:-1]:
            owner = getattr(owner, p)
        a, b = weights[name + '.lora_a'].astype(mx.float32), weights[name + '.lora_b'].astype(mx.float32)
        rank = a.shape[1]
        setattr(owner, parts[-1], cls(getattr(owner, parts[-1]), a, b, scale))
    lm.freeze()
    mx.eval(lm.parameters())
    return len(names), rank


def as_parts(document):
    """A document is a string (text) or a list of parts {'text': str} / {'image': path}."""
    if isinstance(document, str):
        return [{'text': document}]
    if not document or any(set(p) not in ({'text'}, {'image'}) for p in document):
        raise ValueError('parts must be {"text": ...} or {"image": ...}')
    return document


def document_key(document):
    h = hashlib.sha256()
    for p in as_parts(document):
        h.update(b'T' + p['text'].encode() if 'text' in p else b'I' + bytes.fromhex(sha(p['image'])))
    return h.hexdigest()


class ContractEngine(Engine):
    def __init__(self, arithmetic='float32', ledger=None, adapter=None, placement='question'):
        if placement not in ('question', 'full'):
            raise ValueError('placement must be question or full')
        super().__init__(arithmetic, ledger, adapter=None)
        self.placement = placement
        self.adapter = None
        self.rank = None
        if adapter is not None:
            self.wrapped, self.rank = attach_adapter(self.lm, adapter)
            self.adapter = str(adapter)
        self._pad = self.t.convert_tokens_to_ids(IMAGE_PAD)

    # ---- rendering -------------------------------------------------------------------
    def _content(self, parts):
        out = ''
        for i, p in enumerate(parts):
            if 'text' in p:
                out += ('\n' if i and 'image' in parts[i - 1] else '') + p['text']
            else:
                out += ('\n' if i and 'text' in parts[i - 1] else '') + PAGE
        return out

    def _text(self, parts, block):
        msgs = [{'role': 'system', 'content': SYSTEM}, {'role': 'user', 'content': 'Document:\n' + self._content(parts) + '\n\n' + block}]
        return self.t.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True, enable_thinking=False)

    def _expand(self, ids, counts):
        out, k = [], 0
        for tok in ids:
            if tok == self._pad:
                out += [tok] * counts[k]
                k += 1
            else:
                out.append(tok)
        if k != len(counts):
            raise ValueError('image placeholder count differs from the number of images')
        return out

    def _pixels(self, parts):
        from PIL import Image
        images = [Image.open(p['image']).convert('RGB') for p in parts if 'image' in p]
        if not images:
            return None, None, []
        out = self.processor.image_processor(images=images)
        grid = np.array(out['image_grid_thw'])
        merge = self.model.config.vision_config.spatial_merge_size ** 2
        counts = [int(g[0] * g[1] * g[2]) // merge for g in grid]
        return self.mx.array(np.array(out['pixel_values'])), self.mx.array(grid), counts

    def _vision(self, pixels, grid):
        """Vision tower one image at a time (memory stays flat for long scans)."""
        mx = self.mx
        dtype = self.model.vision_tower.patch_embed.proj.weight.dtype
        feats, offset = [], 0
        for i in range(grid.shape[0]):
            n = int(np.prod(np.array(grid[i])))
            h, _ = self.model.vision_tower(pixels[offset:offset + n].astype(dtype), grid[i:i + 1])
            mx.eval(h)
            feats.append(h)
            offset += n
        return mx.concatenate(feats, axis=0)

    # ---- execution -------------------------------------------------------------------
    def _forward(self, ids, pos, cache, embeds=None, chunk=CHUNK, adapter_from=None):
        """Run ids through the language model; logits for the final position. adapter_from indexes into ids."""
        mx = self.mx
        n = len(ids)
        if self.ledger is not None:
            self.ledger.add(forwards=1, input_tokens=n)
        step = n if chunk is None else chunk
        if cache is None and step < n:
            raise ValueError('Chunked execution needs a cache')
        logits = None
        for a in range(0, n, step):
            b = min(a + step, n)
            CTL['start'] = None if adapter_from is None else max(0, adapter_from - a)
            kwargs = dict(cache=cache, position_ids=pos[:, :, a:b], skip_logits=True)
            if embeds is not None:
                kwargs['inputs_embeds'] = embeds[:, a:b]
            if b == n:
                out = self.lm(mx.array([ids[a:b]]), return_hidden=True, **kwargs)
                lg = self.lm.lm_head(out.hidden_states[-1][:, -1:, :])[0, -1].astype(mx.float32)
                mx.eval(lg)
                logits = np.array(lg)
                del out
            else:
                self.lm(mx.array([ids[a:b]]), **kwargs)
                mx.eval([c.state for c in cache])
        CTL['start'] = None
        if not np.isfinite(logits).all():
            raise ValueError('Non-finite model output')
        if mx.get_peak_memory() > 100 * 2**30:
            raise MemoryError('100 GiB MLX allocation cap exceeded')
        return logits

    def _positions(self, start, n):
        mx = self.mx
        return mx.broadcast_to(mx.arange(start, start + n)[None, None, :], (3, 1, n))

    def prefill(self, document, chunk=CHUNK):
        """Encode the task-neutral document prefix once. Returns state format v3."""
        mx = self.mx
        parts = as_parts(document)
        text = self._text(parts, 'X')
        end = text.rfind('\n\nX')
        if end < 0:
            raise ValueError('Document boundary missing')
        raw = self.t.encode(text[:end], add_special_tokens=False)
        if 'text' in parts[-1]:
            raw = raw[:-1]    # leave the boundary token uncached: its tokenisation can depend on what follows
        pixels, grid, counts = self._pixels(parts)
        ids = self._expand(raw, counts)
        if not ids or len(ids) > TOKEN_CAP:
            raise ValueError(f'prefix of {len(ids)} tokens is empty or exceeds the {TOKEN_CAP}-token cap')
        embeds, delta, feats = None, 0, None
        if counts:
            feats = self._vision(pixels, grid)
            f = self.model.get_input_embeddings(mx.array([ids]), pixels, image_grid_thw=grid, cached_image_features=feats)
            embeds, pos = f.inputs_embeds, f.position_ids
            delta = int(np.array(f.rope_deltas).reshape(-1)[0])
            if delta != int(np.array(pos).max()) + 1 - len(ids):
                raise ValueError('rope delta disagrees with the prefix positions')
        else:
            pos = self._positions(0, len(ids))
        cache = self.lm.make_cache()
        self._forward(ids, pos, cache, embeds, chunk, adapter_from=0 if (self.adapter and self.placement == 'full') else None)
        mx.clear_cache()
        return {'format': STATE_FORMAT, 'prefix_ids': ids, 'cache': cache, 'rope_delta': delta, 'parts': parts,
                'image_tokens': int(sum(counts)), 'image_sha256': [sha(p['image']) for p in parts if 'image' in p],
                'layout': ''.join('T' if 'text' in p else 'I' for p in parts),
                'prefix_sha256': hashlib.sha256(json.dumps(ids).encode()).hexdigest(),
                '_pixels': pixels, '_grid': grid, '_counts': counts, '_feats': feats}

    def letters_for(self, text, n):
        ids = self.t.encode(text, add_special_tokens=False)
        out = []
        for letter in LETTERS[:n]:
            if letter not in self._letters:
                ext = self.t.encode(text + letter, add_special_tokens=False)
                if len(ext) != len(ids) + 1 or ext[:-1] != ids:
                    raise ValueError('Unstable answer-letter continuation')
                self._letters[letter] = ext[-1]
            out.append(self._letters[letter])
        return out, ids

    def ask(self, state, block, n_letters, execution='cached', full_chunk=None):
        """Answer-letter distribution for one task block against a prefilled state."""
        mx = self.mx
        text = self._text(state['parts'], block)
        letters, raw = self.letters_for(text, n_letters)
        ids = self._expand(raw, state['_counts'])
        prefix = state['prefix_ids']
        P = len(prefix)
        fallback = None
        if ids[:P] != prefix:
            execution, fallback = 'full', 'prefix token mismatch'
        on = self.adapter is not None
        if execution == 'cached':
            branch = copy.deepcopy(state['cache'])
            logits = self._forward(ids[P:], self._positions(P + state['rope_delta'], len(ids) - P), branch, chunk=None,
                                   adapter_from=0 if on else None)
            del branch
            reused = P
        else:
            embeds = None
            if state['_counts']:
                f = self.model.get_input_embeddings(mx.array([ids]), state['_pixels'], image_grid_thw=state['_grid'],
                                                    cached_image_features=None if len(state['_counts']) <= 6 else state['_feats'])
                embeds, pos = f.inputs_embeds, f.position_ids
            else:
                pos = self._positions(0, len(ids))
            start = None if not on else (0 if self.placement == 'full' else P)
            logits = self._forward(ids, pos, self.lm.make_cache() if full_chunk else None, embeds, full_chunk, adapter_from=start)
            reused = 0
        full = softmax(logits)
        mx.clear_cache()
        return {'letter_logits': logits[letters].astype(np.float64), 'probabilities': softmax(logits[letters]),
                'mass': float(full[letters].sum()), 'top_is_letter': int(logits.argmax()) in letters, 'execution': execution,
                'fallback': fallback, 'prompt_tokens': len(ids), 'reused_prefix_tokens': reused, 'branch_tokens': len(ids) - reused}


def identity(engine):
    """Runtime fingerprint for state format v3. With question placement the adapter is not part of the state."""
    from importlib.metadata import version
    ip = engine.processor.image_processor
    payload = {'contract': CONTRACT, 'state_format': STATE_FORMAT, 'system_sha256': hashlib.sha256(SYSTEM.encode()).hexdigest(),
               'arithmetic': engine.arithmetic, 'engine_sha256': sha(Path(__file__)), 'base_engine_sha256': sha(ROOT / 'solomon/engine_reference.py'),
               'model_config_sha256': sha(MODEL / 'config.json'),
               'image_processor': {'class': type(ip).__name__, 'config_sha256': sha(MODEL / 'preprocessor_config.json')
                                   if (MODEL / 'preprocessor_config.json').exists() else None},
               'state_adapter': (engine.adapter and sha(engine.adapter)) if engine.placement == 'full' else None,
               'versions': {p: version(p) for p in ('mlx', 'mlx-vlm', 'transformers')}}
    payload['fingerprint'] = hashlib.sha256(json.dumps(payload, sort_keys=True).encode()).hexdigest()
    payload['branch_adapter'] = engine.adapter and {'sha256': sha(engine.adapter), 'placement': engine.placement, 'rank': engine.rank}
    return payload