File size: 12,914 Bytes
e1ced61
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Deterministic Qwen3.5-VL prediction for frozen evaluation manifests.

GPU libraries remain lazy imports so manifest/scoring tests run on CPU-only
machines.  The runner is append-resumable by eval ID and never re-hashes model,
adapter, or image trees; paths are checked structurally when they are used.
"""

from __future__ import annotations

import copy
import json
import os
import time
from collections.abc import Mapping, Sequence
from pathlib import Path
from typing import Any

from ..hashing import canonical_json
from ..paths import repo_root
from ..training.answers import parse_answer
from .core import EvaluationError


def resolve_asset(root: Path, raw_path: str) -> Path:
    relative = Path(raw_path)
    if not raw_path or relative.is_absolute() or ".." in relative.parts or "\\" in raw_path:
        raise EvaluationError(f"unsafe evaluation image path: {raw_path!r}")
    resolved_root = root.resolve()
    resolved = (resolved_root / relative).resolve()
    try:
        resolved.relative_to(resolved_root)
    except ValueError as exc:
        raise EvaluationError(f"evaluation image escapes asset root: {raw_path!r}") from exc
    if not resolved.is_file():
        raise EvaluationError(f"evaluation image not found: {resolved}")
    return resolved


def _question_text(question: str, choices: Sequence[Mapping[str, Any]]) -> str:
    if not choices:
        return question
    rendered: list[str] = []
    for choice in choices:
        if "key" not in choice or "text" not in choice:
            raise EvaluationError("evaluation choice requires key and text")
        rendered.append(f"{choice['key']}. {choice['text']}")
    return question + "\n\nChoices:\n" + "\n".join(rendered)


def build_messages(
    row: Mapping[str, Any],
    *,
    images: Sequence[Any],
) -> list[dict[str, Any]]:
    question = row.get("question")
    choices = row.get("choices")
    if not isinstance(question, str) or not question:
        raise EvaluationError("evaluation row has no question")
    if not isinstance(choices, list) or any(not isinstance(choice, Mapping) for choice in choices):
        raise EvaluationError("evaluation row choices are malformed")
    prompt_path = repo_root() / "prompts" / "common_system.txt"
    try:
        system = prompt_path.read_text(encoding="utf-8").strip()
    except OSError as exc:
        raise EvaluationError(f"cannot read common system prompt: {exc}") from exc
    content = [{"type": "image", "image": image} for image in images]
    content.append({"type": "text", "text": _question_text(question, choices)})
    return [
        {"role": "system", "content": [{"type": "text", "text": system}]},
        {"role": "user", "content": content},
    ]


def _load_images(row: Mapping[str, Any], asset_root: Path) -> list[Any]:
    from PIL import Image

    raw_images = row.get("images")
    if not isinstance(raw_images, list):
        raise EvaluationError("evaluation row images must be a list")
    images: list[Any] = []
    for raw in raw_images:
        if not isinstance(raw, Mapping) or not isinstance(raw.get("path"), str):
            raise EvaluationError("evaluation image record is malformed")
        path = resolve_asset(asset_root, str(raw["path"]))
        try:
            with Image.open(path) as image:
                images.append(image.convert("RGB").copy())
        except (OSError, ValueError) as exc:
            raise EvaluationError(f"cannot decode evaluation image {path}: {exc}") from exc
    return images


def _load_model(base_model: Path, adapter: Path | None) -> tuple[Any, Any]:
    try:
        import torch
        from transformers import AutoModelForMultimodalLM, AutoProcessor
    except ImportError as exc:
        raise EvaluationError(f"GPU evaluation dependency is missing: {exc}") from exc
    if not base_model.is_dir():
        raise EvaluationError(f"base model directory not found: {base_model}")
    processor = AutoProcessor.from_pretrained(str(base_model))
    model = AutoModelForMultimodalLM.from_pretrained(
        str(base_model),
        dtype=torch.bfloat16,
        attn_implementation="flash_attention_2",
        device_map="auto",
    )
    if adapter is not None:
        if not adapter.is_dir():
            raise EvaluationError(f"adapter directory not found: {adapter}")
        try:
            from peft import PeftModel
        except ImportError as exc:
            raise EvaluationError(f"PEFT is required for adapter evaluation: {exc}") from exc
        model = PeftModel.from_pretrained(model, str(adapter), is_trainable=False)
    model.eval()
    return model, processor


def _model_device(model: Any) -> Any:
    device = getattr(model, "device", None)
    if device is not None:
        return device
    try:
        return next(model.parameters()).device
    except (AttributeError, StopIteration) as exc:
        raise EvaluationError("cannot determine evaluation model device") from exc


def _predict_one(
    row: Mapping[str, Any],
    *,
    asset_root: Path,
    model: Any,
    processor: Any,
    max_prompt_tokens: int,
    max_new_tokens: int,
    image_row: Mapping[str, Any] | None = None,
    question_only: bool = False,
) -> dict[str, Any]:
    try:
        import torch
        from qwen_vl_utils import process_vision_info
    except ImportError as exc:
        raise EvaluationError(f"GPU evaluation dependency is missing: {exc}") from exc
    images = [] if question_only else _load_images(image_row or row, asset_root)
    messages = build_messages(row, images=images)
    rendered = processor.apply_chat_template(
        copy.deepcopy(messages),
        tokenize=False,
        add_generation_prompt=True,
        enable_thinking=False,
    )
    image_inputs, video_inputs = process_vision_info(messages)
    processor_kwargs: dict[str, Any] = {
        "text": [rendered],
        "padding": True,
        "return_tensors": "pt",
    }
    if image_inputs:
        processor_kwargs["images"] = image_inputs
    if video_inputs:
        processor_kwargs["videos"] = video_inputs
    encoded = processor(**processor_kwargs)
    prompt_tokens = int(encoded["input_ids"].shape[-1])
    if prompt_tokens > max_prompt_tokens:
        raise EvaluationError(
            f"{row.get('eval_id')}: prompt has {prompt_tokens} tokens, limit is {max_prompt_tokens}"
        )
    encoded = encoded.to(_model_device(model))
    started = time.monotonic()
    with torch.inference_mode():
        generated = model.generate(
            **encoded,
            do_sample=False,
            max_new_tokens=max_new_tokens,
            stop_strings=["</answer>"],
            tokenizer=getattr(processor, "tokenizer", processor),
        )
    elapsed = time.monotonic() - started
    completion_ids = generated[:, encoded["input_ids"].shape[-1] :]
    response = processor.batch_decode(
        completion_ids,
        skip_special_tokens=True,
        clean_up_tokenization_spaces=False,
    )[0]
    token_count = int(completion_ids.shape[-1])
    parsed = parse_answer(response)
    return {
        "response": response,
        "prompt_tokens": prompt_tokens,
        "completion_tokens": token_count,
        "prompt_truncated": False,
        "completion_truncated": bool(token_count >= max_new_tokens and not parsed.valid),
        "latency_seconds": elapsed,
    }


def _existing_predictions(path: Path, *, run_id: str) -> dict[str, dict[str, Any]]:
    if not path.exists():
        return {}
    rows: dict[str, dict[str, Any]] = {}
    try:
        with path.open("r", encoding="utf-8") as handle:
            for line_number, line in enumerate(handle, start=1):
                if not line.strip():
                    continue
                value = json.loads(line)
                if not isinstance(value, dict):
                    raise EvaluationError(f"{path}:{line_number}: prediction is not an object")
                if value.get("run_id") != run_id:
                    raise EvaluationError(f"{path}: existing predictions belong to another run")
                eval_id = value.get("eval_id")
                if not isinstance(eval_id, str) or not eval_id or eval_id in rows:
                    raise EvaluationError(f"{path}: duplicate or empty existing eval_id")
                rows[eval_id] = value
    except (OSError, json.JSONDecodeError) as exc:
        raise EvaluationError(f"cannot resume prediction file {path}: {exc}") from exc
    return rows


def predict_run(
    rows: Sequence[Mapping[str, Any]],
    *,
    run_id: str,
    base_model: Path,
    adapter: Path | None,
    asset_root: Path,
    output_path: Path,
    max_prompt_tokens: int = 4096,
    max_new_tokens: int = 256,
    limit: int | None = None,
    input_mode: str = "standard",
    constant_answer: str | None = None,
) -> dict[str, Any]:
    """Generate one exact, resumable prediction row per evaluation view."""

    if not run_id:
        raise EvaluationError("run_id must be non-empty")
    if max_prompt_tokens <= 0 or max_new_tokens <= 0:
        raise EvaluationError("evaluation token limits must be positive")
    if input_mode not in {"standard", "question_only", "full_image_control"}:
        raise EvaluationError(f"unsupported evaluation input mode: {input_mode!r}")
    if constant_answer is not None and not parse_answer(
        f"<answer>{constant_answer}</answer>"
    ).valid:
        raise EvaluationError("constant answer cannot be encoded by the public answer schema")
    selected = list(rows[:limit] if limit is not None else rows)
    existing = _existing_predictions(output_path, run_id=run_id)
    if any(row.get("input_mode", "standard") != input_mode for row in existing.values()):
        raise EvaluationError("existing predictions use another input mode")
    expected_ids = {str(row.get("eval_id", "")) for row in selected}
    if not set(existing).issubset(expected_ids):
        raise EvaluationError("existing prediction file contains IDs outside this evaluation")
    pending = [row for row in selected if str(row.get("eval_id", "")) not in existing]
    model: Any | None = None
    processor: Any | None = None
    if pending and constant_answer is None:
        model, processor = _load_model(base_model, adapter)
    full_by_group: dict[str, Mapping[str, Any]] = {}
    if input_mode == "full_image_control":
        for row in selected:
            if row.get("state") == "FULL":
                full_by_group[str(row.get("group_id", ""))] = row
        missing_full = sorted(
            {
                str(row.get("group_id", ""))
                for row in selected
                if str(row.get("group_id", "")) not in full_by_group
            }
        )
        if missing_full:
            raise EvaluationError(
                f"full-image control lacks FULL rows for groups: {missing_full[:5]}"
            )
    output_path.parent.mkdir(parents=True, exist_ok=True)
    with output_path.open("a", encoding="utf-8") as handle:
        for index, row in enumerate(pending, start=1):
            if constant_answer is None:
                assert model is not None and processor is not None
                prediction = _predict_one(
                    row,
                    asset_root=asset_root,
                    model=model,
                    processor=processor,
                    max_prompt_tokens=max_prompt_tokens,
                    max_new_tokens=max_new_tokens,
                    image_row=full_by_group.get(str(row.get("group_id", ""))),
                    question_only=input_mode == "question_only",
                )
            else:
                prediction = {
                    "response": f"<answer>{constant_answer}</answer>",
                    "prompt_tokens": 0,
                    "completion_tokens": 0,
                    "prompt_truncated": False,
                    "completion_truncated": False,
                    "latency_seconds": 0.0,
                }
            value = {
                "schema_version": 1,
                "run_id": run_id,
                "eval_id": str(row["eval_id"]),
                "input_mode": input_mode,
                **prediction,
            }
            handle.write(canonical_json(value) + "\n")
            handle.flush()
            if index % 32 == 0:
                os.fsync(handle.fileno())
        if pending:
            os.fsync(handle.fileno())
    all_rows = _existing_predictions(output_path, run_id=run_id)
    truncated = sum(bool(row.get("completion_truncated")) for row in all_rows.values())
    return {
        "run_id": run_id,
        "output": str(output_path.resolve()),
        "expected": len(selected),
        "predicted": len(all_rows),
        "new_predictions": len(pending),
        "completion_truncation_count": truncated,
        "input_mode": input_mode,
        "constant_answer": constant_answer,
        "complete": len(all_rows) == len(selected),
    }