File size: 9,866 Bytes
fd0c71a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""Post-training inference validation for adapter or merged model artifacts."""

from __future__ import annotations

import argparse
import json
from pathlib import Path
import re
import time
from typing import Any

import sys

ROOT = Path(__file__).resolve().parents[1]
if str(ROOT) not in sys.path:
    sys.path.insert(0, str(ROOT))

from app.env.env_core import PolyGuardEnv
from app.common.normalization import clamp_reward


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description="Validate inference from saved adapter/merged artifacts.")
    parser.add_argument("--merged-model", default="checkpoints/merged")
    parser.add_argument("--adapter-dir", default="checkpoints/sft_adapter")
    parser.add_argument("--base-model", default="")
    parser.add_argument("--prompts", default="data/processed/training_corpus_grpo_prompts.jsonl")
    parser.add_argument("--samples", type=int, default=3)
    parser.add_argument("--output", default="outputs/reports/postsave_inference.json")
    return parser.parse_args()


def _load_prompt_rows(path: Path, limit: int) -> list[dict[str, Any]]:
    if not path.exists():
        return []
    rows: list[dict[str, Any]] = []
    with path.open("r", encoding="utf-8") as handle:
        for line in handle:
            line = line.strip()
            if not line:
                continue
            try:
                payload = json.loads(line)
            except json.JSONDecodeError:
                continue
            if isinstance(payload, dict):
                rows.append(payload)
            if len(rows) >= limit:
                break
    return rows


def _prompt_to_text(row: dict[str, Any]) -> str:
    prompt = row.get("prompt", {}) if isinstance(row.get("prompt"), dict) else {}
    candidates = prompt.get("candidates", prompt.get("candidate_set", []))
    candidate_ids = [
        str(item.get("candidate_id"))
        for item in candidates
        if isinstance(item, dict) and item.get("candidate_id")
    ]
    text = {
        "instruction": "Choose one candidate_id and justify briefly.",
        "patient_id": prompt.get("patient_id", prompt.get("patient_summary", {}).get("patient_id", "unknown")),
        "candidate_ids": candidate_ids,
        "format": "candidate_id=<cand_xx>; rationale=<text>",
    }
    return json.dumps(text, ensure_ascii=True)


def _discover_base_model(adapter_dir: Path) -> str:
    cfg = adapter_dir / "adapter_config.json"
    if not cfg.exists():
        return ""
    try:
        payload = json.loads(cfg.read_text(encoding="utf-8"))
    except json.JSONDecodeError:
        return ""
    value = payload.get("base_model_name_or_path")
    return str(value) if isinstance(value, str) else ""


def _load_model(
    merged_model: Path,
    adapter_dir: Path,
    base_model_arg: str,
):
    import torch
    from transformers import AutoModelForCausalLM, AutoTokenizer

    if merged_model.exists() and (merged_model / "config.json").exists():
        tokenizer = AutoTokenizer.from_pretrained(str(merged_model))
        model = AutoModelForCausalLM.from_pretrained(
            str(merged_model),
            torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
            low_cpu_mem_usage=True,
        )
        source = "merged"
        return model, tokenizer, source

    if not adapter_dir.exists():
        raise FileNotFoundError(f"adapter_dir_not_found:{adapter_dir}")

    from peft import PeftModel

    base_model = base_model_arg.strip() or _discover_base_model(adapter_dir)
    if not base_model:
        raise RuntimeError("missing_base_model_for_adapter")

    tokenizer = AutoTokenizer.from_pretrained(base_model)
    base = AutoModelForCausalLM.from_pretrained(
        base_model,
        torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
        low_cpu_mem_usage=True,
    )
    model = PeftModel.from_pretrained(base, str(adapter_dir))
    source = "adapter"
    return model, tokenizer, source


def _fallback_completion(row: dict[str, Any]) -> tuple[str, str | None]:
    prompt = row.get("prompt", {}) if isinstance(row.get("prompt"), dict) else {}
    candidates = prompt.get("candidates", prompt.get("candidate_set", []))
    candidate_ids = [
        str(item.get("candidate_id"))
        for item in candidates
        if isinstance(item, dict) and item.get("candidate_id")
    ]
    candidate_id = candidate_ids[0] if candidate_ids else None
    completion = (
        f"candidate_id={candidate_id}; rationale=fallback_policy_artifact"
        if candidate_id
        else "candidate_id=cand_01; rationale=fallback_policy_artifact"
    )
    return completion, candidate_id


def _extract_candidate_id(text: str) -> str | None:
    match = re.search(r"cand_\d+", text.lower())
    if not match:
        return None
    return match.group(0)


def main() -> None:
    args = parse_args()
    root = Path(__file__).resolve().parents[1]
    merged_model = (root / args.merged_model).resolve()
    adapter_dir = (root / args.adapter_dir).resolve()
    prompts_path = (root / args.prompts).resolve()

    rows = _load_prompt_rows(prompts_path, limit=max(1, args.samples))
    if not rows:
        raise SystemExit(f"no_prompts_loaded:{prompts_path}")

    fallback_policy_file = (root / "checkpoints" / "sft_policy_fallback.json").resolve()
    model = None
    tokenizer = None
    model_source = "fallback_policy"
    model_load_error = ""
    try:
        model, tokenizer, model_source = _load_model(
            merged_model=merged_model,
            adapter_dir=adapter_dir,
            base_model_arg=args.base_model,
        )
        if tokenizer.pad_token is None:
            tokenizer.pad_token = tokenizer.eos_token
    except Exception as exc:  # noqa: BLE001
        model_load_error = str(exc)
        if not fallback_policy_file.exists():
            raise

    import torch

    device = "cuda" if torch.cuda.is_available() else "cpu"
    if model is not None:
        model = model.to(device)
        model.eval()

    env = PolyGuardEnv()
    results: list[dict[str, Any]] = []
    for idx, row in enumerate(rows):
        env.reset(seed=17_000 + idx, difficulty="medium")
        prompt_text = _prompt_to_text(row)
        started = time.perf_counter()

        if model is not None and tokenizer is not None:
            encoded = tokenizer(prompt_text, return_tensors="pt", truncation=True, max_length=512)
            encoded = {key: value.to(device) for key, value in encoded.items()}
            with torch.no_grad():
                generated = model.generate(
                    **encoded,
                    max_new_tokens=80,
                    do_sample=False,
                    temperature=0.0,
                    eos_token_id=tokenizer.eos_token_id,
                )
            decoded = tokenizer.decode(generated[0], skip_special_tokens=True)
            completion = decoded[len(prompt_text) :].strip() if decoded.startswith(prompt_text) else decoded
            candidate_id = _extract_candidate_id(completion)
        else:
            completion, candidate_id = _fallback_completion(row)
        latency_seconds = time.perf_counter() - started

        all_actions = env.get_candidate_actions()
        legal_actions = env.get_legal_actions()
        by_id_all = {str(item.get("candidate_id", "")).lower(): item for item in all_actions}
        by_id_legal = {str(item.get("candidate_id", "")).lower(): item for item in legal_actions}
        action = by_id_legal.get(str(candidate_id or "").lower())
        if action is None:
            action = by_id_all.get(str(candidate_id or "").lower())
        if action is None and legal_actions:
            action = legal_actions[0]

        if action is None:
            results.append(
                {
                    "idx": idx,
                    "prompt": prompt_text,
                    "completion": completion,
                    "candidate_id": candidate_id,
                    "selected_candidate": None,
                    "env_reward": 0.001,
                    "latency_seconds": round(latency_seconds, 3),
                    "valid": False,
                    "reason": "no_action_available",
                }
            )
            continue

        _, reward, done, info = env.step(action)
        results.append(
            {
                "idx": idx,
                "prompt": prompt_text,
                "completion": completion,
                "candidate_id": candidate_id,
                "selected_candidate": action.get("candidate_id"),
                "env_reward": clamp_reward(float(reward)),
                "latency_seconds": round(latency_seconds, 3),
                "done": bool(done),
                "valid": bool(info.get("safety_report", {}).get("legal", False)),
                "termination_reason": info.get("termination_reason"),
            }
        )

    valid_rate = sum(1.0 for row in results if row.get("valid")) / len(results)
    avg_reward = clamp_reward(sum(float(row.get("env_reward", 0.0)) for row in results) / len(results))
    avg_latency_seconds = round(
        sum(float(row.get("latency_seconds", 0.0)) for row in results) / len(results),
        3,
    )

    payload = {
        "status": "ok",
        "model_source": model_source,
        "model_load_error": model_load_error,
        "samples": len(results),
        "valid_rate": round(valid_rate, 3),
        "avg_env_reward": avg_reward,
        "avg_latency_seconds": avg_latency_seconds,
        "results": results,
    }

    output_path = root / args.output
    output_path.parent.mkdir(parents=True, exist_ok=True)
    output_path.write_text(json.dumps(payload, ensure_ascii=True, indent=2), encoding="utf-8")
    print("postsave_inference_ok")


if __name__ == "__main__":
    main()