File size: 10,240 Bytes
219ee1a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
bbc71d8
 
 
 
 
 
 
 
219ee1a
 
 
 
 
 
 
 
 
 
 
bbc71d8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
219ee1a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
bbc71d8
 
 
219ee1a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Inference pipeline.

Loads ONLY the trained CriterionWiseScoringSystem checkpoint (best_model_v5.pt)
locally. The CodeT5 encoder lives in a separate Hugging Face Space β€” we call
it over HTTP for hidden states. The Qwen explanation LLM is called via the
Hugging Face Inference API. See model/encoder_client.py and model/explainer.py.

Net effect: the deployment server only needs the small .pt checkpoint plus
the cross-attention layers in memory (~150–200 MB). No transformers / no
CodeT5 / no Qwen on the server.
"""
import logging
import os
import time

import torch

from .architecture import CriterionWiseScoringSystem
from .signals import extract_signals_from_attn
from .explainer import build_explanation, normalize_signals
from .encoder_client import encode_text

logger = logging.getLogger("inference")


# ── Config (overridable via env vars) ──────────────────────────────────────
CHECKPOINT_PATH = os.getenv("CHECKPOINT_PATH", "best_model_v5.pt")

# Optional: pull the checkpoint from a Hugging Face Hub model repo at boot.
# Useful for deployments where you don't want the .pt baked into the git repo.
# If set, takes precedence over CHECKPOINT_PATH.
CHECKPOINT_HF_REPO = os.getenv("CHECKPOINT_HF_REPO", "").strip()
CHECKPOINT_HF_FILENAME = os.getenv("CHECKPOINT_HF_FILENAME", "best_model_v5.pt")
CHECKPOINT_HF_REVISION = os.getenv("CHECKPOINT_HF_REVISION", "main")

EMBEDDING_DIM = int(os.getenv("EMBEDDING_DIM", "1024"))
N_HEADS = int(os.getenv("N_HEADS", "8"))
N_CROSS_LAYERS = int(os.getenv("N_CROSS_LAYERS", "2"))
DEVICE = os.getenv("DEVICE", "cuda" if torch.cuda.is_available() else "cpu")


# ── Module-level singletons (initialised by load_pipeline) ─────────────────
_model = None
_device = None


def _resolve_checkpoint_path() -> str:
    """
    Return a local filesystem path to the checkpoint, downloading from the
    Hugging Face Hub if CHECKPOINT_HF_REPO is set. Otherwise falls back to
    the local CHECKPOINT_PATH.

    huggingface_hub caches by content hash under ~/.cache/huggingface/hub,
    so repeated boots inside the same container don't re-download.
    """
    if CHECKPOINT_HF_REPO:
        # Lazy import: only needed when the env var is actually set, so local
        # dev (with the .pt sitting next to app.py) doesn't pay the import cost.
        from huggingface_hub import hf_hub_download
        token = os.getenv("HF_TOKEN") or None
        logger.info(
            "Downloading checkpoint from HF Hub: repo=%s file=%s rev=%s",
            CHECKPOINT_HF_REPO, CHECKPOINT_HF_FILENAME, CHECKPOINT_HF_REVISION,
        )
        t0 = time.perf_counter()
        path = hf_hub_download(
            repo_id=CHECKPOINT_HF_REPO,
            filename=CHECKPOINT_HF_FILENAME,
            revision=CHECKPOINT_HF_REVISION,
            token=token,
        )
        logger.info(
            "Checkpoint ready at %s (fetch/cache took %.1fs)",
            path, time.perf_counter() - t0,
        )
        return path

    if not os.path.exists(CHECKPOINT_PATH):
        raise FileNotFoundError(
            f"Checkpoint not found at '{CHECKPOINT_PATH}' and CHECKPOINT_HF_REPO is unset. "
            f"Either place best_model_v5.pt in the project root, set CHECKPOINT_PATH, "
            f"or set CHECKPOINT_HF_REPO to a HF model repo (e.g. 'user/pseudoscorex-checkpoint')."
        )
    return CHECKPOINT_PATH


def load_pipeline():
    """Load the trained scoring model. Call once at boot."""
    global _model, _device

    if _model is not None:
        return  # already loaded

    t0 = time.perf_counter()
    device = torch.device(DEVICE)
    logger.info("Loading scoring head on device: %s", device)

    model = CriterionWiseScoringSystem(
        embedding_dim=EMBEDDING_DIM,
        n_heads=N_HEADS,
        n_cross_layers=N_CROSS_LAYERS,
        dropout=0.1,
    ).to(device)

    checkpoint_path = _resolve_checkpoint_path()
    logger.info("Loading checkpoint: %s", checkpoint_path)
    state_dict = torch.load(checkpoint_path, map_location=device)
    model.load_state_dict(state_dict)
    model.eval()

    _model = model
    _device = device
    logger.info("Pipeline ready in %.1fs total.", time.perf_counter() - t0)


def _validate_payload(question, answer, criteria):
    if not isinstance(question, str) or not question.strip():
        raise ValueError("'question' must be a non-empty string")
    if not isinstance(answer, str) or not answer.strip():
        raise ValueError("'answer' must be a non-empty string")
    if not isinstance(criteria, list) or not criteria:
        raise ValueError("'criteria' must be a non-empty list")
    for i, c in enumerate(criteria):
        if not isinstance(c, dict):
            raise ValueError(f"criteria[{i}] must be an object")
        if "name" not in c:
            raise ValueError(f"criteria[{i}] missing 'name'")
        if "max_score" not in c:
            raise ValueError(f"criteria[{i}] missing 'max_score'")
        try:
            float(c["max_score"])
        except (TypeError, ValueError):
            raise ValueError(f"criteria[{i}].max_score must be numeric")


# ── Public API ─────────────────────────────────────────────────────────────
def predict(question, answer, criteria):
    """
    Score one (question, answer) pair against a list of rubric criteria.

    Same signature and return shape as the original implementation; the
    encoder is now remote and the LLM uses the HF Inference API.
    """
    if _model is None:
        raise RuntimeError("Pipeline not loaded. Call load_pipeline() first.")

    _validate_payload(question, answer, criteria)

    t_total = time.perf_counter()
    logger.info(
        "Starting prediction: question=%d chars, answer=%d chars, %d criteria",
        len(question), len(answer), len(criteria),
    )

    # Encode question and answer ONCE β€” they are shared across criteria
    t_enc = time.perf_counter()
    q_hidden, q_mask, question_tokens = encode_text(question, _device)
    a_hidden, a_mask, answer_tokens = encode_text(answer, _device)
    logger.info("Encoded question + answer via Space (%.2fs)", time.perf_counter() - t_enc)

    criterion_results = []
    total_score = 0.0
    max_total = 0.0
    n_crit = len(criteria)

    for idx, crit in enumerate(criteria, start=1):
        t_crit = time.perf_counter()
        name = crit["name"]
        desc = crit.get("description") or name
        max_s = float(crit["max_score"])
        logger.info("[%d/%d] Criterion '%s' (max=%.0f) β†’ scoring…", idx, n_crit, name, max_s)

        # Notebook always prepends the <criterion> sentinel β€” the Space's
        # tokenizer knows about it (we add it during Space boot).
        c_text = "<criterion> " + desc
        c_hidden, c_mask, criterion_tokens = encode_text(c_text, _device)

        t_fwd = time.perf_counter()
        with torch.no_grad():
            pred_norm, attn_dict = _model(
                q_hidden, a_hidden, c_hidden,
                A_mask=a_mask, c_mask=c_mask,
                return_attn=True,
            )

        pred_norm = float(pred_norm.item())
        score_val = pred_norm * max_s
        logger.info(
            "[%d/%d] score=%.2f/%.0f (pred_norm=%.4f) in %.2fs",
            idx, n_crit, score_val, max_s, pred_norm, time.perf_counter() - t_fwd,
        )

        signals = extract_signals_from_attn(
            attn_dict,
            answer_tokens=answer_tokens,
            criterion_tokens=criterion_tokens,
            question_tokens=question_tokens,
        )
        signals = normalize_signals(signals)

        t_expl = time.perf_counter()
        logger.info("[%d/%d] Generating explanation via HF Inference API…", idx, n_crit)
        explanation = build_explanation(
            criterion_name=name,
            pred_norm=pred_norm,
            score=score_val,
            max_score=max_s,
            signals=signals,
            question=question,
            answer=answer,
            criterion_desc=desc,
        )
        logger.info(
            "[%d/%d] Explanation ready in %.2fs (criterion total %.2fs)",
            idx, n_crit, time.perf_counter() - t_expl, time.perf_counter() - t_crit,
        )

        criterion_results.append({
            "criterion": name,
            "criterion_description": desc,
            "score": round(score_val, 3),
            "max_score": max_s,
            "pred_norm": round(pred_norm, 4),
            "explanation": explanation,
        })

        total_score += score_val
        max_total += max_s

    percentage = (total_score / max_total * 100.0) if max_total > 0 else 0.0
    overall_explanation = _build_overall_explanation(
        total_score, max_total, criterion_results
    )

    logger.info(
        "Prediction done: total=%.2f/%.0f (%.1f%%) in %.2fs",
        total_score, max_total, percentage, time.perf_counter() - t_total,
    )

    return {
        "question": question,
        "answer": answer,
        "total_score": round(total_score, 3),
        "max_total_score": max_total,
        "percentage": round(percentage, 2),
        "results": criterion_results,
        "overall_explanation": overall_explanation,
    }


def _build_overall_explanation(total, max_total, criterion_results):
    """Tiny narrative stitched together from the per-criterion results."""
    if max_total == 0:
        return "No criteria were scorable."

    pct = total / max_total * 100.0
    if pct >= 85:
        band = "excellent"
    elif pct >= 70:
        band = "strong"
    elif pct >= 50:
        band = "partial"
    else:
        band = "weak"

    strong = [r["criterion"] for r in criterion_results if r["pred_norm"] >= 0.8]
    weak = [r["criterion"] for r in criterion_results if r["pred_norm"] < 0.5]

    parts = [f"Overall {band} performance: {total:.2f} / {max_total:.0f} ({pct:.1f}%)."]
    if strong:
        parts.append("Strongest on: " + ", ".join(strong) + ".")
    if weak:
        parts.append("Needs work on: " + ", ".join(weak) + ".")
    return " ".join(parts)