File size: 7,250 Bytes
53f0cc2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Component 7: Inference engine for local code generation.

Features:
- Deterministic low-temperature greedy mode.
- Stop rules for clean function completion.
- Syntax-aware retry with up to 3 attempts.
"""

from __future__ import annotations

import ast
from dataclasses import dataclass
from typing import Dict, List, Optional, Tuple

import torch

from src.evaluation_system.code_eval import restore_code_from_structured
from src.model_architecture.code_transformer import CodeTransformerLM
from src.tokenizer.code_tokenizer import CodeTokenizer


@dataclass
class DecodingConfig:
    max_new_tokens: int = 300
    # Mode 1: deterministic output
    greedy_temperature: float = 0.0
    # Retry mode 2
    retry2_temperature: float = 0.25
    retry2_top_p: float = 0.85
    # Retry mode 3
    retry3_temperature: float = 0.35
    retry3_top_p: float = 0.90
    max_retries: int = 3
    min_tokens_before_stop_check: int = 64
    # Stop only when function body is non-trivial.
    min_function_body_statements: int = 2


class InferenceEngine:
    def __init__(self, model: CodeTransformerLM, tokenizer: CodeTokenizer, device: torch.device) -> None:
        self.model = model
        self.tokenizer = tokenizer
        self.device = device
        self.model.eval()

    @staticmethod
    def _syntax_ok_python(code: str) -> bool:
        try:
            ast.parse(code)
            return True
        except Exception:
            return False

    @staticmethod
    def _function_completion_score(code: str) -> int:
        # Higher score = more complete usable function.
        try:
            tree = ast.parse(code)
        except Exception:
            return 0
        funcs = [n for n in tree.body if isinstance(n, ast.FunctionDef)]
        if not funcs:
            return 0
        fn = funcs[-1]
        body_len = len(fn.body)
        has_return = any(isinstance(n, ast.Return) for n in ast.walk(fn))
        return body_len + (2 if has_return else 0)

    def _looks_complete_function(self, code: str, min_body_statements: int) -> bool:
        if "def " not in code:
            return False
        try:
            tree = ast.parse(code)
        except Exception:
            return False
        funcs = [n for n in tree.body if isinstance(n, ast.FunctionDef)]
        if not funcs:
            return False
        fn = funcs[-1]
        if len(fn.body) < min_body_statements:
            return False
        return True

    def _sample_next(
        self,
        logits: torch.Tensor,
        temperature: float,
        top_p: float,
    ) -> torch.Tensor:
        if temperature <= 0:
            return torch.argmax(logits, dim=-1, keepdim=True)

        logits = logits / temperature
        probs = torch.softmax(logits, dim=-1)

        sorted_probs, sorted_idx = torch.sort(probs, descending=True)
        cumulative = torch.cumsum(sorted_probs, dim=-1)
        cutoff = cumulative > top_p
        cutoff[..., 1:] = cutoff[..., :-1].clone()
        cutoff[..., 0] = False
        sorted_probs[cutoff] = 0.0
        denom = sorted_probs.sum(dim=-1, keepdim=True).clamp_min(1e-12)
        sorted_probs = sorted_probs / denom
        sampled = torch.multinomial(sorted_probs, num_samples=1)
        return sorted_idx.gather(-1, sampled)

    @torch.no_grad()
    def _generate_once(
        self,
        prompt: str,
        language: str,
        max_new_tokens: int,
        temperature: float,
        top_p: float,
        min_tokens_before_stop_check: int,
        min_function_body_statements: int,
    ) -> Dict[str, object]:
        prompt_text = self.tokenizer.format_training_sample(prompt=prompt, code="", language=language)
        prompt_text = prompt_text.replace(" <NL>", "").strip()

        ids = self.tokenizer.encode(prompt_text)
        eos_id = self.tokenizer.special_token_ids.get("<EOS>")

        # Remove trailing EOS so generation can continue.
        if eos_id is not None and len(ids) > 1 and ids[-1] == int(eos_id):
            ids = ids[:-1]

        input_ids = torch.tensor([ids], dtype=torch.long, device=self.device)

        generated_steps = 0
        for _ in range(max_new_tokens):
            out = self.model(input_ids=input_ids)
            logits = out["logits"][:, -1, :]
            next_id = self._sample_next(logits, temperature=temperature, top_p=top_p)
            input_ids = torch.cat([input_ids, next_id], dim=1)
            generated_steps += 1

            # Primary stop: EOS token.
            if eos_id is not None and int(next_id.item()) == int(eos_id):
                break

            # Secondary stop: complete parseable function with non-trivial body.
            if generated_steps >= min_tokens_before_stop_check and (generated_steps % 12 == 0):
                decoded = self.tokenizer.decode(input_ids[0].tolist())
                code = restore_code_from_structured(decoded)
                if self._looks_complete_function(code, min_body_statements=min_function_body_statements):
                    break

        decoded = self.tokenizer.decode(input_ids[0].tolist())
        code = restore_code_from_structured(decoded)
        syntax_ok = self._syntax_ok_python(code) if language == "python" else True
        completion_score = self._function_completion_score(code) if language == "python" else 0
        return {
            "code": code,
            "syntax_ok": syntax_ok,
            "generated_tokens": generated_steps,
            "temperature": temperature,
            "top_p": top_p,
            "completion_score": completion_score,
        }

    @torch.no_grad()
    def generate_with_retry(
        self,
        prompt: str,
        language: str = "python",
        cfg: Optional[DecodingConfig] = None,
    ) -> Dict[str, object]:
        cfg = cfg or DecodingConfig()

        attempts: List[Tuple[float, float]] = [
            (cfg.greedy_temperature, 1.0),
            (cfg.retry2_temperature, cfg.retry2_top_p),
            (cfg.retry3_temperature, cfg.retry3_top_p),
        ]

        results = []
        for i in range(min(cfg.max_retries, len(attempts))):
            temp, top_p = attempts[i]
            res = self._generate_once(
                prompt=prompt,
                language=language,
                max_new_tokens=cfg.max_new_tokens,
                temperature=temp,
                top_p=top_p,
                min_tokens_before_stop_check=cfg.min_tokens_before_stop_check,
                min_function_body_statements=cfg.min_function_body_statements,
            )
            res["attempt"] = i + 1
            results.append(res)

            # Syntax-aware retry: stop retries as soon as syntax is valid.
            if bool(res["syntax_ok"]):
                return {
                    "final": res,
                    "attempts": results,
                    "used_retry": i > 0,
                }

        # If all retries fail, choose best completion score then longest generation.
        best = sorted(
            results,
            key=lambda x: (int(x.get("completion_score", 0)), int(x.get("generated_tokens", 0))),
            reverse=True,
        )[0]
        return {
            "final": best,
            "attempts": results,
            "used_retry": True,
        }