File size: 5,341 Bytes
0dd6c2f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import json
from types import TracebackType
from typing import Any

from linalg_zero.generator.models import ComponentResult, DifficultyCategory
from linalg_zero.generator.sympy.template_engine import TemplateEngine
from linalg_zero.shared.types import LibTypes


class ProblemContext:
    """
    Context manager for state information around the resolution process.
    """

    def __init__(self, entropy: float, difficulty_level: DifficultyCategory, step_counter: int):
        self.entropy = entropy
        self.difficulty_level = difficulty_level
        self.used_entropy = 0.0
        self.tool_calls_count = 0
        self.stepwise_results: list[dict[str, Any]] = []
        self.golden_result: dict[str, str] = {}
        self._step_counter = step_counter
        self.constraints: dict[str, Any] = {}

    def __enter__(self) -> "ProblemContext":
        return self

    def __exit__(self, exc_type: type, exc_val: Exception, exc_tb: TracebackType) -> None:
        pass

    def record_entropy_usage(self, amount: float) -> None:
        """
        Record entropy usage for tracking problem complexity.
        """
        self.used_entropy += amount

    def allocate_entropy(self, entropy: float | None) -> float:
        """
        Resolve and consume an entropy amount based on the given value or use
        entire entropy budget if None is provided. The provided value allows to
        allocate entropy multiple times across a context lifetime.
        The chosen amount is recorded against the context budget.
        """

        remaining = self.entropy - self.used_entropy
        if remaining <= 1e-12:
            raise ValueError(f"Entropy budget exceeded: remaining {remaining:.3f}")

        amount: float | None = None

        if entropy is not None:
            # If entropy is provided, use it directly.
            # This can happen during generator lifetime that require multiple variable allocations.
            amount = entropy

        if amount is None:
            amount = remaining

        if amount > remaining + 1e-12:
            raise ValueError(f"Entropy budget exceeded: request {amount:.3f}, remaining {remaining:.3f}")

        self.record_entropy_usage(amount)
        return amount

    def _prepare_verification_data(self, input_data: dict[str, Any]) -> dict[str, Any]:
        """Prepare verification data by extracting dependencies and input fields."""
        verification = {}
        num_inputs = 0
        num_dependencies = 0

        # Handle dependencies
        dependent_on = input_data.pop("dependent_on", None)
        if dependent_on is not None:
            verification["dependent_on"] = dependent_on
            num_dependencies = len(dependent_on)

        # Extract and JSON-encode all input_* fields
        for key in list(input_data.keys()):
            if key.startswith("input_"):
                verification[key] = json.dumps(input_data.pop(key))
                num_inputs += 1

        assert num_inputs == num_dependencies, "Number of inputs and dependencies must match"

        # Add generator type and remaining input data
        verification["generator_type"] = input_data.pop("generator_type")
        verification["input"] = json.dumps(input_data)

        return verification

    def record_tool_call(
        self,
        function_name: str,
        result: LibTypes,
        input_data: dict[str, Any],
        is_final: bool = False,
    ) -> str:
        """
        Record a tool call with its result. It tracks the dependencies between
        steps which will later be used to verify correctness during GRPO.
        """
        self.tool_calls_count += 1
        step_id = str(self._step_counter)

        if result is not None:
            result_json = json.dumps(result)
            step_data = {
                "tool": function_name,
                "result": result_json,
                "step_id": step_id,
                "verification": self._prepare_verification_data(input_data),
            }

            if is_final:
                self.golden_result = {"final_answer": result_json, "from_step_id": step_id}

            self.stepwise_results.append(step_data)

        self._step_counter += 1
        return step_id


class CompositionContext(ProblemContext):
    """
    Extends the base ProblemContext to support shared state and global variables
    across composed problem components.
    """

    def __init__(
        self,
        entropy: float,
        difficulty_level: DifficultyCategory,
        step_counter: int,
        template_engine: TemplateEngine,
        local_index: int,
    ):
        super().__init__(entropy, difficulty_level, step_counter)
        self.component_results: list[ComponentResult] = []
        self.template_engine = template_engine
        self.local_index = local_index

    def record_component_result(self, result: ComponentResult) -> None:
        """Record the result of a component execution."""
        self.component_results.append(result)

        # Update entropy usage and validate budget
        self.used_entropy += result.entropy_consumed
        if self.used_entropy > self.entropy + 1e-12:
            raise ValueError(f"Entropy budget exceeded: used {self.used_entropy:.3f}, available {self.entropy:.3f}")

        self.tool_calls_count += result.tool_calls_used