File size: 9,089 Bytes
8097081
 
 
 
 
 
 
 
 
 
 
 
1435892
8097081
 
 
 
1435892
 
 
8097081
 
 
 
 
 
 
 
 
1435892
 
 
 
8097081
 
 
 
 
 
1435892
8097081
 
 
1435892
 
 
8097081
 
 
 
 
 
1435892
 
 
 
8097081
 
 
 
 
 
1435892
8097081
 
 
 
 
 
 
 
 
 
 
1435892
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8097081
 
 
 
 
 
 
 
 
 
 
1435892
8097081
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1435892
8097081
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1435892
8097081
 
 
 
 
 
 
 
 
 
 
 
1435892
8097081
 
 
 
 
 
 
 
 
1435892
 
 
 
 
 
 
 
8097081
1435892
 
 
8097081
 
 
 
 
 
 
 
 
1435892
8097081
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# src/pytorch_debug_env/environment.py
from __future__ import annotations

from dataclasses import dataclass, field
from typing import Dict, List

from .models import (
    HypothesisRecord,
    PyTorchDebugAction,
    PyTorchDebugObservation,
    PyTorchDebugState,
)
from .reward import clamp_score, compute_step_reward
from .scenario_generator import ScenarioGenerator
from .graders import grade_easy, grade_medium, grade_hard

GRADER_MAP = {"easy": grade_easy, "medium": grade_medium, "hard": grade_hard}
LOSS_WINDOW_STEP = 25
GPU_WINDOW_STEP = 25
LOG_WINDOW_STEP = 10


@dataclass
class RuntimeState:
    scenario: object | None = None
    max_steps: int = 5
    current_step: int = 0
    revealed_files: List[str] = field(default_factory=list)
    hypothesis_history: List[HypothesisRecord] = field(default_factory=list)
    loss_curve_bonus: int = 0
    gpu_profile_bonus: int = 0
    log_tail_bonus: int = 0
    diagnostic_revealed: bool = False
    done: bool = False
    final_score: float = 0.0


class PyTorchDebugEnv:
    def __init__(self, generator: ScenarioGenerator, max_steps: int = 5):
        """Create a PyTorch debugging environment with a scenario generator."""
        self.generator = generator
        self.runtime = RuntimeState(max_steps=max_steps)

    async def reset(self, task_id: str = "easy", seed: int | None = None):
        """Start a new episode and return the initial observation."""
        scenario = self.generator.generate(task_id, seed=seed)
        self.runtime = RuntimeState(
            scenario=scenario,
            max_steps=5 if task_id == "easy" else 6,
            current_step=0,
            revealed_files=["train.py", "config/training_config.yaml"],
            hypothesis_history=[],
            loss_curve_bonus=0,
            gpu_profile_bonus=0,
            log_tail_bonus=0,
            diagnostic_revealed=False,
            done=False,
            final_score=0.0,
        )
        return self._build_observation(last_feedback="Episode reset.")

    async def step(self, action: PyTorchDebugAction):
        """Advance the environment by one step using the provided action."""
        if self.runtime.scenario is None:
            raise RuntimeError("Call /reset before /step")

        if self.runtime.done:
            raise RuntimeError("Episode already completed")

        self.runtime.current_step += 1
        scenario = self.runtime.scenario
        previous_quality = self.runtime.hypothesis_history[-1].quality if self.runtime.hypothesis_history else 0.0

        investigation_target = None
        if action.investigation_action:
            action_type = action.investigation_action.action
            if action_type == "reveal_file":
                investigation_target = action.investigation_action.target
                if (
                    investigation_target in scenario.repo_files
                    and investigation_target not in self.runtime.revealed_files
                ):
                    self.runtime.revealed_files.append(investigation_target)
            elif action_type == "extend_loss_curve":
                self.runtime.loss_curve_bonus += 1
            elif action_type == "extend_gpu_profile":
                self.runtime.gpu_profile_bonus += 1
            elif action_type == "reveal_log_chunk":
                self.runtime.log_tail_bonus += 1
            elif action_type == "run_diagnostic":
                self.runtime.diagnostic_revealed = True

        committed = action.final_diagnosis.model_dump() if action.commit_diagnosis and action.final_diagnosis else None
        reward, components = compute_step_reward(
            previous_quality=previous_quality,
            current_hypothesis=action.current_hypothesis.model_dump(),
            ground_truth=scenario.ground_truth,
            investigation_target=investigation_target,
            committed_diagnosis=None, # Temporarily don't compute diagnosis reward here to use grader
            step_num=self.runtime.current_step,
            max_steps=self.runtime.max_steps,
        )
        reward = clamp_score(reward)

        if committed:
            grader = GRADER_MAP.get(scenario.task_id, grade_easy)
            diagnosis_reward = grader(committed, scenario.ground_truth)

            # Combine the diagnosis reward logic from `compute_step_reward` that applies on top
            if diagnosis_reward > 0.7:
                diagnosis_reward += max(0.0, 0.08 * (self.runtime.max_steps - self.runtime.current_step))

            # Update the total reward incorporating diagnosis
            components["diagnosis_reward"] = round(diagnosis_reward, 4)
            delta = components["hypothesis_delta"]
            inv_reward = components["investigation_reward"]
            conf_bonus = components["confirmation_bonus"]

            total = 0.60 * delta + 0.20 * inv_reward + 0.20 * diagnosis_reward + conf_bonus
            reward = round(clamp_score(min(max(total, 0.0), 1.0)), 4)

        self.runtime.hypothesis_history.append(
            HypothesisRecord(
                step=self.runtime.current_step,
                hypothesis=action.current_hypothesis,
                quality=components["hypothesis_quality"],
            )
        )

        if action.commit_diagnosis or self.runtime.current_step >= self.runtime.max_steps:
            self.runtime.done = True
            self.runtime.final_score = reward

        observation = self._build_observation(
            last_feedback=self._feedback(action, scenario.ground_truth)
        )
        return {
            "observation": observation,
            "reward": reward,
            "done": self.runtime.done,
            "info": components,
        }

    async def state(self):
        """Return the current episode state, or None if not started."""
        scenario = self.runtime.scenario
        if not scenario:
            return None
        return PyTorchDebugState(
            scenario_id=scenario.scenario_id,
            task_id=scenario.task_id,
            max_steps=self.runtime.max_steps,
            current_step=self.runtime.current_step,
            revealed_files=self.runtime.revealed_files,
            remaining_files=[
                f for f in scenario.repo_files.keys() if f not in self.runtime.revealed_files
            ],
            diagnostic_revealed=self.runtime.diagnostic_revealed,
            done=self.runtime.done,
            final_score=self.runtime.final_score,
        )

    def _build_observation(self, last_feedback: str) -> PyTorchDebugObservation:
        scenario = self.runtime.scenario
        revealed = {k: v for k, v in scenario.repo_files.items() if k in self.runtime.revealed_files}
        available = [k for k in scenario.repo_files.keys() if k not in self.runtime.revealed_files]

        loss_window_size = min(
            len(scenario.loss_curve),
            LOSS_WINDOW_STEP * (self.runtime.current_step + 1 + self.runtime.loss_curve_bonus),
        )
        gpu_window_size = min(
            len(scenario.gpu_profile),
            GPU_WINDOW_STEP * (self.runtime.current_step + 1 + self.runtime.gpu_profile_bonus),
        )
        log_lines = scenario.training_log.splitlines()
        log_window = LOG_WINDOW_STEP * (self.runtime.current_step + 1 + self.runtime.log_tail_bonus)
        visible_log = "\n".join(log_lines[-min(len(log_lines), log_window):])
        diagnostic_report = scenario.diagnostic_report if self.runtime.diagnostic_revealed else None

        return PyTorchDebugObservation(
            scenario_id=scenario.scenario_id,
            task_id=scenario.task_id,
            revealed_files=revealed,
            available_files=available,
            loss_curve_window=scenario.loss_curve[:loss_window_size],
            gpu_profile_window=scenario.gpu_profile[:gpu_window_size],
            training_log_tail=visible_log,
            diagnostic_report=diagnostic_report,
            step_num=self.runtime.current_step,
            steps_remaining=max(0, self.runtime.max_steps - self.runtime.current_step),
            investigation_budget=max(0, self.runtime.max_steps - self.runtime.current_step),
            hypothesis_history=self.runtime.hypothesis_history,
            last_feedback=last_feedback,
        )

    def _feedback(self, action: PyTorchDebugAction, gt: Dict) -> str:
        suspected_file = action.current_hypothesis.affected_file
        suspected_bug = action.current_hypothesis.bug_type

        if suspected_file == gt.get("red_herring_file"):
            return "That file contains a plausible symptom, but not the root cause. Investigate upstream causes."
        if suspected_file == gt["primary_bug_file"] and suspected_bug != gt["bug_type"]:
            return "Correct region, wrong failure mode. Re-check the training artifacts more carefully."
        if suspected_bug == gt["bug_type"] and suspected_file != gt["primary_bug_file"]:
            return "The bug class looks right, but the faulty implementation is in another file."
        return "Continue refining the hypothesis using newly revealed evidence."