File size: 9,512 Bytes
bc35a94
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from __future__ import annotations

import statistics
from dataclasses import dataclass
from dataclasses import field
from typing import Any
from typing import Callable
from typing import Sequence

from hpc_gym import EnterpriseHPCEnv
from training.agent_prompt import SYSTEM_PROMPT
from training.agent_prompt import USER_PROMPT
from training.agent_prompt import iter_actions
from training.agent_prompt import parse_action


GenerateFn = Callable[[list[list[dict[str, str]]]], list[str]]


@dataclass
class RolloutRecord:
    # `reward` is the *cumulative* shaped reward collected during the rollout.
    # useful as a dense progress signal when `terminated` is False. the final
    # step's reward (on solve) is also reflected here because the server's
    # evaluate_action returns a positive health_delta on the solving step.
    reward: float
    steps: int
    terminated: bool
    truncated: bool
    task_id: str
    transcript: list[dict[str, str]] = field(default_factory=list)
    # latest grader_health observed during the rollout. server-side graders
    # populate this 0..1 progress signal; clients without server support see 0.
    grader_health: float = 0.0
    # peak health during the rollout. robust to transient regressions.
    best_health: float = 0.0
    ood_http_code: str = ""
    # the reward returned by the final env step. separate from the cumulative
    # `reward` above so reward functions can distinguish "finished on a solve
    # step" from "accumulated shaped progress".
    last_reward: float = 0.0


def score_single_shot(completions: Sequence[str], env: EnterpriseHPCEnv) -> list[RolloutRecord]:
    records: list[RolloutRecord] = []
    for completion in completions:
        env.reset()
        cumulative_reward = 0.0
        last_reward = 0.0
        health = 0.0
        best_health = 0.0
        http_code = ""
        steps = 0
        terminated = False
        truncated = False
        task_id = env.scenario.TASK_ID
        for action in iter_actions(completion):
            _, last_reward, terminated, truncated, info = env.step(action)
            cumulative_reward += float(last_reward)
            steps += 1
            health = float(info.get("grader_health", 0.0))
            best_health = max(best_health, health)
            http_code = str(info.get("ood_http_code", ""))
            if terminated or truncated:
                break
        records.append(
            RolloutRecord(
                reward=cumulative_reward,
                steps=steps,
                terminated=terminated,
                truncated=truncated,
                task_id=task_id,
                grader_health=health,
                best_health=best_health,
                ood_http_code=http_code,
                last_reward=float(last_reward),
            )
        )
    return records


def run_interactive_group(
    group_size: int,
    generate_fn: GenerateFn,
    env_factory: Callable[[], EnterpriseHPCEnv],
    max_turns: int,
    seed_start: int = 0,
    initial_completions: list[str] | None = None,
) -> list[RolloutRecord]:
    envs: list[EnterpriseHPCEnv] = []
    transcripts: list[list[dict[str, str]]] = []
    observations: list[str] = []
    done: list[bool] = []
    cumulative_rewards: list[float] = [0.0] * group_size
    last_rewards: list[float] = [0.0] * group_size
    health: list[float] = [0.0] * group_size
    best_health: list[float] = [0.0] * group_size
    http_codes: list[str] = [""] * group_size
    steps_taken: list[int] = [0] * group_size
    terminated_list: list[bool] = [False] * group_size
    truncated_list: list[bool] = [False] * group_size
    task_ids: list[str] = [""] * group_size

    for idx in range(group_size):
        env = env_factory()
        obs, info = env.reset(seed=seed_start + idx)
        envs.append(env)
        task_ids[idx] = info.get("task_id") or getattr(getattr(env, "scenario", None), "TASK_ID", "") or ""
        initial_health = float(info.get("grader_health", 0.0))
        health[idx] = initial_health
        best_health[idx] = initial_health
        transcripts.append(
            [
                {"role": "system", "content": SYSTEM_PROMPT},
                {"role": "user", "content": f"{USER_PROMPT}\n\ncurrent observation:\n{obs}"},
            ]
        )
        observations.append(obs)
        done.append(False)

    try:
        for turn in range(max_turns):
            active = [i for i in range(group_size) if not done[i]]
            if not active:
                break
            if turn == 0 and initial_completions is not None:
                completions = [initial_completions[i] for i in active]
            else:
                batch = [transcripts[i] for i in active]
                completions = generate_fn(batch)
            if len(completions) != len(active):
                raise RuntimeError(
                    f"generate_fn returned {len(completions)} completions expected {len(active)}"
                )
            for j, idx in enumerate(active):
                completion = completions[j]
                transcripts[idx].append({"role": "assistant", "content": completion})

                command, voluntary_done = parse_action(completion)
                if command is None:
                    transcripts[idx].append(
                        {
                            "role": "user",
                            "content": "error no bash block detected. emit exactly one <bash>...</bash> block",
                        }
                    )
                    steps_taken[idx] += 1
                    if steps_taken[idx] >= max_turns:
                        truncated_list[idx] = True
                        done[idx] = True
                    continue

                obs, reward, terminated, truncated, info = envs[idx].step(command)
                steps_taken[idx] += 1
                last_rewards[idx] = float(reward)
                cumulative_rewards[idx] += float(reward)
                step_health = float(info.get("grader_health", 0.0))
                health[idx] = step_health
                best_health[idx] = max(best_health[idx], step_health)
                http_codes[idx] = str(info.get("ood_http_code", ""))
                terminated_list[idx] = bool(terminated)
                truncated_list[idx] = bool(truncated)

                if voluntary_done or terminated or truncated:
                    done[idx] = True
                else:
                    transcripts[idx].append(
                        {
                            "role": "user",
                            "content": f"step {steps_taken[idx]} observation:\n{obs}",
                        }
                    )
    finally:
        for env in envs:
            try:
                env.close()
            except Exception:
                pass

    return [
        RolloutRecord(
            reward=cumulative_rewards[i],
            steps=steps_taken[i],
            terminated=terminated_list[i],
            truncated=truncated_list[i],
            task_id=task_ids[i],
            transcript=transcripts[i],
            grader_health=health[i],
            best_health=best_health[i],
            ood_http_code=http_codes[i],
            last_reward=last_rewards[i],
        )
        for i in range(group_size)
    ]


def summarize_group(records: Sequence[RolloutRecord]) -> dict[str, float]:
    if not records:
        return {}
    rewards = [r.reward for r in records]
    steps = [r.steps for r in records]
    solved = sum(1 for r in records if r.terminated)
    return {
        "n": float(len(records)),
        "reward_mean": statistics.fmean(rewards),
        "reward_max": max(rewards),
        "solve_rate": solved / len(records),
        "steps_mean": statistics.fmean(steps),
        "health_mean": statistics.fmean(r.grader_health for r in records),
        "best_health_mean": statistics.fmean(r.best_health for r in records),
    }


def run_fixed_policy(
    env: EnterpriseHPCEnv,
    actions: Sequence[str],
    reset_options: dict[str, Any] | None = None,
) -> RolloutRecord:
    obs, info = env.reset(options=reset_options)
    cumulative_reward = 0.0
    last_reward = 0.0
    steps = 0
    terminated = False
    truncated = False
    task_id = info.get("task_id") or getattr(getattr(env, "scenario", None), "TASK_ID", "") or ""
    health = float(info.get("grader_health", 0.0))
    best_health = health
    http_code = ""
    transcript = [
        {"role": "system", "content": SYSTEM_PROMPT},
        {"role": "user", "content": f"{USER_PROMPT}\n\ncurrent observation:\n{obs}"},
    ]
    for command in actions:
        transcript.append({"role": "assistant", "content": f"<bash>{command}</bash>"})
        obs, last_reward, terminated, truncated, info = env.step(command)
        cumulative_reward += float(last_reward)
        steps += 1
        health = float(info.get("grader_health", 0.0))
        best_health = max(best_health, health)
        http_code = str(info.get("ood_http_code", ""))
        transcript.append({"role": "user", "content": f"step {steps} observation:\n{obs}"})
        if terminated or truncated:
            break
    return RolloutRecord(
        reward=cumulative_reward,
        steps=steps,
        terminated=terminated,
        truncated=truncated,
        task_id=task_id,
        transcript=transcript,
        grader_health=health,
        best_health=best_health,
        ood_http_code=http_code,
        last_reward=float(last_reward),
    )