Spaces:
Sleeping
Sleeping
File size: 7,725 Bytes
7952f32 | 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 | """Multi-turn rollout — the bridge between the env and a policy.
For each turn:
1. The policy is sampled, given the conversation so far. It returns a
single text completion.
2. The completion is parsed to extract the tool call. If parsing fails,
a synthetic ``schema_rejection`` step is recorded with the reward
engine's MALFORMED magnitude and the loop continues.
3. The tool call is forwarded to the env via ``EnvClient.step``. The env
returns ``{observation, reward, done, info}``.
4. The observation is appended to the conversation as a user turn.
5. We stop on ``done`` or when ``episode_cap`` is reached.
After the loop we compute discounted returns from each turn and produce a
list of ``TurnSample(prompt_messages, completion_text, reward, return_)``
tuples — exactly the shape ``trl.GRPOTrainer`` consumes when wrapped with
a custom reward function.
The rollout is environment-agnostic via :class:`EnvClient` and
policy-agnostic via :class:`Policy`. Both come from sibling modules; the
rollout function never imports torch or httpx directly.
"""
from __future__ import annotations
from dataclasses import dataclass, field
from typing import Any
from graphforge.reward.engine import (
DUPLICATE_ACTION,
PER_TURN_COST,
SCHEMA_REJECTION,
)
from graphforge.training.client import EnvClient
from graphforge.training.policy import Policy
from graphforge.training.prompt import (
Message,
append_completion,
append_observation,
initial_messages,
)
from graphforge.training.protocol import (
ParseFailure,
ParseSuccess,
parse_completion,
)
# ---- per-turn record -------------------------------------------------
@dataclass
class TurnSample:
"""Single (prompt, completion, reward, return) tuple for the trainer.
``prompt_messages`` is the conversation up to (but not including) the
assistant's completion at this turn.
"""
turn: int
prompt_messages: list[Message]
completion_text: str
reward: float
return_: float = 0.0
# Diagnostics; not consumed by the trainer.
parse_ok: bool = True
parse_failure_code: str | None = None
env_response: dict[str, Any] = field(default_factory=dict)
done: bool = False
@dataclass
class Trajectory:
episode_id: str
task_id: str
samples: list[TurnSample] = field(default_factory=list)
terminated_naturally: bool = False
terminal_total: float | None = None
@property
def total_reward(self) -> float:
return sum(s.reward for s in self.samples)
def __len__(self) -> int:
return len(self.samples)
# ---- rollout ---------------------------------------------------------
def rollout(
*,
policy: Policy,
env: EnvClient,
task_id: str | None = None,
seed: int | None = None,
gamma: float = 0.97,
max_turns: int | None = None,
auto_close: bool = True,
) -> Trajectory:
"""Run one episode end-to-end. Returns a :class:`Trajectory`.
``max_turns`` overrides the task's ``episode_cap`` if specified
(useful for unit tests). Otherwise the env's own cap fires first.
``auto_close`` calls ``env.close`` when the episode ends.
"""
reset_resp = env.reset(task_id=task_id, seed=seed)
episode_id = reset_resp["episode_id"]
task_visible = reset_resp["observation"]["task"]
cap = max_turns or task_visible["episode_cap"]
messages = initial_messages(task_visible)
samples: list[TurnSample] = []
done = False
terminal_total: float | None = None
for turn_idx in range(cap):
# 1. Sample the policy.
completion = policy.sample(messages)
prompt_at_turn = list(messages) # snapshot before appending the assistant turn
# 2. Parse the tool call.
parsed = parse_completion(completion)
if isinstance(parsed, ParseFailure):
# Synthetic step — env never sees the action. Reward mirrors
# the MALFORMED branch of score_turn (no token cost because
# nothing came back from the env).
reward = SCHEMA_REJECTION + PER_TURN_COST
sample = TurnSample(
turn=turn_idx,
prompt_messages=prompt_at_turn,
completion_text=completion,
reward=reward,
parse_ok=False,
parse_failure_code=parsed.code,
)
samples.append(sample)
messages = append_completion(messages, completion)
messages = append_observation(
messages,
{
"ok": False,
"outcome": "malformed",
"is_duplicate": False,
"reward": reward,
"payload": {"error": parsed.code, "message": parsed.message},
"turns_total": turn_idx + 1,
"tokens_used_total": 0,
"budget_remaining": task_visible["budget"],
"episode_cap_remaining": cap - (turn_idx + 1),
},
)
continue
# 3. Forward to env.
assert isinstance(parsed, ParseSuccess)
env_resp = env.step(episode_id, parsed.action)
info = env_resp.get("info", {})
# The env client returns a synthetic response on FastAPI 422 — that's
# a schema_rejection (e.g. unknown kind, missing required field).
# Score it the same as a parse-side malformed completion.
is_schema_rejection = info.get("error") == "schema_rejection"
if is_schema_rejection:
reward = SCHEMA_REJECTION + PER_TURN_COST
else:
reward = float(env_resp.get("reward", 0.0))
done = bool(env_resp.get("done", False))
# The embedded observation carries duplicate flags etc.
obs = env_resp.get("observation", {})
sample = TurnSample(
turn=turn_idx,
prompt_messages=prompt_at_turn,
completion_text=completion,
reward=reward,
env_response=env_resp,
done=done,
parse_ok=not is_schema_rejection,
parse_failure_code="env_schema_rejection" if is_schema_rejection else None,
)
samples.append(sample)
messages = append_completion(messages, completion)
messages = append_observation(messages, obs)
if done:
terminal_total = info.get("terminal", {}).get("total")
break
if auto_close:
try:
env.close(episode_id)
except Exception:
pass
_fill_returns(samples, gamma=gamma)
return Trajectory(
episode_id=episode_id,
task_id=task_visible.get("id", ""),
samples=samples,
terminated_naturally=done,
terminal_total=terminal_total,
)
# ---- discounted returns ---------------------------------------------
def _fill_returns(samples: list[TurnSample], *, gamma: float) -> None:
"""In-place fill of ``return_`` on each sample.
return_t = r_t + gamma * return_{t+1}, with return_{T+1} = 0.
"""
running = 0.0
for s in reversed(samples):
running = s.reward + gamma * running
s.return_ = running
# ---- helper for stub-policy demo ------------------------------------
def trajectory_summary(traj: Trajectory) -> dict[str, Any]:
return {
"episode_id": traj.episode_id,
"task_id": traj.task_id,
"n_turns": len(traj),
"total_reward": traj.total_reward,
"terminated_naturally": traj.terminated_naturally,
"terminal_total": traj.terminal_total,
"parse_failures": sum(1 for s in traj.samples if not s.parse_ok),
}
|