File size: 9,173 Bytes
b74674a | 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 | # Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
"""
Recursive backend abstractions for repl_env.
This module keeps direct LM calls and recursive child spawning out of the
runner and environment. The runner owns the iterative loop; the backend owns
query/query_batched/child-recursion behavior.
"""
from __future__ import annotations
import threading
import time
from concurrent.futures import as_completed, ThreadPoolExecutor
from dataclasses import dataclass, field
from typing import Callable, Protocol
ChatFn = Callable[..., str]
class RecursiveBackend(Protocol):
max_depth: int
depth: int
child_traces: list["ChildTrace"]
def query(self, prompt: str, model: str | None = None) -> str: ...
def query_batched(
self, prompts: list[str], model: str | None = None
) -> list[str]: ...
def recursive_query(self, prompt: str, model: str | None = None) -> str: ...
def recursive_query_batched(
self, prompts: list[str], model: str | None = None
) -> list[str]: ...
@dataclass
class BackendLimits:
max_depth: int = 1
max_batch_workers: int = 8
max_children_total: int | None = None
max_children_per_batch: int | None = None
result_truncation_limit: int | None = None
# Cooperative timeout: checked between iterations, not during LLM calls.
# A slow LLM call within an iteration will not be interrupted — the timeout
# fires at the next iteration boundary. For mid-call cancellation, use
# process-based isolation instead.
per_child_timeout_s: float | None = None
# Tree-global child counter shared across all recursion depths
_children_spawned: int = field(default=0, init=False, repr=False)
_children_lock: threading.Lock = field(
default_factory=threading.Lock, init=False, repr=False
)
@dataclass
class ChildTrace:
depth: int
duration_s: float
prompt_preview: str
result_preview: str | None
error: str | None
class DirectLMBackend:
"""Direct LM backend with no child recursion beyond fallback to itself."""
def __init__(
self,
llm_chat_fn: ChatFn,
*,
depth: int = 0,
limits: BackendLimits | None = None,
) -> None:
self.llm_chat_fn = llm_chat_fn
self.depth = depth
self.limits = limits or BackendLimits()
self.max_depth = self.limits.max_depth
self.child_traces: list[ChildTrace] = []
def query(self, prompt: str, model: str | None = None) -> str:
try:
result = self.llm_chat_fn([{"role": "user", "content": prompt}], model)
except TypeError:
result = self.llm_chat_fn([{"role": "user", "content": prompt}])
return self._truncate(result)
def query_batched(self, prompts: list[str], model: str | None = None) -> list[str]:
if not prompts:
return []
max_workers = min(len(prompts), self.limits.max_batch_workers)
results: list[str] = [""] * len(prompts)
with ThreadPoolExecutor(max_workers=max_workers) as executor:
future_to_idx = {
executor.submit(self.query, prompt, model): idx
for idx, prompt in enumerate(prompts)
}
for future in as_completed(future_to_idx):
idx = future_to_idx[future]
try:
results[idx] = future.result()
except Exception as exc:
results[idx] = f"Error: {exc}"
return results
def recursive_query(self, prompt: str, model: str | None = None) -> str:
return self.query(prompt, model)
def recursive_query_batched(
self, prompts: list[str], model: str | None = None
) -> list[str]:
return self.query_batched(prompts, model)
def _truncate(self, result: str) -> str:
limit = self.limits.result_truncation_limit
if limit is not None and len(result) > limit:
return result[:limit]
return result
class LocalChildRLMBackend(DirectLMBackend):
"""Recursive backend that spawns child LocalRLMRunner instances."""
def __init__(
self,
llm_chat_fn: ChatFn,
*,
runner_factory: Callable[..., object],
system_prompt: str,
max_iterations: int,
env_max_iterations_multiplier: int,
depth: int = 0,
limits: BackendLimits | None = None,
on_subcall_start: Callable[[int, str, str], None] | None = None,
on_subcall_complete: Callable[[int, str, float, str | None], None]
| None = None,
) -> None:
super().__init__(llm_chat_fn, depth=depth, limits=limits)
self.runner_factory = runner_factory
self.system_prompt = system_prompt
self.max_iterations = max_iterations
self.env_max_iterations_multiplier = env_max_iterations_multiplier
self.on_subcall_start = on_subcall_start
self.on_subcall_complete = on_subcall_complete
def recursive_query(self, prompt: str, model: str | None = None) -> str:
next_depth = self.depth + 1
if next_depth >= self.max_depth:
return self.query(prompt, model)
with self.limits._children_lock:
if self.limits.max_children_total is not None:
if self.limits._children_spawned >= self.limits.max_children_total:
return "Error: max_children_total exceeded"
self.limits._children_spawned += 1
start = time.perf_counter()
error: str | None = None
result_text = ""
resolved_model = model or "default"
if self.on_subcall_start is not None:
try:
self.on_subcall_start(next_depth, str(resolved_model), prompt[:80])
except Exception:
pass
try:
child = self.runner_factory(
self.llm_chat_fn,
system_prompt=self.system_prompt,
max_iterations=self.max_iterations,
max_depth=self.max_depth,
depth=next_depth,
env_max_iterations_multiplier=self.env_max_iterations_multiplier,
max_batch_workers=self.limits.max_batch_workers,
backend_factory=self._child_backend_factory,
on_subcall_start=self.on_subcall_start,
on_subcall_complete=self.on_subcall_complete,
)
result = child.run(
prompt, prompt, model=model, timeout_s=self.limits.per_child_timeout_s
)
result_text = self._truncate(result.final_answer or "")
return result_text
except Exception as exc:
error = str(exc)
raise
finally:
duration = time.perf_counter() - start
self.child_traces.append(
ChildTrace(
depth=next_depth,
duration_s=duration,
prompt_preview=prompt[:80],
result_preview=(result_text[:80] if result_text else None),
error=error,
)
)
if self.on_subcall_complete is not None:
try:
self.on_subcall_complete(
next_depth,
str(resolved_model),
duration,
error,
)
except Exception:
pass
def recursive_query_batched(
self, prompts: list[str], model: str | None = None
) -> list[str]:
if not prompts:
return []
batch_limit = self.limits.max_children_per_batch
if batch_limit is not None:
prompts = prompts[:batch_limit]
max_workers = min(len(prompts), self.limits.max_batch_workers)
results: list[str] = [""] * len(prompts)
with ThreadPoolExecutor(max_workers=max_workers) as executor:
future_to_idx = {
executor.submit(self.recursive_query, prompt, model): idx
for idx, prompt in enumerate(prompts)
}
for future in as_completed(future_to_idx):
idx = future_to_idx[future]
try:
results[idx] = future.result()
except Exception as exc:
results[idx] = f"Error: {exc}"
return results
def _child_backend_factory(
self, llm_chat_fn: ChatFn, **kwargs
) -> "LocalChildRLMBackend":
return LocalChildRLMBackend(
llm_chat_fn,
runner_factory=self.runner_factory,
system_prompt=kwargs["system_prompt"],
max_iterations=kwargs["max_iterations"],
env_max_iterations_multiplier=kwargs["env_max_iterations_multiplier"],
depth=kwargs["depth"],
limits=self.limits,
on_subcall_start=self.on_subcall_start,
on_subcall_complete=self.on_subcall_complete,
)
|