Spaces:
Sleeping
Sleeping
File size: 7,916 Bytes
6da8b7c 8271223 6da8b7c d140675 6da8b7c 8271223 6da8b7c 8271223 6da8b7c | 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 | """Trajectory cache system for DispatchR.
Builds, persists, and samples from pre-computed greedy episode trajectories.
This eliminates the need to re-run episode simulation during dataset construction.
Cache format (JSONL, one line per episode):
{
"seed": 123,
"difficulty": "learning",
"episode_reward": 0.45,
"steps": [
{"step": 0, "action": {...}},
{"step": 1, "action": {...}},
...
]
}
We do NOT store full observations in the cache to keep file size reasonable.
Observations are cheaply recomputed by replaying cached actions.
"""
import json
import os
from concurrent.futures import ProcessPoolExecutor, as_completed
from pathlib import Path
from typing import Any, Dict, List, Optional, Tuple
import numpy as np
from .greedy_policy import greedy_action
from .grpo_env_wrapper import DispatchRGRPOEnv
def _run_single_episode(args: Tuple[int, str]) -> Dict[str, Any]:
"""Worker: run one greedy episode. Must be top-level for pickling."""
seed, difficulty = args
env = DispatchRGRPOEnv(seed=seed, difficulty=difficulty)
env.reset(seed=seed, difficulty=difficulty)
steps = []
for step_idx in range(80):
obs = env._obs
if obs is None or obs.get("done"):
break
action = greedy_action(obs)
action_text = json.dumps(action)
env.step(action_text)
steps.append({"step": step_idx, "action": action})
reward = env.reward if env.reward is not None else 0.0
return {
"seed": seed,
"difficulty": difficulty,
"episode_reward": float(reward),
"steps": steps,
}
def build_trajectory_cache(
n_seeds: int,
difficulty: str,
output_path: str,
base_seed: int = 0,
num_workers: int = 8,
) -> str:
"""Build trajectory cache by running greedy episodes in parallel.
Args:
n_seeds: Number of episode seeds to simulate.
difficulty: Curriculum phase string.
output_path: Where to write the JSONL cache file.
base_seed: Offset for seed values.
num_workers: Parallel workers (default 8, set to 1 for debugging).
Returns:
Path to the created cache file.
"""
os.makedirs(os.path.dirname(output_path) or ".", exist_ok=True)
tasks = [(base_seed + i, difficulty) for i in range(n_seeds)]
print(f"[Cache] Building {n_seeds} greedy trajectories (difficulty={difficulty})...")
print(f"[Cache] Workers: {num_workers}")
with open(output_path, "w", encoding="utf-8") as f:
if num_workers <= 1:
# Sequential (useful for debugging)
for task in tasks:
episode = _run_single_episode(task)
f.write(json.dumps(episode) + "\n")
f.flush()
else:
with ProcessPoolExecutor(max_workers=num_workers) as executor:
futures = {executor.submit(_run_single_episode, t): t for t in tasks}
for future in as_completed(futures):
episode = future.result()
f.write(json.dumps(episode) + "\n")
f.flush()
print(f"[Cache] Saved to {output_path}")
return output_path
def load_trajectory_cache(cache_path: str) -> List[Dict[str, Any]]:
"""Load full cache into memory.
Returns list of episode dicts. For large caches consider streaming
with ``iter_trajectory_cache`` instead.
"""
episodes = []
with open(cache_path, "r", encoding="utf-8") as f:
for line in f:
if line.strip():
episodes.append(json.loads(line))
return episodes
def iter_trajectory_cache(cache_path: str):
"""Stream trajectory cache line-by-line (memory-efficient)."""
with open(cache_path, "r", encoding="utf-8") as f:
for line in f:
if line.strip():
yield json.loads(line)
def sample_dataset_rows(
cache_path: str,
tokenizer,
system_prompt: str,
n_samples_per_seed: int = 5,
max_step: int = 79,
build_chat_prompt_fn=None,
) -> List[Dict[str, Any]]:
"""Sample random-step observations from cache to build dataset rows.
For each episode in the cache, sample ``n_samples_per_seed`` random steps
(uniformly), replay the episode to that step, capture the observation,
and format it as a prompt.
Args:
cache_path: Path to JSONL cache file.
tokenizer: HF tokenizer for chat template formatting.
system_prompt: The SYSTEM_PROMPT string.
n_samples_per_seed: How many steps to sample per episode.
max_step: Maximum step index to sample (inclusive).
build_chat_prompt_fn: Callable(tokenizer, system, user) -> str.
Returns:
List of dataset row dicts with keys: seed, difficulty, step_idx, prompt.
"""
rows = []
for episode in iter_trajectory_cache(cache_path):
seed = episode["seed"]
difficulty = episode["difficulty"]
cached_steps = episode["steps"]
n_steps = len(cached_steps)
if n_steps <= 1:
continue
# Sample step indices (exclude step 0, cap at actual episode length)
upper = min(max_step, n_steps - 1)
if upper < 1:
continue
sampled_indices = np.random.choice(
range(3, upper + 1),
size=min(n_samples_per_seed, upper),
replace=False,
).tolist()
# Replay episode and capture observations at sampled steps
env = DispatchRGRPOEnv(seed=seed, difficulty=difficulty)
env.reset(seed=seed, difficulty=difficulty)
for step_idx in range(n_steps):
obs = env._obs
if obs is None or obs.get("done"):
break
if step_idx in sampled_indices:
# Format observation as prompt
from .prompt_utils import format_observation, build_chat_prompt
obs_text = format_observation(obs)
if build_chat_prompt_fn is not None:
prompt = build_chat_prompt_fn(tokenizer, system_prompt, obs_text)
else:
prompt = build_chat_prompt(tokenizer, system_prompt, obs_text)
rows.append(
{
"seed": seed,
"difficulty": difficulty,
"step_idx": step_idx,
"prompt": prompt,
}
)
# Apply cached action to advance env (convert dict to JSON string)
action_dict = cached_steps[step_idx]["action"]
env.step(json.dumps(action_dict))
return rows
def replay_steps(env_ref, cached_steps: List[Dict[str, Any]], up_to_step: int) -> None:
"""Replay cached actions on an env up to (but not including) a target step.
Used in reward_fn to quickly reconstruct the state at step ``up_to_step``.
"""
for step_idx in range(min(up_to_step, len(cached_steps))):
action_dict = cached_steps[step_idx]["action"]
env_ref.step(json.dumps(action_dict))
def get_cache_path(cache_dir: str, difficulty: str, n_seeds: int) -> str:
"""Return standard cache file path."""
return os.path.join(
cache_dir, f"trajectories_{difficulty}_{n_seeds}seeds.jsonl"
)
def ensure_cache_exists(
cache_dir: str,
n_seeds: int,
difficulty: str,
base_seed: int = 0,
num_workers: int = 8,
) -> str:
"""Return cache path, building it first if it doesn't exist."""
path = get_cache_path(cache_dir, difficulty, n_seeds)
if os.path.exists(path):
print(f"[Cache] Reusing existing cache: {path}")
return path
return build_trajectory_cache(
n_seeds=n_seeds,
difficulty=difficulty,
output_path=path,
base_seed=base_seed,
num_workers=num_workers,
)
|