File size: 10,152 Bytes
e0eb79a 5c30113 e0eb79a | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 | """Profile one RL fine-tuning ablation iteration by component.
Runs ``run_ablation`` unmodified and attributes wall-clock time to the
pieces that matter for the speed pass: environment construction, env
stepping, GPU sampling, window extraction, the host-to-device transfer,
the gradient step, each diagnostic, and evaluation.
Everything is measured by monkey-patching the callables ``training.py``
already uses, so the training loop itself is never edited and the
measured run computes exactly what a real run computes.
Usage:
python scripts/profile_ablation.py --checkpoint PATH \
--ablation baseline_rl --max-iter 51
Without ``--config`` this profiles ``configs/defaults.yaml`` merged with
``experiments/rl_finetuning/configs/ablations_final_minihack_gpu_24gb.yaml``.
"""
from __future__ import annotations
import argparse
import json
import logging
import os
import sys
import time
from collections import defaultdict
from pathlib import Path
import torch
_PROJECT_ROOT = Path(__file__).resolve().parents[1]
if str(_PROJECT_ROOT) not in sys.path:
sys.path.insert(0, str(_PROJECT_ROOT))
from experiments.rl_finetuning.ablations import training as T # noqa: E402
from experiments.rl_finetuning.ablations.registry import REGISTRY # noqa: E402
from experiments.rl_finetuning.run_ablations import ( # noqa: E402
_load_yaml,
_merge_to_namespace,
)
from src.envs import minihack_env # noqa: E402
from src.models.denoiser import ModelEMA # noqa: E402
from src.planners.inference import Evaluator # noqa: E402
logger = logging.getLogger("profile_ablation")
class Acc:
"""Accumulating wall-clock timer with a call counter."""
def __init__(self) -> None:
self.total = 0.0
self.calls = 0
def add(self, dt: float) -> None:
"""Record one timed call of duration *dt* seconds."""
self.total += dt
self.calls += 1
def reset(self) -> None:
"""Zero the accumulator."""
self.total = 0.0
self.calls = 0
ACC: dict[str, Acc] = defaultdict(Acc)
WINDOWS: list[int] = []
def _sync() -> None:
if torch.cuda.is_available():
torch.cuda.synchronize()
def _wrap(fn, key: str, sync: bool = False):
"""Return *fn* wrapped so its wall time accumulates under *key*."""
def inner(*a, **kw):
if sync:
_sync()
t0 = time.perf_counter()
try:
return fn(*a, **kw)
finally:
if sync:
_sync()
ACC[key].add(time.perf_counter() - t0)
return inner
def install_patches() -> None:
"""Monkey-patch the callables whose cost we want attributed."""
# -- Collection internals -------------------------------------------
# ``collect/env_acquire`` is the whole cost of getting a usable env,
# whether that is a fresh construction or a pool hit;
# ``collect/env_init`` is the construction alone, so the two diverge
# once the pool starts serving.
if hasattr(T, "acquire_env"):
T.acquire_env = _wrap(T.acquire_env, "collect/env_acquire")
else:
T.make_env = _wrap(T.make_env, "collect/env_acquire")
minihack_env.AdvancedObservationEnv.step = _wrap(
minihack_env.AdvancedObservationEnv.step, "collect/env_step"
)
minihack_env.AdvancedObservationEnv.reset = _wrap(
minihack_env.AdvancedObservationEnv.reset, "collect/env_reset"
)
minihack_env.AdvancedObservationEnv.close = _wrap(
minihack_env.AdvancedObservationEnv.close, "collect/env_close"
)
minihack_env.AdvancedObservationEnv.__init__ = _wrap(
minihack_env.AdvancedObservationEnv.__init__, "collect/env_init"
)
T.remdm_sample = _wrap(T.remdm_sample, "collect/gpu_sample", sync=True)
T._extract_windows = _wrap(T._extract_windows, "collect/extract_windows")
# -- Window count actually available to the gradient step ------------
# ``compute_advantages`` is called once per iteration on every window
# collected after the ablation's filters, so its input length is the
# number the ``local_obs[:batch_size]`` slice draws from.
_adv = T.compute_advantages
def _adv_counting(returns, *a, **kw):
WINDOWS.append(int(returns.shape[0]))
return _adv(returns, *a, **kw)
T.compute_advantages = _adv_counting
# -- Model plumbing --------------------------------------------------
ModelEMA.make_eval_model = _wrap(
ModelEMA.make_eval_model, "train/make_eval_model", sync=True
)
# -- Diagnostics -----------------------------------------------------
T.compute_grad_alignment = _wrap(
T.compute_grad_alignment, "diag/grad_alignment", sync=True
)
T.compute_per_layer_grad_norms = _wrap(
T.compute_per_layer_grad_norms, "diag/per_layer_norms_only", sync=True
)
T.compute_repr_drift = _wrap(T.compute_repr_drift, "diag/repr_drift", sync=True)
T.compute_cka = _wrap(T.compute_cka, "diag/cka", sync=True)
T.compute_t_analysis = _wrap(T.compute_t_analysis, "diag/t_analysis", sync=True)
Evaluator.evaluate = _wrap(Evaluator.evaluate, "eval/evaluate", sync=True)
ITER_ROWS: list[dict] = []
def install_metric_capture() -> None:
"""Capture each iteration's metric dict plus the timer deltas.
``training.py`` stops its own ``speed/iter_time_sec`` clock straight
after the gradient step, before the diagnostics and the eval, so it
understates a diagnostic iteration badly. ``prof/wall_iter_sec`` is
the gap between consecutive log calls, which is the real thing.
"""
prev_total: dict[str, float] = {}
prev_calls: dict[str, int] = {}
last = [time.perf_counter()]
seen_windows = [0]
def _log(metrics: dict, step: int) -> None:
now = time.perf_counter()
row = dict(metrics)
row["prof/wall_iter_sec"] = now - last[0]
last[0] = now
# First new advantage call of the iteration is the online batch;
# mixed_replay adds a second call for the buffer sample.
if len(WINDOWS) > seen_windows[0]:
row["prof/windows_collected"] = WINDOWS[seen_windows[0]]
seen_windows[0] = len(WINDOWS)
for key, acc in ACC.items():
row[f"prof/{key}"] = acc.total - prev_total.get(key, 0.0)
row[f"prof/{key}_calls"] = acc.calls - prev_calls.get(key, 0)
prev_total[key] = acc.total
prev_calls[key] = acc.calls
ITER_ROWS.append(row)
T._wandb_log = _log
def main() -> None:
"""Entry point."""
p = argparse.ArgumentParser(description=__doc__)
p.add_argument("--config", default=str(_PROJECT_ROOT / "configs/defaults.yaml"))
p.add_argument(
"--ablations-config",
default=str(
_PROJECT_ROOT
/ "experiments/rl_finetuning/configs/ablations_final_minihack_gpu_24gb.yaml"
),
)
p.add_argument("--checkpoint", required=True)
p.add_argument("--ablation", default="baseline_rl")
p.add_argument("--max-iter", type=int, default=51)
p.add_argument("--seed", type=int, default=0)
p.add_argument("--device", default="cuda")
p.add_argument("--out", default=None, help="Write per-iteration JSON here.")
p.add_argument(
"--override",
action="append",
default=[],
help="Config override key=value (parsed as JSON, then as str).",
)
args = p.parse_args()
logging.basicConfig(
level=logging.INFO,
format="%(asctime)s %(levelname)s %(message)s",
)
overrides: dict = {"max_iter": args.max_iter}
for item in args.override:
key, _, raw = item.partition("=")
try:
overrides[key] = json.loads(raw)
except json.JSONDecodeError:
overrides[key] = raw
cfg = _merge_to_namespace(
_load_yaml(args.config),
_load_yaml(args.ablations_config),
overrides,
)
device = torch.device(args.device)
if device.type == "cuda":
torch.set_float32_matmul_precision("high")
logger.info("TMPDIR=%s", os.environ.get("TMPDIR", "<unset>"))
logger.info("config=%s", args.config)
logger.info("ablations_config=%s", args.ablations_config)
logger.info("checkpoint=%s", args.checkpoint)
logger.info(
"n_embd=%s n_head=%s n_layer=%s batch_size=%s episodes_per_iter=%s",
cfg.n_embd,
cfg.n_head,
cfg.n_layer,
cfg.batch_size,
cfg.episodes_per_iter,
)
install_patches()
install_metric_capture()
spec = REGISTRY[args.ablation]
t0 = time.perf_counter()
_, final_score, _, _ = T.run_ablation(
spec, cfg, args.checkpoint, device, seed=args.seed
)
wall = time.perf_counter() - t0
print("\n" + "=" * 78)
print(f"ablation={args.ablation} iters={args.max_iter} wall={wall:.1f}s")
print(f"final_score={final_score:.4f}")
print("=" * 78)
print(f"{'component':38s} {'total_s':>10s} {'calls':>9s} {'ms/call':>10s}")
for key in sorted(ACC):
acc = ACC[key]
per = 1000.0 * acc.total / acc.calls if acc.calls else 0.0
print(f"{key:38s} {acc.total:10.2f} {acc.calls:9d} {per:10.3f}")
if args.out:
Path(args.out).write_text(
json.dumps(
{
"ablation": args.ablation,
"max_iter": args.max_iter,
"seed": args.seed,
"wall_sec": wall,
"final_score": final_score,
"config": {
"n_embd": cfg.n_embd,
"n_head": cfg.n_head,
"n_layer": cfg.n_layer,
"batch_size": cfg.batch_size,
"episodes_per_iter": cfg.episodes_per_iter,
},
"totals": {
k: {"total_sec": v.total, "calls": v.calls}
for k, v in ACC.items()
},
"iters": ITER_ROWS,
},
indent=1,
)
)
print(f"\nwrote {args.out}")
if __name__ == "__main__":
main()
|