File size: 12,595 Bytes
0523608 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 | """
Standalone CORE evaluation script.
Loads one or more nanochat checkpoints directly from their directories and
reports per-task accuracy, centred score, and the aggregate CORE metric.
Unlike base_eval.py this script:
* Accepts raw checkpoint directories (no nanochat model-tag indirection).
* Can evaluate several checkpoints in one shot (pass multiple --checkpoint-dir
values or use shell globs).
* Writes a CSV per checkpoint to --output-dir (default: ./core_results/).
* Prints a compact summary table at the end.
Examples
--------
# Single checkpoint, last step, all tasks, auto device
python -m scripts.eval_core --checkpoint-dir /path/to/checkpoints/d12
# Multiple checkpoints (shell glob)
python -m scripts.eval_core --checkpoint-dir /path/to/checkpoints/d*
# Specific step, fast approximate run, explicit tokeniser
python -m scripts.eval_core \\
--checkpoint-dir /path/to/checkpoints/d24 \\
--step 50000 \\
--max-per-task 100 \\
--tokenizer-dir /path/to/nanochat/tokenizer \\
--output-dir ./core_results
# Multi-GPU via torchrun
torchrun --nproc_per_node=4 -m scripts.eval_core --checkpoint-dir /path/to/checkpoints/d24
"""
import os
import csv
import glob
import json
import time
import yaml
import random
import shutil
import zipfile
import tempfile
import argparse
import urllib.request
from pathlib import Path
import torch
from filelock import FileLock
from nanochat.common import (
compute_init,
compute_cleanup,
autodetect_device_type,
get_base_dir,
download_file_with_lock,
)
from nanochat.checkpoint_manager import build_model, find_last_step
from nanochat.core_eval import evaluate_task
# ---------------------------------------------------------------------------
# Constants
# ---------------------------------------------------------------------------
EVAL_BUNDLE_URL = "https://karpathy-public.s3.us-west-2.amazonaws.com/eval_bundle.zip"
# ---------------------------------------------------------------------------
# Helpers
# ---------------------------------------------------------------------------
def print0(msg: str = "", **kwargs):
"""Print only on rank-0 to avoid duplicate output in distributed runs."""
if int(os.environ.get("RANK", 0)) == 0:
print(msg, **kwargs)
def _ensure_eval_bundle() -> str:
"""Download and unpack the CORE eval bundle if it isn't already present."""
base_dir = get_base_dir()
eval_bundle_dir = os.path.join(base_dir, "eval_bundle")
if os.path.exists(eval_bundle_dir):
return eval_bundle_dir
def _place(zip_path: str):
with tempfile.TemporaryDirectory() as tmpdir:
with zipfile.ZipFile(zip_path, "r") as zf:
zf.extractall(tmpdir)
extracted = os.path.join(tmpdir, "eval_bundle")
shutil.move(extracted, eval_bundle_dir)
print0(f"Eval bundle placed at: {eval_bundle_dir}")
download_file_with_lock(EVAL_BUNDLE_URL, "eval_bundle.zip", postprocess_fn=_place)
return eval_bundle_dir
def _load_core_config(eval_bundle_dir: str):
"""Return (tasks list, random_baselines dict) from the CORE bundle."""
config_path = os.path.join(eval_bundle_dir, "core.yaml")
meta_path = os.path.join(eval_bundle_dir, "eval_meta_data.csv")
with open(config_path, "r", encoding="utf-8") as f:
config = yaml.safe_load(f)
tasks = config["icl_tasks"]
random_baselines: dict[str, float] = {}
with open(meta_path, "r", encoding="utf-8") as f:
reader = csv.DictReader(f)
for row in reader:
random_baselines[row["Eval Task"]] = float(row["Random baseline"])
return tasks, random_baselines
def _run_core(model, tokenizer, device, tasks, random_baselines,
eval_bundle_dir: str, max_per_task: int = -1) -> dict:
"""
Run all CORE tasks and return a results dictionary:
{
"results": {label: accuracy},
"centered_results": {label: centered},
"core_metric": float,
}
"""
data_base = os.path.join(eval_bundle_dir, "eval_data")
results: dict[str, float] = {}
centered_results: dict[str, float] = {}
for task in tasks:
t0 = time.time()
label = task["label"]
task_meta = {
"task_type": task["icl_task_type"],
"dataset_uri": task["dataset_uri"],
"num_fewshot": task["num_fewshot"][0],
"continuation_delimiter": task.get("continuation_delimiter", " "),
}
print0(
f" {label:<35} "
f"({task_meta['num_fewshot']}-shot, {task_meta['task_type']}) ... ",
end="",
flush=True,
)
data_path = os.path.join(data_base, task_meta["dataset_uri"])
with open(data_path, "r", encoding="utf-8") as f:
data = [json.loads(line.strip()) for line in f]
shuffle_rng = random.Random(1337)
shuffle_rng.shuffle(data)
if max_per_task > 0:
data = data[:max_per_task]
accuracy = evaluate_task(model, tokenizer, data, device, task_meta)
random_bl = random_baselines.get(label, 0.0)
centered = (accuracy - 0.01 * random_bl) / (1.0 - 0.01 * random_bl)
results[label] = accuracy
centered_results[label] = centered
elapsed = time.time() - t0
print0(f"acc={accuracy:.4f} centred={centered:.4f} ({elapsed:.1f}s)")
core_metric = sum(centered_results.values()) / len(centered_results)
return {"results": results, "centered_results": centered_results, "core_metric": core_metric}
def _write_csv(out: dict, csv_path: str, checkpoint_dir: str, step: int):
"""Write per-task results to a CSV file."""
os.makedirs(os.path.dirname(csv_path), exist_ok=True)
with open(csv_path, "w", newline="", encoding="utf-8") as f:
writer = csv.writer(f)
writer.writerow(["checkpoint", "step", "task", "accuracy", "centered"])
for label in out["results"]:
writer.writerow([
checkpoint_dir,
step,
label,
f"{out['results'][label]:.6f}",
f"{out['centered_results'][label]:.6f}",
])
writer.writerow([checkpoint_dir, step, "CORE", "", f"{out['core_metric']:.6f}"])
print0(f" → CSV saved to: {csv_path}")
# ---------------------------------------------------------------------------
# Main
# ---------------------------------------------------------------------------
def parse_args():
p = argparse.ArgumentParser(
description="Standalone CORE evaluation for nanochat checkpoints.",
formatter_class=argparse.RawDescriptionHelpFormatter,
epilog=__doc__,
)
p.add_argument(
"--checkpoint-dir",
nargs="+",
required=True,
metavar="DIR",
help=(
"One or more checkpoint directories to evaluate. "
"Each directory must contain model_<step>.pt + meta_<step>.json. "
"Shell globs are supported (e.g. '/runs/d*')."
),
)
p.add_argument(
"--step",
type=int,
default=None,
help="Checkpoint step to load (default: last available step).",
)
p.add_argument(
"--tokenizer-dir",
type=str,
default=None,
metavar="DIR",
help="Path to tokenizer directory (default: nanochat's built-in tokenizer).",
)
p.add_argument(
"--max-per-task",
type=int,
default=-1,
metavar="N",
help="Max examples per task (-1 = all examples, default). Use ~100 for quick runs.",
)
p.add_argument(
"--output-dir",
type=str,
default="./core_results",
metavar="DIR",
help="Directory to write per-checkpoint CSV results (default: ./core_results).",
)
p.add_argument(
"--device-type",
type=str,
default="",
choices=["", "cuda", "cpu", "mps"],
help="Device to use: cuda | cpu | mps (default: auto-detect).",
)
p.add_argument(
"--compile",
action="store_true",
default=False,
help="torch.compile the model before evaluation (much faster on GPU, ~2-3x speedup).",
)
return p.parse_args()
def main():
args = parse_args()
# ---- Device / distributed setup ----------------------------------------
device_type = autodetect_device_type() if args.device_type == "" else args.device_type
ddp, ddp_rank, ddp_local_rank, ddp_world_size, device = compute_init(device_type)
# ---- Expand checkpoint dirs (support globs) ----------------------------
raw_dirs: list[str] = []
for pattern in args.checkpoint_dir:
expanded = sorted(glob.glob(pattern))
if expanded:
raw_dirs.extend(expanded)
else:
# treat as literal path even if glob matched nothing
raw_dirs.append(pattern)
checkpoint_dirs = [d for d in raw_dirs if os.path.isdir(d)]
if not checkpoint_dirs:
print0("ERROR: No valid checkpoint directories found.")
print0(f" Patterns given: {args.checkpoint_dir}")
raise SystemExit(1)
print0(f"\nFound {len(checkpoint_dirs)} checkpoint director(y/ies) to evaluate.")
# ---- CORE bundle --------------------------------------------------------
print0("\nEnsuring CORE eval bundle is present...")
eval_bundle_dir = _ensure_eval_bundle()
tasks, random_baselines = _load_core_config(eval_bundle_dir)
print0(f"Loaded {len(tasks)} CORE tasks.")
# ---- Summary table accumulator -----------------------------------------
summary: list[dict] = [] # [{name, step, core_metric}]
# ---- Evaluate each checkpoint ------------------------------------------
for checkpoint_dir in checkpoint_dirs:
checkpoint_dir = os.path.abspath(checkpoint_dir)
name = os.path.basename(checkpoint_dir)
step = args.step
if step is None:
try:
step = find_last_step(checkpoint_dir)
except FileNotFoundError as e:
print0(f"\nWARNING: Skipping {checkpoint_dir} — {e}")
continue
print0(f"\n{'='*70}")
print0(f"Checkpoint : {checkpoint_dir}")
print0(f"Step : {step:,}")
print0(f"{'='*70}")
# Load model
try:
model, tokenizer, meta_data = build_model(
checkpoint_dir, step, device, phase="eval",
tokenizer_dir=args.tokenizer_dir,
)
except Exception as e:
print0(f"ERROR loading checkpoint: {e}")
continue
# Compile for speed (huge difference on H100 / Ampere+)
if args.compile:
print0(" Compiling model with torch.compile ...")
model = torch.compile(model)
# Warmup pass to trigger compilation before timing starts
print0(" Warmup forward pass ...", end="", flush=True)
warmup_ids = torch.randint(0, 256, (1, 64), device=device)
with torch.no_grad():
_ = model(warmup_ids)
del warmup_ids
if device_type == "cuda":
torch.cuda.synchronize()
print0(" done.")
# Run CORE
out = _run_core(
model, tokenizer, device,
tasks, random_baselines, eval_bundle_dir,
max_per_task=args.max_per_task,
)
core_metric = out["core_metric"]
print0(f"\n CORE metric: {core_metric:.4f}")
# Write CSV (rank-0 only)
if ddp_rank == 0:
csv_name = f"{name}_step{step:06d}_core.csv"
csv_path = os.path.join(args.output_dir, csv_name)
_write_csv(out, csv_path, checkpoint_dir, step)
summary.append({"name": name, "step": step, "core_metric": core_metric})
# Free model memory before next checkpoint
del model
if device_type == "cuda":
torch.cuda.empty_cache()
# ---- Summary table (rank-0 only) ---------------------------------------
if ddp_rank == 0 and summary:
print0(f"\n{'='*70}")
print0("SUMMARY")
print0(f"{'='*70}")
print0(f" {'Checkpoint':<30} {'Step':>8} {'CORE':>8}")
print0(f" {'-'*30} {'-'*8} {'-'*8}")
for row in summary:
print0(f" {row['name']:<30} {row['step']:>8,} {row['core_metric']:>8.4f}")
print0()
compute_cleanup()
if __name__ == "__main__":
main()
|