Spaces:
Sleeping
Sleeping
File size: 13,137 Bytes
5412d82 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 | """
Per-stage latency and peak VRAM benchmark.
Measures wall-clock latency and peak GPU memory across Stage 1, Stage 2,
and Stage 3 on a single representative image. Warmup runs prime CUDA caches
so the recorded runs reflect steady-state inference, not first-call overhead.
The headline target for the abstract: Stage 3 total latency < 700 ms.
Typical usage (library)::
from src.evaluation.latency_benchmark import benchmark
results = benchmark(Path("data/test_images/arkit_41159529_0000.jpg"))
CLI usage::
python -m src.evaluation.latency_benchmark \\
--image data/test_images/arkit_41159529_0000.jpg \\
--output outputs/results/latency.csv
"""
from __future__ import annotations
import argparse
import csv
import time
from pathlib import Path
from typing import Optional
import numpy as np
import torch
from PIL import Image
from ..config import N_LATENCY_RUNS, RESULTS_DIR
from ..pipeline import Pipeline
# ---------------------------------------------------------------------------
# Core benchmark function
# ---------------------------------------------------------------------------
def benchmark(
image_path: Path,
n_warmup: int = 2,
n_runs: int = 10,
stages: list[int] | None = None,
force_model: Optional[str] = None,
) -> dict:
"""Measure per-stage latency and peak VRAM on a single image.
For each requested stage:
1. Run ``n_warmup`` times (results discarded) to prime CUDA kernel
caches and avoid measuring JIT / first-call overhead.
2. Reset peak VRAM counter.
3. Run ``n_runs`` times, recording wall-clock time bracketed by
``torch.cuda.synchronize()`` calls so the timer measures only
completed GPU work.
4. Capture ``torch.cuda.max_memory_allocated()`` as peak VRAM.
All models are loaded once before the warmup loop and reused across
stages, matching real deployment behaviour.
Args:
image_path: Path to the image file to benchmark.
n_warmup: Number of throwaway runs before recording. Default 2.
n_runs: Number of timed runs to average over. Default 10.
stages: List of stage numbers to benchmark (subset of [1, 2, 3]).
Defaults to ``[1, 2, 3]``.
force_model: Passed to Pipeline β ``"moondream"`` or ``"qwen"`` to
override the VRAM-based auto-selection.
Returns:
Dict keyed by stage number (int). Each value is a dict with::
mean_ms β mean wall-clock latency in milliseconds
std_ms β standard deviation of latency
min_ms β minimum observed latency
max_ms β maximum observed latency
peak_vram_mb β peak CUDA memory allocated during timed runs (MB)
breakdown β sub-timing means in ms (stage-dependent keys)
A top-level ``"meta"`` key holds image path, n_warmup, n_runs.
Raises:
FileNotFoundError: If ``image_path`` does not exist.
ValueError: If ``stages`` contains a value outside [1, 2, 3].
"""
if stages is None:
stages = [1, 2, 3]
bad = [s for s in stages if s not in (1, 2, 3)]
if bad:
raise ValueError(f"stages must be subset of {{1, 2, 3}}, got: {bad}")
image_path = Path(image_path)
if not image_path.exists():
raise FileNotFoundError(f"Image not found: {image_path}")
frame_rgb = np.array(Image.open(image_path).convert("RGB"))
# Load all models once before any timing starts.
print(f"Loading pipeline models (force_model={force_model!r})...")
pipeline = Pipeline(force_model=force_model)
# Trigger lazy model loading now so first warmup isn't the load call.
_preload_models(pipeline, stages, frame_rgb)
results: dict = {
"meta": {
"image": str(image_path),
"n_warmup": n_warmup,
"n_runs": n_runs,
}
}
for stage in stages:
run_fn = _stage_runner(pipeline, stage, frame_rgb)
print(
f"\nStage {stage}: {n_warmup} warmup + {n_runs} timed runs...",
flush=True,
)
# ββ Warmup ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
for _ in range(n_warmup):
run_fn()
# ββ Timed runs ββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
if torch.cuda.is_available():
torch.cuda.reset_peak_memory_stats()
wall_times_ms: list[float] = []
sub_timings: list[dict[str, float]] = []
for r in range(n_runs):
if torch.cuda.is_available():
torch.cuda.synchronize()
t0 = time.perf_counter()
_, timing = run_fn()
if torch.cuda.is_available():
torch.cuda.synchronize()
elapsed_ms = (time.perf_counter() - t0) * 1000.0
wall_times_ms.append(elapsed_ms)
sub_timings.append(timing)
print(f" run {r + 1:2d}/{n_runs}: {elapsed_ms:7.1f} ms", flush=True)
peak_vram_mb = (
torch.cuda.max_memory_allocated() / (1024 ** 2)
if torch.cuda.is_available()
else 0.0
)
arr = np.array(wall_times_ms)
results[stage] = {
"mean_ms": float(np.mean(arr)),
"std_ms": float(np.std(arr)),
"min_ms": float(np.min(arr)),
"max_ms": float(np.max(arr)),
"peak_vram_mb": peak_vram_mb,
"breakdown": _mean_breakdown(sub_timings),
}
return results
# ---------------------------------------------------------------------------
# Helpers
# ---------------------------------------------------------------------------
def _preload_models(
pipeline: Pipeline,
stages: list[int],
frame_rgb: np.ndarray,
) -> None:
"""Trigger lazy model loading with a single throwaway run per stage."""
print(" Preloading models (one throwaway run per stage)...", flush=True)
if 1 in stages:
pipeline.run_stage1(frame_rgb)
if 2 in stages:
pipeline.run_stage2(frame_rgb)
if 3 in stages:
pipeline.run_stage3(frame_rgb)
print(" Models loaded.", flush=True)
def _stage_runner(pipeline: Pipeline, stage: int, frame_rgb: np.ndarray):
"""Return a zero-argument callable that runs the requested stage.
Returns ``(description, timing)`` so the caller can inspect sub-timings.
Stage 2 and 3 return three values; we normalise to two here.
"""
if stage == 1:
def _run():
desc, t = pipeline.run_stage1(frame_rgb)
return desc, t
elif stage == 2:
def _run():
desc, _ctx, t = pipeline.run_stage2(frame_rgb)
return desc, t
else:
def _run():
desc, _ctx, t = pipeline.run_stage3(frame_rgb)
return desc, t
return _run
def _mean_breakdown(timings: list[dict[str, float]]) -> dict[str, float]:
"""Average each sub-timing key across all runs, converting to ms."""
if not timings:
return {}
keys = [k for k in timings[0] if k not in ("vram_mb", "n_detections")]
result: dict[str, float] = {}
for k in keys:
vals = [t[k] * 1000.0 for t in timings if k in t]
if vals:
result[f"{k}_mean_ms"] = float(np.mean(vals))
return result
# ---------------------------------------------------------------------------
# CSV writer
# ---------------------------------------------------------------------------
_CSV_FIELDNAMES = [
"stage", "mean_ms", "std_ms", "min_ms", "max_ms", "peak_vram_mb",
]
def write_csv(results: dict, output_csv: Path) -> None:
"""Write benchmark results to a CSV file.
Args:
results: Return value of :func:`benchmark`.
output_csv: Destination path.
"""
output_csv.parent.mkdir(parents=True, exist_ok=True)
rows = []
for stage in (1, 2, 3):
if stage not in results:
continue
r = results[stage]
rows.append({
"stage": stage,
"mean_ms": round(r["mean_ms"], 2),
"std_ms": round(r["std_ms"], 2),
"min_ms": round(r["min_ms"], 2),
"max_ms": round(r["max_ms"], 2),
"peak_vram_mb": round(r["peak_vram_mb"], 1),
})
with open(output_csv, "w", newline="", encoding="utf-8") as fh:
writer = csv.DictWriter(fh, fieldnames=_CSV_FIELDNAMES)
writer.writeheader()
writer.writerows(rows)
# ---------------------------------------------------------------------------
# Pretty-print summary
# ---------------------------------------------------------------------------
def print_summary(results: dict) -> None:
"""Print a formatted three-stage comparison table to stdout.
Highlights whether Stage 3 meets the <700 ms abstract target.
Args:
results: Return value of :func:`benchmark`.
"""
meta = results.get("meta", {})
n_runs = meta.get("n_runs", "?")
image = Path(meta.get("image", "?")).name
sep = "-" * 74
print(f"\n{sep}")
print(f" Latency Benchmark β {image} ({n_runs} runs each)")
print(sep)
print(
f" {'Stage':<22} {'Mean ms':>9} {'Std ms':>8} "
f"{'Min ms':>8} {'Max ms':>8} {'Peak VRAM':>10}"
)
print(sep)
stage_labels = {
1: "Stage 1 (VLM only)",
2: "Stage 2 (VLM+Depth)",
3: "Stage 3 (VLM+Depth+YOLO)",
}
target_ms = 700.0
for stage in (1, 2, 3):
if stage not in results:
continue
r = results[stage]
tag = ""
if stage == 3:
tag = " β <700ms" if r["mean_ms"] < target_ms else " β >700ms"
print(
f" {stage_labels[stage]:<22} "
f"{r['mean_ms']:>9.1f} "
f"{r['std_ms']:>8.1f} "
f"{r['min_ms']:>8.1f} "
f"{r['max_ms']:>8.1f} "
f"{r['peak_vram_mb']:>8.0f} MB"
f"{tag}"
)
print(sep)
# Sub-timing breakdown
for stage in (1, 2, 3):
if stage not in results:
continue
bd = results[stage].get("breakdown", {})
if not bd:
continue
parts = " | ".join(
f"{k.replace('_mean_ms', '')}: {v:.1f} ms"
for k, v in sorted(bd.items())
)
print(f" S{stage} breakdown: {parts}")
print(sep)
# Abstract target verdict
if 3 in results:
mean3 = results[3]["mean_ms"]
verdict = (
f"PASS ({mean3:.1f} ms < 700 ms)"
if mean3 < target_ms
else f"FAIL ({mean3:.1f} ms β₯ 700 ms)"
)
print(f" Abstract target (Stage 3 < 700 ms): {verdict}")
print(sep)
# ---------------------------------------------------------------------------
# CLI
# ---------------------------------------------------------------------------
def _parse_args(argv: list[str] | None = None) -> argparse.Namespace:
p = argparse.ArgumentParser(
description=(
"Benchmark per-stage latency and peak VRAM. "
"Headline target: Stage 3 total < 700 ms."
)
)
p.add_argument(
"--image",
required=True,
help="Path to the representative image to benchmark.",
)
p.add_argument(
"--output",
default=None,
help=(
"Destination CSV "
"(default: outputs/results/latency.csv)."
),
)
p.add_argument(
"--n-warmup",
type=int,
default=2,
dest="n_warmup",
help="Number of warmup runs before recording (default: 2).",
)
p.add_argument(
"--n-runs",
type=int,
default=N_LATENCY_RUNS,
dest="n_runs",
help=f"Number of timed runs to average (default: {N_LATENCY_RUNS}).",
)
p.add_argument(
"--stages",
nargs="+",
type=int,
choices=[1, 2, 3],
default=[1, 2, 3],
help="Stages to benchmark (default: 1 2 3).",
)
p.add_argument(
"--force-model",
choices=["moondream", "qwen"],
default=None,
dest="force_model",
help="Override VRAM-based VLM selection.",
)
return p.parse_args(argv)
def main(argv: list[str] | None = None) -> None:
"""CLI entry point."""
args = _parse_args(argv)
output_csv = Path(args.output) if args.output else RESULTS_DIR / "latency.csv"
results = benchmark(
image_path=Path(args.image),
n_warmup=args.n_warmup,
n_runs=args.n_runs,
stages=sorted(set(args.stages)),
force_model=args.force_model,
)
print_summary(results)
write_csv(results, output_csv)
print(f"\n CSV written to: {output_csv}")
if __name__ == "__main__":
main()
|