File size: 19,797 Bytes
18b382b |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 |
# Copyright (c) Delanoe Pirard / Aedelon
# Licensed under the Apache License, Version 2.0
"""
Adaptive Batching Module for Depth Anything 3.
This module provides intelligent batch size selection based on available GPU memory,
model parameters, and input resolution. It maximizes throughput by dynamically
adjusting batch sizes to utilize as much GPU memory as safely possible.
Key features:
- Memory profiling for accurate estimation
- Model-specific memory coefficients
- Resolution-aware scaling
- Safety margins to prevent OOM
- Support for CUDA and MPS devices
"""
from __future__ import annotations
import gc
import math
from dataclasses import dataclass, field
from typing import Callable, Iterator, Sequence, TypeVar
import torch
from depth_anything_3.utils.logger import logger
T = TypeVar("T")
# =============================================================================
# Model Memory Profiles
# =============================================================================
@dataclass
class ModelMemoryProfile:
"""Memory profile for a specific model variant.
Attributes:
base_memory_mb: Fixed memory overhead (model weights, buffers)
per_image_mb_at_504: Memory per image at 504px resolution
activation_scale: Scaling factor for activations (quadratic with resolution)
safety_margin: Safety margin to prevent OOM (0.0 to 1.0)
"""
base_memory_mb: float
per_image_mb_at_504: float
activation_scale: float = 1.0
safety_margin: float = 0.15 # 15% safety margin by default
# Empirically measured memory profiles for each model variant
# Values calibrated on various GPU configurations
MODEL_MEMORY_PROFILES: dict[str, ModelMemoryProfile] = {
# Small models (ViT-S backbone)
"da3-small": ModelMemoryProfile(
base_memory_mb=350,
per_image_mb_at_504=180,
activation_scale=0.8,
),
# Base models (ViT-B backbone)
"da3-base": ModelMemoryProfile(
base_memory_mb=800,
per_image_mb_at_504=350,
activation_scale=1.0,
),
# Large models (ViT-L backbone)
"da3-large": ModelMemoryProfile(
base_memory_mb=1600,
per_image_mb_at_504=600,
activation_scale=1.2,
),
"da3metric-large": ModelMemoryProfile(
base_memory_mb=1700,
per_image_mb_at_504=650,
activation_scale=1.2,
),
"da3mono-large": ModelMemoryProfile(
base_memory_mb=1700,
per_image_mb_at_504=650,
activation_scale=1.2,
),
# Giant models (ViT-G backbone)
"da3-giant": ModelMemoryProfile(
base_memory_mb=4500,
per_image_mb_at_504=1200,
activation_scale=1.5,
),
"da3nested-giant-large": ModelMemoryProfile(
base_memory_mb=6000,
per_image_mb_at_504=1500,
activation_scale=1.8,
),
}
# =============================================================================
# Memory Utilities
# =============================================================================
def get_available_memory_mb(device: torch.device) -> float:
"""Get available GPU memory in MB.
Args:
device: Target device
Returns:
Available memory in MB, or float('inf') for CPU
"""
if device.type == "cuda":
torch.cuda.synchronize(device)
total = torch.cuda.get_device_properties(device).total_memory
reserved = torch.cuda.memory_reserved(device)
return (total - reserved) / (1024 * 1024)
elif device.type == "mps":
# MPS doesn't expose free memory directly
# Use system memory as a rough proxy (conservative estimate)
try:
allocated = torch.mps.current_allocated_memory()
# Assume 8GB available for MPS (conservative for most Apple Silicon)
# This can be overridden via environment variable
import os
max_mps_memory_gb = float(os.environ.get("DA3_MPS_MAX_MEMORY_GB", "8"))
max_mps_memory_mb = max_mps_memory_gb * 1024
return max(0, max_mps_memory_mb - (allocated / (1024 * 1024)))
except Exception:
return 6000 # Conservative fallback: 6GB
else:
return float("inf") # CPU has no practical limit for batching
def get_total_memory_mb(device: torch.device) -> float:
"""Get total GPU memory in MB.
Args:
device: Target device
Returns:
Total memory in MB
"""
if device.type == "cuda":
return torch.cuda.get_device_properties(device).total_memory / (1024 * 1024)
elif device.type == "mps":
import os
return float(os.environ.get("DA3_MPS_MAX_MEMORY_GB", "8")) * 1024
else:
return float("inf")
# =============================================================================
# Adaptive Batch Size Calculator
# =============================================================================
@dataclass
class AdaptiveBatchConfig:
"""Configuration for adaptive batching.
Attributes:
min_batch_size: Minimum batch size (default: 1)
max_batch_size: Maximum batch size cap (default: 64)
target_memory_utilization: Target GPU memory usage (0.0 to 1.0)
enable_profiling: Enable runtime memory profiling for calibration
profile_warmup_batches: Number of warmup batches before profiling
"""
min_batch_size: int = 1
max_batch_size: int = 64
target_memory_utilization: float = 0.85 # Use 85% of available memory
enable_profiling: bool = True
profile_warmup_batches: int = 2
class AdaptiveBatchSizeCalculator:
"""
Calculates optimal batch sizes based on available GPU memory.
This class provides intelligent batch size selection that:
1. Estimates memory requirements based on model and resolution
2. Measures actual memory usage during runtime (optional profiling)
3. Adjusts batch sizes dynamically to maximize throughput
Example:
>>> from depth_anything_3.utils.adaptive_batching import AdaptiveBatchSizeCalculator
>>> calc = AdaptiveBatchSizeCalculator(
... model_name="da3-large",
... device=torch.device("cuda"),
... )
>>> # Get optimal batch size for 100 images at 518px
>>> batch_size = calc.compute_optimal_batch_size(
... num_images=100,
... process_res=518,
... )
>>> print(f"Optimal batch size: {batch_size}")
"""
def __init__(
self,
model_name: str,
device: torch.device,
config: AdaptiveBatchConfig | None = None,
):
"""Initialize the adaptive batch size calculator.
Args:
model_name: Name of the DA3 model variant
device: Target device for inference
config: Optional configuration overrides
"""
self.model_name = model_name
self.device = device
self.config = config or AdaptiveBatchConfig()
# Get memory profile for this model
self.profile = MODEL_MEMORY_PROFILES.get(
model_name,
# Fallback to large model profile for unknown models
MODEL_MEMORY_PROFILES["da3-large"]
)
# Runtime calibration data
self._measured_per_image_mb: float | None = None
self._profiling_complete: bool = False
self._batch_count: int = 0
def compute_optimal_batch_size(
self,
num_images: int,
process_res: int = 504,
reserved_memory_mb: float = 0,
) -> int:
"""Compute optimal batch size for given workload.
Args:
num_images: Total number of images to process
process_res: Processing resolution
reserved_memory_mb: Additional memory to reserve (e.g., for other operations)
Returns:
Optimal batch size
"""
# Get available memory
available_mb = get_available_memory_mb(self.device)
if available_mb == float("inf"):
# CPU: return reasonable batch size based on image count
return min(num_images, self.config.max_batch_size)
# Apply target utilization and reserve
usable_mb = (available_mb * self.config.target_memory_utilization) - reserved_memory_mb
# Subtract base model memory
usable_mb -= self.profile.base_memory_mb
if usable_mb <= 0:
logger.warn(
f"Insufficient memory for model. "
f"Available: {available_mb:.0f} MB, "
f"Model base: {self.profile.base_memory_mb:.0f} MB"
)
return self.config.min_batch_size
# Calculate per-image memory requirement
per_image_mb = self._estimate_per_image_memory(process_res)
# Apply safety margin
per_image_mb *= (1 + self.profile.safety_margin)
# Calculate optimal batch size
optimal_batch = int(usable_mb / per_image_mb)
# Clamp to configured bounds
optimal_batch = max(self.config.min_batch_size, optimal_batch)
optimal_batch = min(self.config.max_batch_size, optimal_batch)
optimal_batch = min(num_images, optimal_batch)
logger.debug(
f"Adaptive batch: {optimal_batch} "
f"(available: {available_mb:.0f} MB, "
f"per_image: {per_image_mb:.0f} MB @ {process_res}px)"
)
return optimal_batch
def _estimate_per_image_memory(self, process_res: int) -> float:
"""Estimate memory per image at given resolution.
Memory scales approximately quadratically with resolution.
Args:
process_res: Processing resolution
Returns:
Estimated memory per image in MB
"""
# Use measured value if available from profiling
if self._measured_per_image_mb is not None and self._profiling_complete:
base_per_image = self._measured_per_image_mb
else:
base_per_image = self.profile.per_image_mb_at_504
# Scale quadratically with resolution
resolution_scale = (process_res / 504) ** 2
# Apply model-specific activation scale
return base_per_image * resolution_scale * self.profile.activation_scale
def update_from_profiling(self, batch_size: int, memory_used_mb: float, process_res: int) -> None:
"""Update memory estimates from actual profiling data.
Called after inference to calibrate memory estimates.
Args:
batch_size: Batch size used
memory_used_mb: Actual memory consumed
process_res: Resolution used
"""
if not self.config.enable_profiling:
return
self._batch_count += 1
if self._batch_count <= self.config.profile_warmup_batches:
# Skip warmup batches (memory not stable)
return
# Calculate per-image memory at reference resolution (504)
resolution_scale = (process_res / 504) ** 2
memory_per_image = (memory_used_mb - self.profile.base_memory_mb) / batch_size
memory_at_504 = memory_per_image / resolution_scale / self.profile.activation_scale
# Exponential moving average for stability
alpha = 0.3
if self._measured_per_image_mb is None:
self._measured_per_image_mb = memory_at_504
else:
self._measured_per_image_mb = (
alpha * memory_at_504 +
(1 - alpha) * self._measured_per_image_mb
)
self._profiling_complete = True
logger.debug(
f"Profiling update: measured {memory_at_504:.0f} MB/img @ 504px "
f"(running avg: {self._measured_per_image_mb:.0f} MB)"
)
def get_memory_estimate(self, batch_size: int, process_res: int) -> float:
"""Get estimated total memory for a batch.
Args:
batch_size: Batch size
process_res: Processing resolution
Returns:
Estimated memory in MB
"""
per_image = self._estimate_per_image_memory(process_res)
return self.profile.base_memory_mb + (batch_size * per_image)
# =============================================================================
# Batch Iterator
# =============================================================================
@dataclass
class BatchInfo:
"""Information about a batch for processing.
Attributes:
batch_idx: Index of this batch (0-indexed)
start_idx: Start index in original sequence
end_idx: End index in original sequence (exclusive)
items: Items in this batch
batch_size: Size of this batch
is_last: Whether this is the last batch
"""
batch_idx: int
start_idx: int
end_idx: int
items: list
batch_size: int = field(init=False)
is_last: bool = False
def __post_init__(self):
self.batch_size = len(self.items)
def adaptive_batch_iterator(
items: Sequence[T],
calculator: AdaptiveBatchSizeCalculator,
process_res: int = 504,
reserved_memory_mb: float = 0,
) -> Iterator[BatchInfo]:
"""
Iterate over items with adaptive batch sizes.
This iterator dynamically adjusts batch sizes based on available memory,
potentially increasing throughput compared to fixed batch sizes.
Args:
items: Sequence of items to batch
calculator: Adaptive batch size calculator
process_res: Processing resolution
reserved_memory_mb: Additional memory to reserve
Yields:
BatchInfo objects containing batch data and metadata
Example:
>>> calc = AdaptiveBatchSizeCalculator("da3-large", device)
>>> for batch_info in adaptive_batch_iterator(images, calc, process_res=518):
... result = model.inference(batch_info.items, process_res=518)
... # Process result...
"""
total = len(items)
idx = 0
batch_idx = 0
while idx < total:
remaining = total - idx
# Compute optimal batch size for remaining items
batch_size = calculator.compute_optimal_batch_size(
num_images=remaining,
process_res=process_res,
reserved_memory_mb=reserved_memory_mb,
)
end_idx = min(idx + batch_size, total)
batch_items = list(items[idx:end_idx])
yield BatchInfo(
batch_idx=batch_idx,
start_idx=idx,
end_idx=end_idx,
items=batch_items,
is_last=(end_idx >= total),
)
idx = end_idx
batch_idx += 1
# =============================================================================
# High-Level API
# =============================================================================
def process_with_adaptive_batching(
items: Sequence[T],
process_fn: Callable[[list[T]], list],
model_name: str,
device: torch.device,
process_res: int = 504,
config: AdaptiveBatchConfig | None = None,
progress_callback: Callable[[int, int], None] | None = None,
) -> list:
"""
Process items with adaptive batching for optimal GPU utilization.
This function handles the complete workflow of:
1. Computing optimal batch sizes
2. Processing batches
3. Collecting and returning results
4. Memory cleanup between batches
Args:
items: Sequence of items to process
process_fn: Function to process a batch of items
model_name: Name of the DA3 model
device: Target device
process_res: Processing resolution
config: Optional batching configuration
progress_callback: Optional callback(processed, total) for progress updates
Returns:
List of all results concatenated
Example:
>>> def inference_fn(batch):
... return model.inference(batch, process_res=518)
>>>
>>> results = process_with_adaptive_batching(
... items=image_paths,
... process_fn=inference_fn,
... model_name="da3-large",
... device=torch.device("cuda"),
... process_res=518,
... )
"""
calculator = AdaptiveBatchSizeCalculator(
model_name=model_name,
device=device,
config=config,
)
all_results = []
total = len(items)
for batch_info in adaptive_batch_iterator(items, calculator, process_res):
# Process batch
results = process_fn(batch_info.items)
all_results.extend(results if isinstance(results, list) else [results])
# Progress callback
if progress_callback:
progress_callback(batch_info.end_idx, total)
# Memory cleanup between batches (except last)
if not batch_info.is_last:
gc.collect()
if device.type == "cuda":
torch.cuda.empty_cache()
elif device.type == "mps":
torch.mps.empty_cache()
# Optional: profile memory usage for calibration
if calculator.config.enable_profiling and device.type == "cuda":
memory_used = torch.cuda.max_memory_allocated(device) / (1024 * 1024)
calculator.update_from_profiling(
batch_size=batch_info.batch_size,
memory_used_mb=memory_used,
process_res=process_res,
)
torch.cuda.reset_peak_memory_stats(device)
return all_results
# =============================================================================
# Utility Functions
# =============================================================================
def estimate_max_batch_size(
model_name: str,
device: torch.device,
process_res: int = 504,
target_utilization: float = 0.85,
) -> int:
"""
Estimate maximum batch size for a given model and resolution.
Quick utility function for one-off batch size estimation.
Args:
model_name: Name of the DA3 model
device: Target device
process_res: Processing resolution
target_utilization: Target memory utilization (0.0 to 1.0)
Returns:
Estimated maximum batch size
Example:
>>> max_batch = estimate_max_batch_size("da3-large", torch.device("cuda"), 518)
>>> print(f"Max batch size at 518px: {max_batch}")
"""
config = AdaptiveBatchConfig(target_memory_utilization=target_utilization)
calculator = AdaptiveBatchSizeCalculator(model_name, device, config)
# Return estimate for a large number of images
return calculator.compute_optimal_batch_size(num_images=1000, process_res=process_res)
def log_batch_plan(
num_images: int,
model_name: str,
device: torch.device,
process_res: int = 504,
) -> None:
"""
Log the planned batching strategy for a workload.
Useful for debugging and understanding how images will be batched.
Args:
num_images: Number of images to process
model_name: Name of the DA3 model
device: Target device
process_res: Processing resolution
"""
calculator = AdaptiveBatchSizeCalculator(model_name, device)
total_memory = get_total_memory_mb(device)
available_memory = get_available_memory_mb(device)
batch_size = calculator.compute_optimal_batch_size(num_images, process_res)
num_batches = math.ceil(num_images / batch_size)
memory_per_batch = calculator.get_memory_estimate(batch_size, process_res)
logger.info(
f"Batch Plan for {model_name}:\n"
f" Images: {num_images} @ {process_res}px\n"
f" Device: {device} ({total_memory:.0f} MB total, {available_memory:.0f} MB available)\n"
f" Batch Size: {batch_size}\n"
f" Num Batches: {num_batches}\n"
f" Est. Memory/Batch: {memory_per_batch:.0f} MB"
)
|