starflow / sample.py
leoeric's picture
Fix GPU abort: Add memory optimizations (half precision, better monitoring)
fa69135
#
# For licensing see accompanying LICENSE file.
# Copyright (C) 2025 Apple Inc. All Rights Reserved.
#
#!/usr/bin/env python3
"""
Scalable Transformer Autoregressive Flow (STARFlow) Sampling Script
This script provides functionality for sampling from trained transformer autoregressive flow models.
Supports both image and video generation with various conditioning options.
Usage:
python sample.py --model_config_path config.yaml --checkpoint_path model.pth --caption "A cat"
"""
import argparse
import copy
import pathlib
import time
import gc
from typing import Dict, List, Optional, Tuple, Union
import numpy as np
import torch
import torch.nn.functional as F
import torch.utils.data
import torchvision as tv
import tqdm
import yaml
from einops import repeat
from PIL import Image
# Local imports
import transformer_flow
import utils
from dataset import aspect_ratio_to_image_size
from train import get_tarflow_parser
from utils import process_denoising, save_samples_unified, load_model_config, encode_text, add_noise
from transformer_flow import KVCache
from misc import print
# Default caption templates for testing and demonstrations
DEFAULT_CAPTIONS = {
'template1': "In the image, a corgi dog is wearing a Santa hat and is laying on a fluffy rug. The dog's tongue is sticking out and it appears to be happy. There are two pumpkins and a basket of leaves nearby, indicating that the scene takes place during the fall season. The background features a Christmas tree, further suggesting the holiday atmosphere. The image has a warm and cozy feel to it, with the dog looking adorable in its hat and the pumpkins adding a festive touch.",
'template2': "A close-up portrait of a cheerful Corgi dog, showcasing its fluffy, sandy-brown fur and perky ears. The dog has a friendly expression with a slight smile, looking directly into the camera. Set against a soft, natural green background, the image is captured in a high-definition, realistic photography style, emphasizing the texture of the fur and the vibrant colors.",
'template3': "A high-resolution, wide-angle selfie photograph of Albert Einstein in a garden setting. Einstein looks directly into the camera with a gentle, knowing smile. His distinctive wild white hair and bushy mustache frame a face marked by thoughtful wrinkles. He wears a classic tweed jacket over a simple shirt. In the background, lush greenery and flowering bushes under soft daylight create a serene, scholarly atmosphere. Ultra-realistic style, 4K detail.",
'template4': 'A close-up, high-resolution selfie of a red panda perched on a tree branch, its large dark eyes looking directly into the lens. Rich reddish-orange fur with white facial markings contrasts against the lush green bamboo forest behind. Soft sunlight filters through the leaves, casting a warm, natural glow over the scene. Ultra-realistic detail, digital photograph style, 4K resolution.',
'template5': "A realistic selfie of a llama standing in front of a classic Ivy League building on the Princeton University campus. He is smiling gently, wearing his iconic wild hair and mustache, dressed in a wool sweater and collared shirt. The photo has a vintage, slightly sepia tone, with soft natural lighting and leafy trees in the background, capturing an academic and historical vibe.",
}
def setup_model_and_components(args: argparse.Namespace) -> Tuple[torch.nn.Module, Optional[torch.nn.Module], tuple]:
"""Initialize and load the model, VAE, and text encoder."""
# Initialize distributed training context
# For single GPU inference, we still need to initialize process group
# because the model code uses torch.distributed.all_reduce
dist = utils.Distributed()
# If not running with torchrun, initialize single-process group
# This is needed because the model code uses torch.distributed.all_reduce
# Works for both CUDA and CPU modes
if not dist.distributed:
import os
# Initialize single-process process group for model compatibility
if not torch.distributed.is_initialized():
os.environ['MASTER_ADDR'] = 'localhost'
os.environ['MASTER_PORT'] = '12355'
os.environ['RANK'] = '0'
os.environ['LOCAL_RANK'] = '0'
os.environ['WORLD_SIZE'] = '1'
# Use 'nccl' for CUDA, 'gloo' for CPU
backend = 'nccl' if torch.cuda.is_available() else 'gloo'
torch.distributed.init_process_group(
backend=backend,
init_method='env://',
world_size=1,
rank=0,
)
print(f"✅ Initialized single-process distributed group (backend: {backend}) for model compatibility")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# Set random seed
utils.set_random_seed(args.seed + dist.rank)
# Setup text encoder
print("Loading text encoder...")
tokenizer, text_encoder = utils.setup_encoder(args, dist, device)
torch.cuda.empty_cache() # Clear cache after text encoder
# Setup VAE if specified
vae = None
if args.vae is not None:
print("Loading VAE...")
vae = utils.setup_vae(args, dist, device)
args.img_size = args.img_size // vae.downsample_factor
torch.cuda.empty_cache() # Clear cache after VAE
else:
args.finetuned_vae = 'none'
# Setup main transformer model
print("Loading main transformer model...")
model = utils.setup_transformer(
args, dist,
txt_dim=text_encoder.config.hidden_size,
use_checkpoint=1
)
# Load checkpoint to CPU first, then move to GPU
print(f"Loading checkpoint from: {args.checkpoint_path}")
state_dict = torch.load(args.checkpoint_path, map_location='cpu')
model.load_state_dict(state_dict, strict=False)
del state_dict
gc.collect() # Force garbage collection
torch.cuda.empty_cache() # Clear any GPU cache
# Move model to GPU after loading weights
print("Moving model to GPU...")
model = model.to(device)
torch.cuda.empty_cache() # Clear cache after moving to GPU
# Set model to eval mode and disable gradients
for p in model.parameters():
p.requires_grad = False
model.eval()
# Parallelize model for multi-GPU sampling (do this before half precision conversion)
_, model = utils.parallelize_model(args, model, dist, device)
torch.cuda.empty_cache()
# Convert model to half precision for memory efficiency (if CUDA available)
# Do this AFTER parallelization to avoid issues
if torch.cuda.is_available():
# Use bfloat16 if supported, otherwise float16
if torch.cuda.is_bf16_supported():
model = model.to(torch.bfloat16)
print("✅ Converted model to bfloat16 for memory efficiency")
else:
model = model.to(torch.float16)
print("✅ Converted model to float16 for memory efficiency")
torch.cuda.empty_cache()
# Print memory usage
allocated = torch.cuda.memory_allocated(0) / 1024**3
reserved = torch.cuda.memory_reserved(0) / 1024**3
total = torch.cuda.get_device_properties(0).total_memory / 1024**3
print(f"📊 GPU Memory: {allocated:.2f} GB allocated, {reserved:.2f} GB reserved, {total:.2f} GB total")
torch.cuda.empty_cache() # Final cache clear
return model, vae, (tokenizer, text_encoder, dist, device)
def prepare_captions(args: argparse.Namespace, dist) -> Tuple[List[str], List[int], int, str]:
"""Prepare captions for sampling from file or template."""
if args.caption.endswith('.txt'):
with open(args.caption, 'r') as f:
lines = [line.strip() for line in f.readlines()]
num_samples = len(lines)
fixed_y = lines[dist.rank:][::dist.world_size]
fixed_idxs = list(range(len(lines)))[dist.rank:][::dist.world_size]
caption_name = args.caption.split('/')[-1][:-4]
else:
caption_text = DEFAULT_CAPTIONS.get(args.caption, args.caption)
fixed_y = [caption_text] * args.sample_batch_size
fixed_idxs = []
num_samples = args.sample_batch_size * dist.world_size
caption_name = args.caption
return fixed_y, fixed_idxs, num_samples, caption_name
def get_noise_shape(args: argparse.Namespace, vae) -> callable:
"""Generate noise tensor with appropriate shape for sampling."""
def _get_noise_func(b: int, x_shape: tuple) -> torch.Tensor:
rand_shape = [args.channel_size, x_shape[0], x_shape[1]]
if len(x_shape) == 3:
rand_shape = [x_shape[2]] + rand_shape
if vae is not None:
if args.vid_size is not None:
rand_shape[0] = (rand_shape[0] - 1) // vae.temporal_downsample_factor + 1
rand_shape[-2] //= vae.downsample_factor
rand_shape[-1] //= vae.downsample_factor
return torch.randn(b, *rand_shape)
return _get_noise_func
def prepare_input_image(args: argparse.Namespace, x_shape: tuple, vae, device: torch.device, noise_std: float) -> Optional[torch.Tensor]:
"""Load and preprocess input image for conditional generation."""
input_image = Image.open(args.input_image).convert('RGB')
# Resize and crop to target shape
scale = max(x_shape[0] / input_image.height, x_shape[1] / input_image.width)
transform = tv.transforms.Compose([
tv.transforms.Resize((int(input_image.height * scale), int(input_image.width * scale))),
tv.transforms.CenterCrop(x_shape[:2]),
tv.transforms.ToTensor(),
tv.transforms.Normalize([0.5]*3, [0.5]*3)
])
input_image = transform(input_image).unsqueeze(0).to(device)
# Encode with VAE if available
with torch.no_grad():
if vae is not None:
input_image = vae.encode(input_image)
# Add noise
input_image = add_noise(input_image, noise_std)[0]
return input_image
def build_sampling_kwargs(args: argparse.Namespace, caption_name: str) -> dict:
"""Build sampling keyword arguments based on configuration."""
sampling_kwargs = {
'guidance': args.cfg,
'guide_top': args.guide_top,
'verbose': not caption_name.endswith('/'),
'return_sequence': args.return_sequence,
'jacobi': args.jacobi,
'context_length': args.context_length
}
if args.jacobi:
sampling_kwargs.update({
'jacobi_th': args.jacobi_th,
'jacobi_block_size': args.jacobi_block_size,
'jacobi_max_iter': args.jacobi_max_iter
})
else:
sampling_kwargs.update({
'attn_temp': args.attn_temp,
'annealed_guidance': False
})
return sampling_kwargs
def main(args: argparse.Namespace) -> None:
"""Main sampling function."""
# Load model configuration and merge with command line args
trainer_args = load_model_config(args.model_config_path)
trainer_dict = vars(trainer_args)
trainer_dict.update(vars(args))
args = argparse.Namespace(**trainer_dict)
# Handle target length configuration for video
if args.target_length is not None:
assert args.vid_size is not None, "it must be a video model to use target_length"
assert args.jacobi == 1, "target_length is only supported with jacobi sampling"
if args.target_length == 1: # generate single image
args.vid_size = None
args.out_fps = 0
else:
args.local_attn_window = (int(args.vid_size.split(':')[0]) - 1) // 4 + 1
args.vid_size = f"{args.target_length}:16"
if args.context_length is None:
args.context_length = args.local_attn_window - 1
# Override some settings for sampling
# Disable FSDP for single GPU inference (FSDP can cause CPU fallback)
args.fsdp = 0 # Disable FSDP for single GPU - use regular GPU inference
if args.use_pretrained_lm is not None:
args.text = args.use_pretrained_lm
# Setup model and components
model, vae, (tokenizer, text_encoder, dist, device) = setup_model_and_components(args)
# Setup output directory
model_name = pathlib.Path(args.checkpoint_path).stem
sample_dir: pathlib.Path = args.logdir / f'{model_name}'
if dist.local_rank == 0:
sample_dir.mkdir(parents=True, exist_ok=True)
dist.barrier()
print(f'{" Load ":-^80} {model_name}')
# Prepare captions and sampling parameters
fixed_y, fixed_idxs, num_samples, caption_name = prepare_captions(args, dist)
print(f'Sampling {num_samples} from {args.caption} on {dist.world_size} GPU(s)')
get_noise = get_noise_shape(args, vae)
sampling_kwargs = build_sampling_kwargs(args, caption_name)
noise_std = args.target_noise_std if args.target_noise_std else args.noise_std
# Start sampling
print(f'Starting sampling with global batch size {args.sample_batch_size}x{dist.world_size} GPUs')
if torch.cuda.is_available():
torch.cuda.synchronize()
start_time = time.time()
with torch.no_grad():
device_type = 'cuda' if torch.cuda.is_available() else 'cpu'
# Use bfloat16 for CUDA (memory efficient), float32 for CPU
if torch.cuda.is_available() and torch.cuda.is_bf16_supported():
autocast_dtype = torch.bfloat16
elif torch.cuda.is_available():
autocast_dtype = torch.float16
else:
autocast_dtype = torch.float32
with torch.autocast(device_type=device_type, dtype=autocast_dtype):
for i in tqdm.tqdm(range(int(np.ceil(num_samples / (args.sample_batch_size * dist.world_size))))):
# Determine aspect ratio and image shape
x_aspect = args.aspect_ratio if args.mix_aspect else None
if x_aspect == "random":
x_aspect = np.random.choice([
"1:1", "2:3", "3:2", "16:9", "9:16", "4:5", "5:4", "21:9", "9:21"
])
x_shape = aspect_ratio_to_image_size(
args.img_size * vae.downsample_factor, x_aspect,
multiple=vae.downsample_factor * args.patch_size
)
# Setup text encoder kwargs
text_encoder_kwargs = dict(
aspect_ratio=x_aspect,
fps=args.out_fps if args.fps_cond else None,
noise_std=noise_std if args.cond_noise_level else None
)
# Handle video dimensions
if args.vid_size is not None:
vid_size = tuple(map(int, args.vid_size.split(':')))
out_fps = args.out_fps if args.fps_cond else vid_size[1]
num_frames = vid_size[0]
x_shape = (x_shape[0], x_shape[1], num_frames)
else:
out_fps = args.out_fps
# Prepare batch and captions
b = args.sample_batch_size
y = fixed_y[i * b : (i + 1) * b]
y_caption = copy.deepcopy(y)
# Add null captions for CFG
if args.cfg > 0:
y += [""] * len(y)
# Prepare text & noise
y = encode_text(text_encoder, tokenizer, y, args.txt_size, device, **text_encoder_kwargs)
noise = get_noise(len(y_caption), x_shape).to(device)
# Prepare input image if specified
if args.input_image is not None:
input_image = prepare_input_image(args, x_shape, vae, device, noise_std)
input_image = repeat(input_image, '1 c h w -> b c h w', b=b)
assert args.cfg > 0, "CFG is required for image conditioned generation"
kv_caches = model(input_image.unsqueeze(1), y, context=True)
else:
input_image, kv_caches = None, None
# Generate samples
samples = model(noise, y, reverse=True, kv_caches=kv_caches, **sampling_kwargs)
del kv_caches; torch.cuda.empty_cache() # free up memory
# Apply denoising if enabled
samples = process_denoising(
samples, y_caption, args, model, text_encoder,
tokenizer, text_encoder_kwargs, noise_std
)
# Decode with VAE if available
if args.vae is not None:
dec_fn = vae.decode
else:
dec_fn = lambda x: x
if isinstance(samples, list):
samples = torch.cat([dec_fn(s) for s in samples], dim=-1)
else:
samples = dec_fn(samples)
# Save samples using unified function
print(f' Saving samples ... {sample_dir}')
# Determine save mode based on args
if args.save_folder and args.caption.endswith('.txt'):
grid_mode = "individual" # Save individual files when using caption file
else:
grid_mode = "auto" # Use automatic grid arrangement
save_samples_unified(
samples=samples,
save_dir=sample_dir,
filename_prefix=caption_name[:200] if len(caption_name) > 0 else "samples",
epoch_or_iter=i,
fps=out_fps,
dist=dist,
wandb_log=False, # Let sample.py handle its own wandb logging
grid_arrangement=grid_mode
)
# Print timing statistics
if torch.cuda.is_available():
torch.cuda.synchronize()
elapsed_time = time.time() - start_time
print(f'{model_name} cfg {args.cfg:.2f}, bsz={args.sample_batch_size}x{dist.world_size}, '
f'time={elapsed_time:.2f}s, speed={num_samples / elapsed_time:.2f} images/s')
if __name__ == '__main__':
parser = argparse.ArgumentParser()
# Model config
parser.add_argument('--model_config_path', required=True, type=str, help='path to YAML config file or directory containing config file')
parser.add_argument('--checkpoint_path', required=True, type=str, help='path to local checkpoint file (required when using model_config_path)')
parser.add_argument('--logdir', default='./logs', type=pathlib.Path, help='output directory for generated samples')
parser.add_argument('--save_folder', default=0, type=int)
# Caption, condition
parser.add_argument('--caption', type=str, required=True, help='Caption input (required)')
parser.add_argument('--input_image', default=None, type=str, help='path to the input image for image-conditioned generation')
parser.add_argument('--aspect_ratio', default="1:1", type=str, choices=["random", "1:1", "2:3", "3:2", "16:9", "9:16", "4:5", "5:4", "21:9", "9:21"])
parser.add_argument('--out_fps', default=8, type=int, help='fps for video datasets, only useful if fps_cond is set to 1')
# Sampling parameters
parser.add_argument('--seed', default=191, type=int)
parser.add_argument('--denoising_batch_size', default=1, type=int)
parser.add_argument('--self_denoising_lr', default=1, type=float)
parser.add_argument('--disable_learnable_denoiser', default=0, type=int)
parser.add_argument('--attn_temp', default=1, type=float)
parser.add_argument('--jacobi_th', default=0.005, type=float)
parser.add_argument('--jacobi', default=0, type=int)
parser.add_argument('--jacobi_block_size', default=64, type=int)
parser.add_argument('--jacobi_max_iter', default=32, type=int)
parser.add_argument('--num_samples', default=50000, type=int)
parser.add_argument('--sample_batch_size', default=16, type=int)
parser.add_argument('--return_sequence', default=0, type=int)
parser.add_argument('--cfg', default=5, type=float)
parser.add_argument('--guide_top', default=None, type=int)
parser.add_argument('--finetuned_vae', default="px82zaheuu", type=str)
parser.add_argument('--vae_adapter', default=None)
parser.add_argument('--target_noise_std', default=None, help="option to use different noise_std from the config")
# Video-specific parameters
parser.add_argument('--target_length', default=None, type=int, help="target length maybe longer than training")
parser.add_argument('--context_length', default=16, type=int, help="context length used for consective sampling")
args = parser.parse_args()
if args.input_image and args.input_image == 'none':
args.input_image = None
main(args)