import argparse from optimum.quanto import freeze, qfloat8, qint4, qint8, quantize import torch import json import torch.utils.benchmark as benchmark from diffusers import DiffusionPipeline import gc WARM_UP_ITERS = 5 PROMPT = "ghibli style, a fantasy landscape with castles" TORCH_DTYPES = {"fp32": torch.float32, "fp16": torch.float16, "bf16": torch.bfloat16} QTYPES = {"fp8": qfloat8, "int8": qint8, "int4": qint4, "none": None} PREFIXES = { "stabilityai/stable-diffusion-3-medium-diffusers": "sd3", } def flush(): """Wipes off memory.""" gc.collect() torch.cuda.empty_cache() torch.cuda.reset_max_memory_allocated() torch.cuda.reset_peak_memory_stats() def load_pipeline( ckpt_id, torch_dtype, qtype=None, exclude_layers=None, qte=False, first=False, second=False, third=False ): pipe = DiffusionPipeline.from_pretrained(ckpt_id, torch_dtype=torch_dtype).to("cuda") if qtype: quantize(pipe.transformer, weights=qtype, exclude=exclude_layers) freeze(pipe.transformer) if qte: if first: quantize(pipe.text_encoder, weights=qtype) freeze(pipe.text_encoder) if second: quantize(pipe.text_encoder_2, weights=qtype) freeze(pipe.text_encoder) if third: quantize(pipe.text_encoder_3, weights=qtype) freeze(pipe.text_encoder_3) pipe.set_progress_bar_config(disable=True) return pipe def run_inference(pipe, batch_size=1): _ = pipe( prompt=PROMPT, num_images_per_prompt=batch_size, generator=torch.manual_seed(0), ) def benchmark_fn(f, *args, **kwargs): t0 = benchmark.Timer(stmt="f(*args, **kwargs)", globals={"args": args, "kwargs": kwargs, "f": f}) return f"{(t0.blocked_autorange().mean):.3f}" def bytes_to_giga_bytes(bytes): return f"{(bytes / 1024 / 1024 / 1024):.3f}" if __name__ == "__main__": parser = argparse.ArgumentParser() parser.add_argument( "--ckpt_id", type=str, default="stabilityai/stable-diffusion-3-medium-diffusers", choices=list(PREFIXES.keys()), ) parser.add_argument("--batch_size", type=int, default=1) parser.add_argument("--torch_dtype", type=str, default="fp16", choices=list(TORCH_DTYPES.keys())) parser.add_argument("--qtype", type=str, default="none", choices=list(QTYPES.keys())) parser.add_argument("--qte", type=int, default=0, help="Quantize text encoder") parser.add_argument("--first", type=int, default=0, help="Quantize first text encoder") parser.add_argument("--second", type=int, default=0, help="Quantize second text encoder") parser.add_argument("--third", type=int, default=0, help="Quantize third text encoder") parser.add_argument("--exclude_layers", metavar="N", type=str, nargs="*", default=None) args = parser.parse_args() flush() print( f"Running with ckpt_id: {args.ckpt_id}, batch_size: {args.batch_size}, torch_dtype: {args.torch_dtype}, qtype: {args.qtype}, qte: {bool(args.qte)}" ) pipeline = load_pipeline( ckpt_id=args.ckpt_id, torch_dtype=TORCH_DTYPES[args.torch_dtype], qtype=QTYPES[args.qtype], exclude_layers=args.exclude_layers, qte=args.qte, first=args.first, second=args.second, third=args.third, ) for _ in range(WARM_UP_ITERS): run_inference(pipeline, args.batch_size) time = benchmark_fn(run_inference, pipeline, args.batch_size) torch.cuda.empty_cache() memory = bytes_to_giga_bytes(torch.cuda.memory_allocated()) # in GBs. print( f"ckpt: {args.ckpt_id} batch_size: {args.batch_size}, qte: {args.qte}, " f"torch_dtype: {args.torch_dtype}, qtype: {args.qtype} in {time} seconds and {memory} GBs." ) ckpt_id = PREFIXES[args.ckpt_id] img_name = f"ckpt@{ckpt_id}-bs@{args.batch_size}-dtype@{args.torch_dtype}-qtype@{args.qtype}-qte@{args.qte}" if args.exclude_layers: exclude_layers = "_".join(args.exclude_layers) img_name += f"-exclude@{exclude_layers}" if args.first: img_name += f"-first@{args.first}" if args.second: img_name += f"-second@{args.second}" if args.third: img_name += f"-third@{args.third}" image = pipeline( prompt=PROMPT, num_images_per_prompt=args.batch_size, generator=torch.manual_seed(0), ).images[0] image.save(f"{img_name}.png") info = dict( batch_size=args.batch_size, memory=memory, time=time, dtype=args.torch_dtype, qtype=args.qtype, qte=args.qte, exclude_layers=args.exclude_layers, first=args.first, second=args.second, third=args.third, ) info_file = f"{img_name}_info.json" with open(info_file, "w") as f: json.dump(info, f)