text_to_image_sdxl_enterprise.py

import argparse
import os
import time

import torch
import torch.nn as nn

# oneflow_compile should be imported before importing any diffusers
from onediff.infer_compiler import oneflow_compile, OneflowCompileOptions


def parse_args():
    parser = argparse.ArgumentParser()

    parser.add_argument("--model", type=str, required=True)
    parser.add_argument("--saved_image", type=str, required=True)
    parser.add_argument("--save_graph", action="store_true")
    parser.add_argument("--load_graph", action="store_true")
    parser.add_argument(
        "--prompt",
        type=str,
        default="photo of a rhino dressed suit and tie sitting at a table in a bar with a bar stools, award winning photography, Elke vogelsang",
    )
    parser.add_argument("--height", type=int, default=1024)
    parser.add_argument("--width", type=int, default=1024)
    parser.add_argument("--steps", type=int, default=30)
    parser.add_argument("--bits", type=int, default=8)
    parser.add_argument(
        "--compile",
        default=True,
        type=(lambda x: str(x).lower() in ["true", "1", "yes"]),
    )
    parser.add_argument(
        "--compile_text_encoder",
        default=False,
        type=(lambda x: str(x).lower() in ["true", "1", "yes"]),
        help=(
            "Switch controls whether text_encoder is compiled (default: False). "
            "If your CPU is powerful, turning it on will shorten end-to-end time."
        ),
    )
    parser.add_argument(
        "--graph",
        default=True,
        type=(lambda x: str(x).lower() in ["true", "1", "yes"]),
    )
    parser.add_argument("--seed", type=int, default=1)
    parser.add_argument("--warmup", type=int, default=1)

    args = parser.parse_args()
    return args


args = parse_args()

assert os.path.isfile(
    os.path.join(args.model, "calibrate_info.txt")
), f"calibrate_info.txt is required in args.model ({args.model})"

import onediff_quant
from diffusers import StableDiffusionXLPipeline
from onediff_quant.utils import replace_sub_module_with_quantizable_module

onediff_quant.enable_load_quantized_model()
infer_args = {
    "prompt": args.prompt,
    "height": args.height,
    "width": args.width,
    "num_inference_steps": args.steps,
}
calibrate_info = {}
with open(os.path.join(args.model, "calibrate_info.txt"), "r") as f:
    for line in f.readlines():
        line = line.strip()
        items = line.split(" ")
        calibrate_info[items[0]] = [
            float(items[1]),
            int(items[2]),
            [float(x) for x in items[3].split(",")],
        ]

pipe = StableDiffusionXLPipeline.from_pretrained(
    args.model,
    torch_dtype=torch.float16,
    use_safetensors=True,
)
pipe.to("cuda")


for sub_module_name, sub_calibrate_info in calibrate_info.items():
    replace_sub_module_with_quantizable_module(
        pipe.unet,
        sub_module_name,
        sub_calibrate_info,
        False,
        False,
        args.bits,
    )

compile_options = OneflowCompileOptions()
compile_options.use_graph = args.graph

if args.compile_text_encoder:
    if pipe.text_encoder is not None:
        pipe.text_encoder = oneflow_compile(pipe.text_encoder, options=compile_options)
    if pipe.text_encoder_2 is not None:
        pipe.text_encoder_2 = oneflow_compile(
            pipe.text_encoder_2, options=compile_options
        )

if args.compile:
    pipe.unet = oneflow_compile(pipe.unet, options=compile_options)
    pipe.vae.decoder = oneflow_compile(pipe.vae.decoder, options=compile_options)


if args.load_graph:
    print("Loading graphs to avoid compilation...")
    start_t = time.time()
    pipe.unet.load_graph("base_unet_compiled", run_warmup=True)
    pipe.vae.decoder.load_graph("base_vae_compiled", run_warmup=True)
    end_t = time.time()
    print(f"warmup with loading graph elapsed: {end_t - start_t} s")
    start_t = time.time()
    for _ in range(args.warmup):
        torch.manual_seed(args.seed)
        image = pipe(**infer_args).images[0]
    end_t = time.time()
    print(f"warmup with run elapsed: {end_t - start_t} s")
else:
    start_t = time.time()
    for _ in range(args.warmup):
        torch.manual_seed(args.seed)
        image = pipe(**infer_args).images[0]
    end_t = time.time()
    print(f"warmup with run elapsed: {end_t - start_t} s")

start_t = time.time()

torch.manual_seed(args.seed)
torch.cuda.cudart().cudaProfilerStart()
image = pipe(**infer_args).images[0]
torch.cuda.cudart().cudaProfilerStop()

end_t = time.time()
print(f"e2e ({args.steps} steps) elapsed: {end_t - start_t} s")

image.save(args.saved_image)

if args.save_graph:
    print("Saving graphs...")
    start_t = time.time()
    pipe.unet.save_graph("base_unet_compiled")
    pipe.vae.decoder.save_graph("base_vae_compiled")
    end_t = time.time()
    print(f"save graphs elapsed: {end_t - start_t} s")