benchmark_image.py

import torch

# Set high precision for float32 matrix multiplications.
# This setting optimizes performance on NVIDIA GPUs with Ampere architecture (e.g., A100, RTX 30 series) or newer.
torch.set_float32_matmul_precision("high")

from diffusers import DiffusionPipeline

from torchao.quantization import quantize_, autoquant
import argparse
import json

from utils import cleanup_tmp_directory, benchmark_fn, pretty_print_results, reset_memory, bytes_to_giga_bytes


PROMPT = "Eiffel Tower was Made up of more than 2 million translucent straws to look like a cloud, with the bell tower at the top of the building, Michel installed huge foam-making machines in the forest to blow huge amounts of unpredictable wet clouds in the building's classic architecture."
PREFIXES = {
    "stabilityai/stable-diffusion-3-medium-diffusers": "sd3",
    "PixArt-alpha/PixArt-Sigma-XL-2-1024-MS": "pixart",
    "fal/AuraFlow": "auraflow",
    "black-forest-labs/FLUX.1-dev": "flux-dev",
}


def load_pipeline(
    ckpt_id: str,
    fuse_attn_projections: bool,
    compile: bool,
    quantization: str,
    sparsify: bool,
    compile_vae: bool = False,
) -> DiffusionPipeline:
    pipeline = DiffusionPipeline.from_pretrained(ckpt_id, torch_dtype=torch.bfloat16).to("cuda")

    if fuse_attn_projections:
        pipeline.transformer.fuse_qkv_projections()
        if compile_vae:
            pipeline.vae.fuse_qkv_projections()

    if quantization == "autoquant" and compile:
        pipeline.transformer.to(memory_format=torch.channels_last)
        pipeline.transformer = torch.compile(pipeline.transformer, mode="max-autotune", fullgraph=True)
        if compile_vae:
            pipeline.vae.to(memory_format=torch.channels_last)
            pipeline.vae.decode = torch.compile(pipeline.vae.decode, mode="max-autotune", fullgraph=True)

    if not sparsify:
        if quantization == "int8dq":
            from torchao.quantization import int8_dynamic_activation_int8_weight

            quantize_(pipeline.transformer, int8_dynamic_activation_int8_weight())
            if compile_vae:
                quantize_(pipeline.vae, int8_dynamic_activation_int8_weight())
        elif quantization == "int8wo":
            from torchao.quantization import int8_weight_only

            quantize_(pipeline.transformer, int8_weight_only())
            if compile_vae:
                quantize_(pipeline.vae, int8_weight_only())
        elif quantization == "int4wo":
            from torchao.quantization import int4_weight_only

            quantize_(pipeline.transformer, int4_weight_only())
            if compile_vae:
                quantize_(pipeline.vae, int4_weight_only())
        elif quantization == "fp6_e3m2":
            from torchao.quantization import fpx_weight_only

            quantize_(pipeline.transformer, fpx_weight_only(3, 2))
            if compile_vae:
                quantize_(pipeline.vae, fpx_weight_only(3, 2))

        elif quantization == "fp5_e2m2":
            from torchao.quantization import fpx_weight_only

            quantize_(pipeline.transformer, fpx_weight_only(2, 2))
            if compile_vae:
                quantize_(pipeline.vae, fpx_weight_only(2, 2))

        elif quantization == "fp4_e2m1":
            from torchao.quantization import fpx_weight_only

            quantize_(pipeline.transformer, fpx_weight_only(2, 1))
            if compile_vae:
                quantize_(pipeline.vae, fpx_weight_only(2, 1))
        elif quantization == "fp8wo":
            from torchao.quantization import float8_weight_only

            quantize_(pipeline.transformer, float8_weight_only())
            if compile_vae:
                quantize_(pipeline.vae, float8_weight_only())
        elif quantization == "fp8dq":
            from torchao.quantization import float8_dynamic_activation_float8_weight

            quantize_(pipeline.transformer, float8_dynamic_activation_float8_weight())
            if compile_vae:
                quantize_(pipeline.vae, float8_dynamic_activation_float8_weight())
        elif quantization == "fp8dqrow":
            from torchao.quantization import float8_dynamic_activation_float8_weight
            from torchao.quantization.quant_api import PerRow

            quantize_(pipeline.transformer, float8_dynamic_activation_float8_weight(granularity=PerRow()))
            if compile_vae:
                quantize_(pipeline.vae, float8_dynamic_activation_float8_weight(granularity=PerRow()))
        elif quantization == "autoquant":
            pipeline.transformer = autoquant(pipeline.transformer, error_on_unseen=False)
            if compile_vae:
                pipeline.vae = autoquant(pipeline.vae, error_on_unseen=False)

    if sparsify:
        from torchao.sparsity import sparsify_, int8_dynamic_activation_int8_semi_sparse_weight

        sparsify_(pipeline.transformer, int8_dynamic_activation_int8_semi_sparse_weight())
        if compile_vae:
            sparsify_(pipeline.vae, int8_dynamic_activation_int8_semi_sparse_weight())

    if quantization != "autoquant" and compile:
        pipeline.transformer.to(memory_format=torch.channels_last)
        pipeline.transformer = torch.compile(pipeline.transformer, mode="max-autotune", fullgraph=True)
        if compile_vae:
            pipeline.vae.to(memory_format=torch.channels_last)
            pipeline.vae.decode = torch.compile(pipeline.vae.decode, mode="max-autotune", fullgraph=True)

    pipeline.set_progress_bar_config(disable=True)
    return pipeline


def run_inference(pipe, batch_size):
    _ = pipe(
        prompt=PROMPT,
        num_images_per_prompt=batch_size,
        generator=torch.manual_seed(2024),
    )


def run_benchmark(pipeline, args):
    model_memory = bytes_to_giga_bytes(torch.cuda.memory_allocated())  # in GBs.

    for _ in range(5):
        run_inference(pipeline, batch_size=args.batch_size)

    time = benchmark_fn(run_inference, pipeline, args.batch_size)
    torch.cuda.empty_cache()
    inference_memory = bytes_to_giga_bytes(torch.cuda.max_memory_allocated())

    info = dict(
        ckpt_id=args.ckpt_id,
        batch_size=args.batch_size,
        fuse=args.fuse_attn_projections,
        compile=args.compile,
        compile_vae=args.compile_vae,
        quantization=args.quantization,
        sparsify=args.sparsify,
        model_memory=model_memory,
        inference_memory=inference_memory,
        time=time,
    )

    pretty_print_results(info)
    return info


def serialize_artifacts(info: dict, pipeline, args):
    ckpt_id = PREFIXES[args.ckpt_id]
    prefix = f"ckpt@{ckpt_id}-bs@{args.batch_size}-fuse@{args.fuse_attn_projections}-compile@{args.compile}-compile_vae@{args.compile_vae}-quant@{args.quantization}-sparsify@{args.sparsify}"
    info_file = f"{prefix}_info.json"
    with open(info_file, "w") as f:
        json.dump(info, f)

    image = pipeline(
        prompt=PROMPT,
        num_images_per_prompt=args.batch_size,
        generator=torch.manual_seed(0),
    ).images[0]
    image.save(f"{prefix}.png")


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--ckpt_id",
        default="black-forest-labs/FLUX.1-dev",
        type=str,
        help="Hub model or path to local model for which the benchmark is to be run.",
    )
    parser.add_argument(
        "--fuse_attn_projections",
        action="store_true",
        help="Whether or not to fuse the QKV projection layers into one larger layer.",
    )
    parser.add_argument("--compile", action="store_true", help="Whether or not to torch.compile the models.")
    parser.add_argument("--compile_vae", action="store_true", help="If compiling, should VAE be compiled too?")
    parser.add_argument(
        "--quantization",
        default="None",
        choices=[
            "int8dq",
            "int8wo",
            "int4wo",
            "autoquant",
            "fp8wo",
            "fp8dq",
            "fp8dqrow",
            "fp6_e3m2",
            "fp5_e2m2",
            "fp4_e2m1",
            "None",
        ],
        help="Which quantization technique to apply",
    )
    parser.add_argument("--sparsify", action="store_true")
    parser.add_argument(
        "--batch_size",
        default=1,
        type=int,
        choices=[1, 4, 8, 16],
        help="Number of images to generate for the testing prompt.",
    )
    args = parser.parse_args()

    reset_memory("cuda")

    pipeline = load_pipeline(
        ckpt_id=args.ckpt_id,
        fuse_attn_projections=args.fuse_attn_projections,
        compile=args.compile,
        compile_vae=args.compile_vae,
        quantization=args.quantization,
        sparsify=args.sparsify,
    )

    info = run_benchmark(pipeline, args)
    serialize_artifacts(info, pipeline, args)
    cleanup_tmp_directory()