The result of the StableDiffusionAITPipeline cannot be fixed with a seed value.

[ssotabe]

Even if Seed is given to generator to generate, the result changes slightly each time.

def torch_fix_seed(seed=42):
    # Python random
    random.seed(seed)
    # Numpy
    np.random.seed(seed)
    # Pytorch
    torch.manual_seed(seed)
    torch.cuda.manual_seed(seed)
    torch.backends.cudnn.deterministic = True
    torch.use_deterministic_algorithms = True
torch_fix_seed()

Fixing torch seed with such code does not seem to fix the result.

example:

Where is the randomness left?

[hl475]

Thanks for reporting! @terrychenism , can you please help take a look?

[terrychenism]

where did you apply torch_fix_seed?

[ssotabe]

in demo.py I tried two patterns, one before the pipe was created and the other just before the image was generated, but neither made any difference.

def torch_fix_seed(seed=42):
        # Python random
        random.seed(seed)
        # Numpy
        np.random.seed(seed)
        # Pytorch
        torch.manual_seed(seed)
        torch.cuda.manual_seed(seed)
        torch.backends.cudnn.deterministic = True
        torch.use_deterministic_algorithms = True

def run(local_dir, width, height, prompt, benchmark):
    torch_fix_seed() # <-
    pipe = StableDiffusionAITPipeline.from_pretrained(
        local_dir,
        scheduler=EulerDiscreteScheduler.from_pretrained(
            local_dir, subfolder="scheduler"
        ),
        revision="fp16",
        torch_dtype=torch.float16,
    ).to("cuda")
    generator = torch.Generator("cuda").manual_seed(1)
    with torch.autocast("cuda"):
        torch_fix_seed() # <-
        image = pipe(prompt, height, width,generator=generator).images[0]
        if benchmark:
            t = benchmark_torch_function(10, pipe, prompt, height=height, width=width)
            print(f"sd e2e: {t} ms")

    image.save("example_ait.png")

if __name__ == "__main__":
    run()

[terrychenism]

what if put torch_fix_seed inside main?

[ssotabe]

There was no change. code

def torch_fix_seed(seed=42):
    # Python random
    random.seed(seed)
    # Numpy
    np.random.seed(seed)
    # Pytorch
    torch.manual_seed(seed)
    torch.cuda.manual_seed(seed)
    torch.backends.cudnn.deterministic = True
    torch.use_deterministic_algorithms = True
@click.command()
@click.option(
    "--local-dir",
    default="./tmp/diffusers-pipeline/stabilityai/stable-diffusion-v2",
    help="the local diffusers pipeline directory",
)
@click.option("--width", default=512, help="Width of generated image")
@click.option("--height", default=512, help="Height of generated image")
@click.option("--prompt", default="A vision of paradise, Unreal Engine", help="prompt")
@click.option(
    "--benchmark", type=bool, default=False, help="run stable diffusion e2e benchmark"
)
def run(local_dir, width, height, prompt, benchmark):

    pipe = StableDiffusionAITPipeline.from_pretrained(
        local_dir,
        scheduler=EulerDiscreteScheduler.from_pretrained(
            local_dir, subfolder="scheduler"
        ),
        revision="fp16",
        torch_dtype=torch.float16,
    ).to("cuda")
    generator = torch.Generator("cuda").manual_seed(1)
    with torch.autocast("cuda"):
        image = pipe(prompt, height, width,generator=generator).images[0]
        if benchmark:
            t = benchmark_torch_function(10, pipe, prompt, height=height, width=width)
            print(f"sd e2e: {t} ms")

    image.save("example_ait.png")

def main():
    torch_fix_seed()
    run()

if __name__ == "__main__":
    main()

[CanyonWind]

hi any update on this? Is there any chance triton causes the issue

[terrychenism]

you can try to disable use_mem_eff https://github.com/facebookincubator/AITemplate/blob/main/examples/05_stable_diffusion/src/modeling/attention.py#L72

[Stax124]

you can try to disable use_mem_eff https://github.com/facebookincubator/AITemplate/blob/main/examples/05_stable_diffusion/src/modeling/attention.py#L72

I tried that but it is still non deterministic. Latents generate correctly with the fixed seed. It must be somewhere in the AIT code, not the SD example, as this is the only point where torch.randn is called (and correct generator is passed down to the steps process). I will try to do some more debugging this week, I really need that to work for VoltaML.