why this code generating image slower that another script in trainer program import…
- User
why this code generating image slower that another script in trainer program
import gradio as gr import numpy as np import random import torch from diffusers import PixArtSigmaPipeline, EulerAncestralDiscreteScheduler,DPMSolverMultistepScheduler
device = "cuda" if torch.cuda.is_available() else "cpu"
#torch.set_float32_matmul_precision("high")
#torch._inductor.config.conv_1x1_as_mm = True #torch._inductor.config.coordinate_descent_tuning = True #torch._inductor.config.epilogue_fusion = False #torch._inductor.config.coordinate_descent_check_all_directions = True
pipe = PixArtSigmaPipeline.from_pretrained( "H:/OneTrainer/models/test_sigma", torch_dtype=torch.float16, #max_sequence_length = 120 ).to("cuda") pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config)
#pipe.transformer.to(memory_format=torch.channels_last) #pipe.vae.to(memory_format=torch.channels_last)
#pipe.transformer = torch.compile(pipe.transformer, mode="max-autotune", fullgraph=True) #pipe.vae.decode = torch.compile(pipe.vae.decode, mode="max-autotune", fullgraph=True)
MAX_SEED = np.iinfo(np.int32).max MAX_IMAGE_SIZE = 2048
def infer(prompt, negative_prompt="", seed=42, randomize_seed=False, width=1024, height=1024, guidance_scale=5.0, num_inference_steps=28, progress=gr.Progress(track_tqdm=True)):
if randomize_seed: seed = random.randint(0, MAX_SEED) generator = torch.Generator().manual_seed(seed) image = pipe( prompt = prompt, negative_prompt = negative_prompt, width=width, height=height, guidance_scale = guidance_scale, num_inference_steps = num_inference_steps, generator = generator ).images[0] return image, seedexamples = [ "A taco food cart in front of a japanese castle", "The spirit of a tamagotchi wandering in the city of Prague", "A flourecent cat on the moon", "A delicious gummy bear cheesecake slice", ]
css=""" #col-container { margin: 0 auto; max-width: 520px; } """
with gr.Blocks(css=css) as demo:
with gr.Column(elem_id="col-container"): gr.Markdown(f""" # PixArt Sigma 900M Demo of the [PixArt Sigma 900M](https://huggingface.co/dataautogpt3/PixArt-Sigma-900M) model, expanded from [PixArt Sigma 600M](https://huggingface.co/PixArt-alpha/PixArt-Sigma-XL-2-1024-MS) """) with gr.Row(): prompt = gr.Text( label="Prompt", show_label=False, max_lines=1, placeholder="Enter your prompt", container=False, ) run_button = gr.Button("Run", scale=0) result = gr.Image(label="Result", show_label=False) with gr.Accordion("Advanced Settings", open=False): negative_prompt = gr.Text( label="Negative prompt", max_lines=1, value="low quality, bad, watermark", placeholder="Enter a negative prompt", ) seed = gr.Slider( label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0, ) randomize_seed = gr.Checkbox(label="Randomize seed", value=True) with gr.Row(): width = gr.Slider( label="Width", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=2048, ) height = gr.Slider( label="Height", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=2048, ) with gr.Row(): guidance_scale = gr.Slider( label="Guidance scale", minimum=0.0, maximum=10.0, step=0.1, value=5.0, ) num_inference_steps = gr.Slider( label="Number of inference steps", minimum=1, maximum=50, step=1, value=28, ) gr.Examples( examples = examples, fn = infer, inputs = [prompt], outputs = [result, seed], cache_examples="lazy" ) gr.on( triggers=[run_button.click, prompt.submit, negative_prompt.submit], fn = infer, inputs = [prompt, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps], outputs = [result, seed] )demo.queue().launch()
- Assistant
The slower generation of images compared to another script in a training program can be attributed to multiple factors. Here are several potential reasons and improvement suggestions:
-
Model Initialization and Device Loading:
- Model loading and device transfer often take a significant amount of time. Ensure this only happens once and not repeatedly during inferences.
-
Commented Optimizations:
- Several commented lines could potentially speed up your script if enabled. For example:
#torch.set_float32_matmul_precision("high") #torch._inductor.config.conv_1x1_as_mm = True #torch._inductor.config.coordinate_descent_tuning = True #torch._inductor.config.epilogue_fusion = False #torch._inductor.config.coordinate_descent_check_all_directions = True #pipe.transformer.to(memory_format=torch.channels_last) #pipe.vae.to(memory_format=torch.channels_last) #pipe.transformer = torch.compile(pipe.transformer, mode="max-autotune", fullgraph=True) #pipe.vae.decode = torch.compile(pipe.vae.decode, mode="max-autotune", fullgraph=True) - Uncommenting and carefully testing these configurations and optimizations could lead to significant speed improvements.
- Several commented lines could potentially speed up your script if enabled. For example:
-
Scheduler Configuration:
- The selection of the scheduler can impact performance. Make sure
EulerAncestralDiscreteScheduleris optimal for your use case.
- The selection of the scheduler can impact performance. Make sure
-
Image Size:
- Generating larger images (
2048x2048) is more computationally expensive. Compare the performance impact by testing with smaller sizes.
- Generating larger images (
-
Batch Processing:
- Ensure no unnecessary computations are repeated within the infer function. For example, the
torch.Generator()can be instantiated outside the inference function if not dependent on a variable seed.
- Ensure no unnecessary computations are repeated within the infer function. For example, the
-
Scheduler Configuration:
- Make sure the scheduler and model configuration are optimal and up-to-date.
-
Model Precision:
- While you set
torch_dtype=torch.float16, further optimizations are possible depending on the hardware and framework support.
- While you set
Here's a revised version of your script with some optimizations re-enabled:
import gradio as gr import numpy as np import random import torch from diffusers import PixArtSigmaPipeline, EulerAncestralDiscreteScheduler device = "cuda" if torch.cuda.is_available() else "cpu" torch.set_float32_matmul_precision("high") torch._inductor.config.conv_1x1_as_mm = True torch._inductor.config.coordinate_descent_tuning = True torch._inductor.config.epilogue_fusion = False torch._inductor.config.coordinate_descent_check_all_directions = True pipe = PixArtSigmaPipeline.from_pretrained( "H:/OneTrainer/models/test_sigma", torch_dtype=torch.float16, # max_sequence_length = 120 ).to("cuda") pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config) pipe.transformer.to(memory_format=torch.channels_last) pipe.vae.to(memory_format=torch.channels_last) # Using `torch.compile` to optimize transformer and vae.decode pipe.transformer = torch.compile(pipe.transformer, mode="max-autotune", fullgraph=True) pipe.vae.decode = torch.compile(pipe.vae.decode, mode="max-autotune", fullgraph=True) MAX_SEED = np.iinfo(np.int32).max MAX_IMAGE_SIZE = 2048 def infer(prompt, negative_prompt="", seed=42, randomize_seed=False, width=1024, height=1024, guidance_scale=5.0, num_inference_steps=28, progress=gr.Progress(track_tqdm=True)): if randomize_seed: seed = random.randint(0, MAX_SEED) generator = torch.Generator().manual_seed(seed) image = pipe( prompt=prompt, negative_prompt=negative_prompt, width=width, height=height, guidance_scale=guidance_scale, num_inference_steps=num_inference_steps, generator=generator ).images[0] return image, seed examples = [ "A taco food cart in front of a japanese castle", "The spirit of a tamagotchi wandering in the city of Prague", "A flourecent cat on the moon", "A delicious gummy bear cheesecake slice", ] css = """ #col-container { margin: 0 auto; max-width: 520px; } """ with gr.Blocks(css=css) as demo: with gr.Column(elem_id="col-container"): gr.Markdown(f""" # PixArt Sigma 900M Demo of the [PixArt Sigma 900M](https://huggingface.co/dataautogpt3/PixArt-Sigma-900M) model, expanded from [PixArt Sigma 600M](https://huggingface.co/PixArt-alpha/PixArt-Sigma-XL-2-1024-MS) """) with gr.Row(): prompt = gr.Text( label="Prompt", show_label=False, max_lines=1, placeholder="Enter your prompt", container=False, ) run_button = gr.Button("Run", scale=0) result = gr.Image(label="Result", show_label=False) with gr.Accordion("Advanced Settings", open=False): negative_prompt = gr.Text( label="Negative prompt", max_lines=1, value="low quality, bad, watermark", placeholder="Enter a negative prompt", ) seed = gr.Slider( label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0, ) randomize_seed = gr.Checkbox(label="Randomize seed", value=True) with gr.Row(): width = gr.Slider( label="Width", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=2048, ) height = gr.Slider( label="Height", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=2048, ) with gr.Row(): guidance_scale = gr.Slider( label="Guidance scale", minimum=0.0, maximum=10.0, step=0.1, value=5.0, ) num_inference_steps = gr.Slider( label="Number of inference steps", minimum=1, maximum=50, step=1, value=28, ) gr.Examples( examples=examples, fn=infer, inputs=[prompt], outputs=[result, seed], cache_examples="lazy" ) gr.on( triggers=[run_button.click, prompt.submit, negative_prompt.submit], fn=infer, inputs=[prompt, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps], outputs=[result, seed] ) demo.queue().launch()To summarize:
- Specific optimizations and configurations related to the PyTorch backend (Inductor, floating point precision).
- Memory format optimizations in the pipeline.
- Utilizing
torch.compilefor critical model components.
Always profile before and after these changes to ensure they provide the intended performance improvements.
-