The code snippet shows how to run Stable Diffusion 3 with a 8bit T5-xxl, drastically reducing the memory requirements.

run_sd3_8bit.py

from diffusers import StableDiffusion3Pipeline
from transformers import T5EncoderModel
import torch 
import time
import gc

def flush():
    gc.collect()
    torch.cuda.empty_cache()

def bytes_to_giga_bytes(bytes):
    return bytes / 1024 / 1024 / 1024

id = "stabilityai/stable-diffusion-3-medium-diffusers"
text_encoder = T5EncoderModel.from_pretrained(
    id, 
    subfolder="text_encoder_3",
    load_in_8bit=True,
    device_map="auto"
)
pipeline = StableDiffusion3Pipeline.from_pretrained(
    id, 
    text_encoder_3=text_encoder,
    transformer=None,
    vae=None,
    device_map="balanced",
)
with torch.no_grad():
    for _ in range(3):
        prompt = "a photo of a cat"
        (
            prompt_embeds,
            negative_prompt_embeds,
            pooled_prompt_embeds,
            negative_pooled_prompt_embeds,
        ) = pipeline.encode_prompt(prompt=prompt, prompt_2=None, prompt_3=None)
    start = time.time()
    for _ in range(10):
        (
            prompt_embeds,
            negative_prompt_embeds,
            pooled_prompt_embeds,
            negative_pooled_prompt_embeds,
        ) = pipeline.encode_prompt(prompt=prompt, prompt_2=None, prompt_3=None)
    end = time.time()
    avg_prompt_encoding_time = (end - start) / 10

del text_encoder
del pipeline
flush()

pipeline = StableDiffusion3Pipeline.from_pretrained(
    id, 
    text_encoder=None,
    text_encoder_2=None,
    text_encoder_3=None,
    tokenizer=None,
    tokenizer_2=None,
    tokenizer_3=None,
    torch_dtype=torch.float16
).to("cuda")
pipeline.set_progress_bar_config(disable=True)

for _ in range(3):
    _ = pipeline(
        prompt_embeds=prompt_embeds.half(),
        negative_prompt_embeds=negative_prompt_embeds.half(),
        pooled_prompt_embeds=pooled_prompt_embeds.half(),
        negative_pooled_prompt_embeds=negative_pooled_prompt_embeds.half(),
    )
start = time.time()
for _ in range(10):
    _ = pipeline(
        prompt_embeds=prompt_embeds.half(),
        negative_prompt_embeds=negative_prompt_embeds.half(),
        pooled_prompt_embeds=pooled_prompt_embeds.half(),
        negative_pooled_prompt_embeds=negative_pooled_prompt_embeds.half(),
    )
end = time.time()
avg_inference_time = (end - start) / 10

print(f"Average prompt encoding time: {avg_prompt_encoding_time:.3f} seconds.")
print(f"Average inference time: {avg_inference_time:.3f} seconds.")
print(f"Total time: {(avg_prompt_encoding_time + avg_inference_time):.3f} seconds.")
print(
    f"Max memory allocated: {bytes_to_giga_bytes(torch.cuda.max_memory_allocated())} GB"
)

image = pipeline(
    prompt_embeds=prompt_embeds.half(),
    negative_prompt_embeds=negative_prompt_embeds.half(),
    pooled_prompt_embeds=pooled_prompt_embeds.half(),
    negative_pooled_prompt_embeds=negative_pooled_prompt_embeds.half(),
).images[0]
image.save("output_8bit.png")