vram_limit is not taking effect

[2502128021]

define vram config

vram_config_light = {
"offload_dtype": torch.bfloat16,
"offload_device": "cpu",
"onload_dtype": torch.bfloat16,
"onload_device": "cpu",
"preparing_dtype": torch.bfloat16,
"preparing_device": device,
"computation_dtype": torch.bfloat16,
"computation_device": device,
}
vram_config = vram_config_light
vram_limit = 22.0

initialize Qwen Image Pipeline

pipe = QwenImagePipeline_2511.from_pretrained(
scheduler_template="Qwen-Image-LightX2v",
torch_dtype=torch.bfloat16,
device=device,
model_configs=[
ModelConfig_2511(
model_id="Qwen-Image-Edit-2511",
origin_file_pattern="transformer/diffusion_pytorch_model*.safetensors",
local_model_path=local_model_path,
skip_download=True,
*vram_config
),
ModelConfig_2511(
model_id="Qwen-Image-Edit-2511",
origin_file_pattern="text_encoder/model.safetensors",
local_model_path=local_model_path,
skip_download=True,
**vram_config
),
ModelConfig_2511(
model_id="Qwen-Image-Edit-2511",
origin_file_pattern="vae/diffusion_pytorch_model.safetensors",
local_model_path=local_model_path,
skip_download=True,
**vram_config
),
],
tokenizer_config=ModelConfig_2511(
model_id="Qwen-Image-Edit-2511",
origin_file_pattern="tokenizer/",
local_model_path=local_model_path,
skip_download=True
),
processor_config=ModelConfig_2511(
model_id="Qwen-Image-Edit-2511",
origin_file_pattern="processor/",
local_model_path=local_model_path,
skip_download=True
),
vram_limit=vram_limit,
)

when I set vram_limit 22.0 and run qwen-image-edit-2511 on 4090 24G, it's out of memory.
it seems vram_limit is not taking effect.
anybody know how to run qwen-image-edit-2511 on 4090 24G?

[mi804]

Referring to your settings, I conducted a test. With vram_limit set to 20 GB, the peak VRAM usage was approximately 21 GB. You could lower your vram_limit a bit further—this won't significantly affect generation speed. My test code is as follows:

from diffsynth.pipelines.qwen_image import QwenImagePipeline, ModelConfig
from PIL import Image
import torch

vram_config_light = {
"offload_dtype": torch.bfloat16,
"offload_device": "cpu",
"onload_dtype": torch.bfloat16,
"onload_device": "cpu",
"preparing_dtype": torch.bfloat16,
"preparing_device": "cuda",
"computation_dtype": torch.bfloat16,
"computation_device": "cuda",
}
vram_config = vram_config_light
vram_limit = 20.0

print(torch.cuda.mem_get_info("cuda")[1] / (1024 ** 3) - 0.5)

pipe = QwenImagePipeline.from_pretrained(
    torch_dtype=torch.bfloat16,
    device="cuda",
    model_configs=[
        ModelConfig(model_id="Qwen/Qwen-Image-Edit-2511", origin_file_pattern="transformer/diffusion_pytorch_model*.safetensors", **vram_config),
        ModelConfig(model_id="Qwen/Qwen-Image", origin_file_pattern="text_encoder/model*.safetensors", **vram_config),
        ModelConfig(model_id="Qwen/Qwen-Image", origin_file_pattern="vae/diffusion_pytorch_model.safetensors", **vram_config),
    ],
    processor_config=ModelConfig(model_id="Qwen/Qwen-Image-Edit", origin_file_pattern="processor/"),
    vram_limit=vram_limit,
)
prompt = "let this girl smile"
edit_image = [Image.open("image.jpg")]
image = pipe(
    prompt,
    edit_image=edit_image,
    seed=2,
    num_inference_steps=2,
    height=1024,
    width=1536,
    edit_image_auto_resize=True,
    zero_cond_t=True, # This is a special parameter introduced by Qwen-Image-Edit-2511
)
image.resize((1536, 1024)).save("image_smile_1.jpg")