fix 4K OOM with VAE-tiling

README.md

@@ -36,6 +36,7 @@ As a result, Sana-0.6B is very competitive with modern giant diffusion model (e.
 
 ## 🔥🔥 News
 
+- (🔥 New) \[2025/1/12\] DC-AE tiling makes Sana-4K inferences 4096x4096px images within 22GB GPU memory.[\[Guidance\]](asset/docs/model_zoo.md#-3-4k-models)
 - (🔥 New) \[2025/1/11\] Sana code-base license changed to Apache 2.0.
 - (🔥 New) \[2025/1/10\] Inference Sana with 8bit quantization.[\[Guidance\]](asset/docs/8bit_sana.md#quantization)
 - (🔥 New) \[2025/1/8\] 4K resolution [Sana models](asset/docs/model_zoo.md) is supported in [Sana-ComfyUI](https://github.com/Efficient-Large-Model/ComfyUI_ExtraModels) and [work flow](asset/docs/ComfyUI/Sana_FlowEuler_4K.json) is also prepared. [\[4K guidance\]](asset/docs/ComfyUI/comfyui.md)

asset/docs/model_zoo.md

@@ -77,11 +77,9 @@ image = pipe(
 image[0].save('sana.png')
 ```
 
-#### 2). For 4K models
+## ❗ 3. 4K models
 
-4K models need [patch_conv](https://github.com/mit-han-lab/patch_conv) to avoid OOM issue.(80GB GPU is recommended)
-
-run `pip install patch_conv` first, then
+4K models need VAE tiling to avoid OOM issue.(24 GPU is recommended)
 
 ```python
 # run `pip install git+https://github.com/huggingface/diffusers` before use Sana in diffusers
@@ -98,10 +96,9 @@ pipe.to("cuda")
 pipe.vae.to(torch.bfloat16)
 pipe.text_encoder.to(torch.bfloat16)
 
-# for 4096x4096 image generation OOM issue
+# for 4096x4096 image generation OOM issue, feel free adjust the tile size
 if pipe.transformer.config.sample_size == 128:
-    from patch_conv import convert_model
-    pipe.vae = convert_model(pipe.vae, splits=32)
+    pipe.vae.enable_tiling(tile_sample_min_height=1024, tile_sample_min_width=1024)
 
 prompt = 'a cyberpunk cat with a neon sign that says "Sana"'
 image = pipe(

configs/sana_config/2048ms/Sana_1600M_img2048_bf16.yaml

@@ -41,8 +41,8 @@ model:
     - 8
 # VAE setting
 vae:
-  vae_type: dc-ae
-  vae_pretrained: mit-han-lab/dc-ae-f32c32-sana-1.0
+  vae_type: AutoencoderDC
+  vae_pretrained: mit-han-lab/dc-ae-f32c32-sana-1.0-diffusers
   scale_factor: 0.41407
   vae_latent_dim: 32
   vae_downsample_rate: 32

configs/sana_config/4096ms/Sana_1600M_img4096_bf16.yaml

@@ -41,8 +41,8 @@ model:
     - 8
 # VAE setting
 vae:
-  vae_type: dc-ae
-  vae_pretrained: mit-han-lab/dc-ae-f32c32-sana-1.0
+  vae_type: AutoencoderDC
+  vae_pretrained: mit-han-lab/dc-ae-f32c32-sana-1.0-diffusers
   scale_factor: 0.41407
   vae_latent_dim: 32
   vae_downsample_rate: 32

diffusion/model/builder.py

@@ -15,6 +15,7 @@
 # SPDX-License-Identifier: Apache-2.0
 
 import torch
+from diffusers import AutoencoderDC
 from diffusers.models import AutoencoderKL
 from mmcv import Registry
 from termcolor import colored
@@ -87,6 +88,10 @@ def get_vae(name, model_path, device="cuda"):
         print(colored(f"[DC-AE] Loading model from {model_path}", attrs=["bold"]))
         dc_ae = DCAE_HF.from_pretrained(model_path).to(device).eval()
         return dc_ae
+    elif "AutoencoderDC" in name:
+        print(colored(f"[AutoencoderDC] Loading model from {model_path}", attrs=["bold"]))
+        dc_ae = AutoencoderDC.from_pretrained(model_path).to(device).eval()
+        return dc_ae
     else:
         print("error load vae")
         exit()
@@ -102,8 +107,14 @@ def vae_encode(name, vae, images, sample_posterior, device):
         z = (z - vae.config.shift_factor) * vae.config.scaling_factor
     elif "dc-ae" in name:
         ae = vae
+        scaling_factor = ae.cfg.scaling_factor if ae.cfg.scaling_factor else 0.41407
+        z = ae.encode(images.to(device))
+        z = z * scaling_factor
+    elif "AutoencoderDC" in name:
+        ae = vae
+        scaling_factor = ae.config.scaling_factor if ae.config.scaling_factor else 0.41407
         z = ae.encode(images.to(device))
-        z = z * ae.cfg.scaling_factor
+        z = z * scaling_factor
     else:
         print("error load vae")
         exit()
@@ -116,7 +127,26 @@ def vae_decode(name, vae, latent):
         samples = vae.decode(latent).sample
     elif "dc-ae" in name:
         ae = vae
-        samples = ae.decode(latent.detach() / ae.cfg.scaling_factor)
+        vae_scale_factor = (
+            2 ** (len(ae.config.encoder_block_out_channels) - 1)
+            if hasattr(ae, "config") and ae.config is not None
+            else 32
+        )
+        scaling_factor = ae.cfg.scaling_factor if ae.cfg.scaling_factor else 0.41407
+        if latent.shape[-1] * vae_scale_factor > 4000 or latent.shape[-2] * vae_scale_factor > 4000:
+            from patch_conv import convert_model
+
+            ae = convert_model(ae, splits=4)
+        samples = ae.decode(latent.detach() / scaling_factor)
+    elif "AutoencoderDC" in name:
+        ae = vae
+        scaling_factor = ae.config.scaling_factor if ae.config.scaling_factor else 0.41407
+        try:
+            samples = ae.decode(latent / scaling_factor, return_dict=False)[0]
+        except torch.cuda.OutOfMemoryError as e:
+            print("Warning: Ran out of memory when regular VAE decoding, retrying with tiled VAE decoding.")
+            ae.enable_tiling(tile_sample_min_height=1024, tile_sample_min_width=1024)
+            samples = ae.decode(latent / scaling_factor, return_dict=False)[0]
     else:
         print("error load vae")
         exit()