Merge branch 'main' into bria_3_2_pipeline

Author: galbria

docs/source/en/api/pipelines/qwenimage.md

@@ -16,7 +16,12 @@
 
 Qwen-Image from the Qwen team is an image generation foundation model in the Qwen series that achieves significant advances in complex text rendering and precise image editing. Experiments show strong general capabilities in both image generation and editing, with exceptional performance in text rendering, especially for Chinese.
 
-Check out the model card [here](https://huggingface.co/Qwen/Qwen-Image) to learn more.
+Qwen-Image comes in the following variants:
+
+| model type | model id |
+|:----------:|:--------:|
+| Qwen-Image | [`Qwen/Qwen-Image`](https://huggingface.co/Qwen/Qwen-Image) |
+| Qwen-Image-Edit | [`Qwen/Qwen-Image-Edit`](https://huggingface.co/Qwen/Qwen-Image-Edit) |
 
 <Tip>
 
@@ -81,16 +86,18 @@ image.save("qwen_fewsteps.png")
 
 </details>
 
+<Tip>
+
+The `guidance_scale` parameter in the pipeline is there to support future guidance-distilled models when they come up. Note that passing `guidance_scale` to the pipeline is ineffective. To enable classifier-free guidance, please pass `true_cfg_scale` and `negative_prompt` (even an empty negative prompt like " ") should enable classifier-free guidance computations.
+
+</Tip>
+
 ## QwenImagePipeline
 
 [[autodoc]] QwenImagePipeline
   - all
   - __call__
 
-## QwenImagePipelineOutput
-
-[[autodoc]] pipelines.qwenimage.pipeline_output.QwenImagePipelineOutput
-
 ## QwenImageImg2ImgPipeline
 
 [[autodoc]] QwenImageImg2ImgPipeline
@@ -102,3 +109,13 @@ image.save("qwen_fewsteps.png")
 [[autodoc]] QwenImageInpaintPipeline
   - all
   - __call__
+
+## QwenImageEditPipeline
+
+[[autodoc]] QwenImageEditPipeline
+  - all
+  - __call__
+
+## QwenImagePipelineOutput
+
+[[autodoc]] pipelines.qwenimage.pipeline_output.QwenImagePipelineOutput

examples/dreambooth/README_qwen.md

@@ -75,9 +75,9 @@ Now, we can launch training using:
 ```bash
 export MODEL_NAME="Qwen/Qwen-Image"
 export INSTANCE_DIR="dog"
-export OUTPUT_DIR="trained-sana-lora"
+export OUTPUT_DIR="trained-qwenimage-lora"
 
-accelerate launch train_dreambooth_lora_sana.py \
+accelerate launch train_dreambooth_lora_qwenimage.py \
   --pretrained_model_name_or_path=$MODEL_NAME  \
   --instance_data_dir=$INSTANCE_DIR \
   --output_dir=$OUTPUT_DIR \

src/diffusers/__init__.py

@@ -491,10 +491,10 @@
             "PixArtAlphaPipeline",
             "PixArtSigmaPAGPipeline",
             "PixArtSigmaPipeline",
+            "QwenImageEditPipeline",
             "QwenImageImg2ImgPipeline",
             "QwenImageInpaintPipeline",
             "QwenImagePipeline",
-            "QwenImageEditPipeline",
             "ReduxImageEncoder",
             "SanaControlNetPipeline",
             "SanaPAGPipeline",

src/diffusers/models/autoencoders/autoencoder_dc.py

@@ -299,6 +299,7 @@ def __init__(
         act_fn: Union[str, Tuple[str]] = "silu",
         upsample_block_type: str = "pixel_shuffle",
         in_shortcut: bool = True,
+        conv_act_fn: str = "relu",
     ):
         super().__init__()
 
@@ -349,7 +350,7 @@ def __init__(
         channels = block_out_channels[0] if layers_per_block[0] > 0 else block_out_channels[1]
 
         self.norm_out = RMSNorm(channels, 1e-5, elementwise_affine=True, bias=True)
-        self.conv_act = nn.ReLU()
+        self.conv_act = get_activation(conv_act_fn)
         self.conv_out = None
 
         if layers_per_block[0] > 0:
@@ -414,6 +415,12 @@ class AutoencoderDC(ModelMixin, ConfigMixin, FromOriginalModelMixin):
             The normalization type(s) to use in the decoder.
         decoder_act_fns (`Union[str, Tuple[str]]`, defaults to `"silu"`):
             The activation function(s) to use in the decoder.
+        encoder_out_shortcut  (`bool`, defaults to `True`):
+            Whether to use shortcut at the end of the encoder.
+        decoder_in_shortcut (`bool`, defaults to `True`):
+            Whether to use shortcut at the beginning of the decoder.
+        decoder_conv_act_fn (`str`, defaults to `"relu"`):
+            The activation function to use at the end of the decoder.
         scaling_factor (`float`, defaults to `1.0`):
             The multiplicative inverse of the root mean square of the latent features. This is used to scale the latent
             space to have unit variance when training the diffusion model. The latents are scaled with the formula `z =
@@ -441,6 +448,9 @@ def __init__(
         downsample_block_type: str = "pixel_unshuffle",
         decoder_norm_types: Union[str, Tuple[str]] = "rms_norm",
         decoder_act_fns: Union[str, Tuple[str]] = "silu",
+        encoder_out_shortcut: bool = True,
+        decoder_in_shortcut: bool = True,
+        decoder_conv_act_fn: str = "relu",
         scaling_factor: float = 1.0,
     ) -> None:
         super().__init__()
@@ -454,6 +464,7 @@ def __init__(
             layers_per_block=encoder_layers_per_block,
             qkv_multiscales=encoder_qkv_multiscales,
             downsample_block_type=downsample_block_type,
+            out_shortcut=encoder_out_shortcut,
         )
         self.decoder = Decoder(
             in_channels=in_channels,
@@ -466,6 +477,8 @@ def __init__(
             norm_type=decoder_norm_types,
             act_fn=decoder_act_fns,
             upsample_block_type=upsample_block_type,
+            in_shortcut=decoder_in_shortcut,
+            conv_act_fn=decoder_conv_act_fn,
         )
 
         self.spatial_compression_ratio = 2 ** (len(encoder_block_out_channels) - 1)

src/diffusers/models/model_loading_utils.py

@@ -726,23 +726,29 @@ def _caching_allocator_warmup(
     very large margin.
     """
     factor = 2 if hf_quantizer is None else hf_quantizer.get_cuda_warm_up_factor()
-    # Remove disk and cpu devices, and cast to proper torch.device
+
+    # Keep only accelerator devices
     accelerator_device_map = {
         param: torch.device(device)
         for param, device in expanded_device_map.items()
         if str(device) not in ["cpu", "disk"]
     }
-    total_byte_count = defaultdict(lambda: 0)
+    if not accelerator_device_map:
+        return
+
+    elements_per_device = defaultdict(int)
     for param_name, device in accelerator_device_map.items():
         try:
-            param = model.get_parameter(param_name)
+            p = model.get_parameter(param_name)
         except AttributeError:
-            param = model.get_buffer(param_name)
-        # The dtype of different parameters may be different with composite models or `keep_in_fp32_modules`
-        param_byte_count = param.numel() * param.element_size()
+            try:
+                p = model.get_buffer(param_name)
+            except AttributeError:
+                raise AttributeError(f"Parameter or buffer with name={param_name} not found in model")
         # TODO: account for TP when needed.
-        total_byte_count[device] += param_byte_count
+        elements_per_device[device] += p.numel()
 
     # This will kick off the caching allocator to avoid having to Malloc afterwards
-    for device, byte_count in total_byte_count.items():
-        _ = torch.empty(byte_count // factor, dtype=dtype, device=device, requires_grad=False)
+    for device, elem_count in elements_per_device.items():
+        warmup_elems = max(1, elem_count // factor)
+        _ = torch.empty(warmup_elems, dtype=dtype, device=device, requires_grad=False)

src/diffusers/models/transformers/transformer_cogview4.py

@@ -28,7 +28,7 @@
 from ..embeddings import CogView3CombinedTimestepSizeEmbeddings
 from ..modeling_outputs import Transformer2DModelOutput
 from ..modeling_utils import ModelMixin
-from ..normalization import AdaLayerNormContinuous
+from ..normalization import LayerNorm, RMSNorm
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
@@ -584,6 +584,38 @@ def forward(self, hidden_states: torch.Tensor) -> Tuple[torch.Tensor, torch.Tens
         return (freqs.cos(), freqs.sin())
 
 
+class CogView4AdaLayerNormContinuous(nn.Module):
+    """
+    CogView4-only final AdaLN: LN(x) -> Linear(cond) -> chunk -> affine. Matches Megatron: **no activation** before the
+    Linear on conditioning embedding.
+    """
+
+    def __init__(
+        self,
+        embedding_dim: int,
+        conditioning_embedding_dim: int,
+        elementwise_affine: bool = True,
+        eps: float = 1e-5,
+        bias: bool = True,
+        norm_type: str = "layer_norm",
+    ):
+        super().__init__()
+        self.linear = nn.Linear(conditioning_embedding_dim, embedding_dim * 2, bias=bias)
+        if norm_type == "layer_norm":
+            self.norm = LayerNorm(embedding_dim, eps, elementwise_affine, bias)
+        elif norm_type == "rms_norm":
+            self.norm = RMSNorm(embedding_dim, eps, elementwise_affine)
+        else:
+            raise ValueError(f"unknown norm_type {norm_type}")
+
+    def forward(self, x: torch.Tensor, conditioning_embedding: torch.Tensor) -> torch.Tensor:
+        # *** NO SiLU here ***
+        emb = self.linear(conditioning_embedding.to(x.dtype))
+        scale, shift = torch.chunk(emb, 2, dim=1)
+        x = self.norm(x) * (1 + scale)[:, None, :] + shift[:, None, :]
+        return x
+
+
 class CogView4Transformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, CacheMixin):
     r"""
     Args:
@@ -666,7 +698,7 @@ def __init__(
         )
 
         # 4. Output projection
-        self.norm_out = AdaLayerNormContinuous(inner_dim, time_embed_dim, elementwise_affine=False)
+        self.norm_out = CogView4AdaLayerNormContinuous(inner_dim, time_embed_dim, elementwise_affine=False)
         self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * out_channels, bias=True)
 
         self.gradient_checkpointing = False

src/diffusers/models/transformers/transformer_qwenimage.py

@@ -219,6 +219,7 @@ def forward(self, video_fhw, txt_seq_lens, device):
                 video_freq = self.rope_cache[rope_key]
             else:
                 video_freq = self._compute_video_freqs(frame, height, width, idx)
+            video_freq = video_freq.to(device)
             vid_freqs.append(video_freq)
 
             if self.scale_rope:

src/diffusers/pipelines/qwenimage/__init__.py

@@ -24,9 +24,9 @@
 else:
     _import_structure["modeling_qwenimage"] = ["ReduxImageEncoder"]
     _import_structure["pipeline_qwenimage"] = ["QwenImagePipeline"]
+    _import_structure["pipeline_qwenimage_edit"] = ["QwenImageEditPipeline"]
     _import_structure["pipeline_qwenimage_img2img"] = ["QwenImageImg2ImgPipeline"]
     _import_structure["pipeline_qwenimage_inpaint"] = ["QwenImageInpaintPipeline"]
-    _import_structure["pipeline_qwenimage_edit"] = ["QwenImageEditPipeline"]
 
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     try:

src/diffusers/pipelines/qwenimage/pipeline_qwenimage.py

@@ -480,6 +480,11 @@ def __call__(
                 of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting
                 `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to
                 the text `prompt`, usually at the expense of lower image quality.
+
+                This parameter in the pipeline is there to support future guidance-distilled models when they come up.
+                Note that passing `guidance_scale` to the pipeline is ineffective. To enable classifier-free guidance,
+                please pass `true_cfg_scale` and `negative_prompt` (even an empty negative prompt like " ") should
+                enable classifier-free guidance computations.
             num_images_per_prompt (`int`, *optional*, defaults to 1):
                 The number of images to generate per prompt.
             generator (`torch.Generator` or `List[torch.Generator]`, *optional*):

src/diffusers/pipelines/qwenimage/pipeline_qwenimage_edit.py

@@ -46,15 +46,20 @@
         >>> import torch
         >>> from PIL import Image
         >>> from diffusers import QwenImageEditPipeline
+        >>> from diffusers.utils import load_image
 
         >>> pipe = QwenImageEditPipeline.from_pretrained("Qwen/Qwen-Image-Edit", torch_dtype=torch.bfloat16)
         >>> pipe.to("cuda")
-        >>> prompt = "Change the cat to a dog"
-        >>> image = Image.open("cat.png")
+        >>> image = load_image(
+        ...     "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/yarn-art-pikachu.png"
+        ... ).convert("RGB")
+        >>> prompt = (
+        ...     "Make Pikachu hold a sign that says 'Qwen Edit is awesome', yarn art style, detailed, vibrant colors"
+        ... )
         >>> # Depending on the variant being used, the pipeline call will slightly vary.
         >>> # Refer to the pipeline documentation for more details.
         >>> image = pipe(image, prompt, num_inference_steps=50).images[0]
-        >>> image.save("qwenimageedit.png")
+        >>> image.save("qwenimage_edit.png")
         ```
 """
 PREFERRED_QWENIMAGE_RESOLUTIONS = [
@@ -178,7 +183,7 @@ def calculate_dimensions(target_area, ratio):
 
 class QwenImageEditPipeline(DiffusionPipeline, QwenImageLoraLoaderMixin):
     r"""
-    The QwenImage pipeline for text-to-image generation.
+    The Qwen-Image-Edit pipeline for image editing.
 
     Args:
         transformer ([`QwenImageTransformer2DModel`]):
@@ -217,8 +222,8 @@ def __init__(
             transformer=transformer,
             scheduler=scheduler,
         )
-        self.latent_channels = 16
         self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8
+        self.latent_channels = self.vae.config.z_dim if getattr(self, "vae", None) else 16
         # QwenImage latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible
         # by the patch size. So the vae scale factor is multiplied by the patch size to account for this
         self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2)
@@ -592,6 +597,11 @@ def __call__(
                 of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting
                 `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to
                 the text `prompt`, usually at the expense of lower image quality.
+
+                This parameter in the pipeline is there to support future guidance-distilled models when they come up.
+                Note that passing `guidance_scale` to the pipeline is ineffective. To enable classifier-free guidance,
+                please pass `true_cfg_scale` and `negative_prompt` (even an empty negative prompt like " ") should
+                enable classifier-free guidance computations.
             num_images_per_prompt (`int`, *optional*, defaults to 1):
                 The number of images to generate per prompt.
             generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
@@ -635,7 +645,9 @@ def __call__(
             [`~pipelines.qwenimage.QwenImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When
             returning a tuple, the first element is a list with the generated images.
         """
-        calculated_width, calculated_height, _ = calculate_dimensions(1024 * 1024, image.width / image.height)
+        image_size = image[0].size if isinstance(image, list) else image.size
+        width, height = image_size
+        calculated_width, calculated_height, _ = calculate_dimensions(1024 * 1024, width / height)
         height = height or calculated_height
         width = width or calculated_width