Merge branch 'main' into remote-utils

Author: hlky

.github/workflows/pr_style_bot.yml

@@ -53,9 +53,9 @@ jobs:
           HEADREF: ${{ steps.pr_info.outputs.headRef }}
           PRNUMBER: ${{ steps.pr_info.outputs.prNumber }}
         run: |
-          echo "PR number: ${{ env.PRNUMBER }}"
-          echo "Head Ref: ${{ env.HEADREF }}"
-          echo "Head Repo Full Name: ${{ env.HEADREPOFULLNAME }}"
+          echo "PR number: $PRNUMBER"
+          echo "Head Ref: $HEADREF"
+          echo "Head Repo Full Name: $HEADREPOFULLNAME"
 
       - name: Set up Python
         uses: actions/setup-python@v4
@@ -89,20 +89,20 @@ jobs:
           PRNUMBER: ${{ steps.pr_info.outputs.prNumber }}
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         run: |
-          echo "HEADREPOFULLNAME: ${{ env.HEADREPOFULLNAME }}, HEADREF: ${{ env.HEADREF }}"
+          echo "HEADREPOFULLNAME: $HEADREPOFULLNAME, HEADREF: $HEADREF"
           # Configure git with the Actions bot user
           git config user.name "github-actions[bot]"
           git config user.email "github-actions[bot]@users.noreply.github.com"
 
           # Make sure your 'origin' remote is set to the contributor's fork
-          git remote set-url origin "https://x-access-token:${GITHUB_TOKEN}@github.com/${{ env.HEADREPOFULLNAME }}.git"
+          git remote set-url origin "https://x-access-token:${GITHUB_TOKEN}@github.com/$HEADREPOFULLNAME.git"
 
           # If there are changes after running style/quality, commit them
           if [ -n "$(git status --porcelain)" ]; then
             git add .
             git commit -m "Apply style fixes"
             # Push to the original contributor's forked branch
-            git push origin HEAD:${{ env.HEADREF }}
+            git push origin HEAD:$HEADREF
             echo "changes_pushed=true" >> $GITHUB_OUTPUT
           else
             echo "No changes to commit."

.github/workflows/pr_tests_gpu.yml

@@ -11,6 +11,8 @@ on:
       - "src/diffusers/loaders/lora_base.py"
       - "src/diffusers/loaders/lora_pipeline.py"
       - "src/diffusers/loaders/peft.py"
+      - "tests/pipelines/test_pipelines_common.py"
+      - "tests/models/test_modeling_common.py"
   workflow_dispatch:
 
 concurrency:
@@ -104,11 +106,18 @@ jobs:
           # https://pytorch.org/docs/stable/notes/randomness.html#avoiding-nondeterministic-algorithms
           CUBLAS_WORKSPACE_CONFIG: :16:8
         run: |
-          pattern=$(cat ${{ steps.extract_tests.outputs.pattern_file }})
-          python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile \
-            -s -v -k "not Flax and not Onnx and $pattern" \
-            --make-reports=tests_pipeline_${{ matrix.module }}_cuda \
-            tests/pipelines/${{ matrix.module }}
+          if [ "${{ matrix.module }}" = "ip_adapters" ]; then 
+              python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile \
+              -s -v -k "not Flax and not Onnx" \
+              --make-reports=tests_pipeline_${{ matrix.module }}_cuda \
+              tests/pipelines/${{ matrix.module }}
+          else 
+              pattern=$(cat ${{ steps.extract_tests.outputs.pattern_file }})
+              python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile \
+              -s -v -k "not Flax and not Onnx and $pattern" \
+              --make-reports=tests_pipeline_${{ matrix.module }}_cuda \
+              tests/pipelines/${{ matrix.module }}
+          fi 
 
       - name: Failure short reports
         if: ${{ failure() }}

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

@@ -65,7 +65,7 @@ The table below lists all the pipelines currently available in 🤗 Diffusers an
 | [Latte](latte) | text2image |
 | [LEDITS++](ledits_pp) | image editing |
 | [Lumina-T2X](lumina) | text2image |
-| [Marigold](marigold) | depth |
+| [Marigold](marigold) | depth-estimation, normals-estimation, intrinsic-decomposition |
 | [MultiDiffusion](panorama) | text2image |
 | [MusicLDM](musicldm) | text2audio |
 | [PAG](pag) | text2image |

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

@@ -345,6 +345,7 @@
             "Lumina2Text2ImgPipeline",
             "LuminaText2ImgPipeline",
             "MarigoldDepthPipeline",
+            "MarigoldIntrinsicsPipeline",
             "MarigoldNormalsPipeline",
             "MochiPipeline",
             "MusicLDMPipeline",
@@ -845,6 +846,7 @@
             Lumina2Text2ImgPipeline,
             LuminaText2ImgPipeline,
             MarigoldDepthPipeline,
+            MarigoldIntrinsicsPipeline,
             MarigoldNormalsPipeline,
             MochiPipeline,
             MusicLDMPipeline,

docs/source/en/using-diffusers/marigold_usage.md

@@ -23,7 +23,9 @@
 from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_state_dict
 from ..utils import (
     USE_PEFT_BACKEND,
+    _get_detailed_type,
     _get_model_file,
+    _is_valid_type,
     is_accelerate_available,
     is_torch_version,
     is_transformers_available,
@@ -577,29 +579,36 @@ def LinearStrengthModel(start, finish, size):
         pipeline.set_ip_adapter_scale(ip_strengths)
         ```
         """
-        transformer = self.transformer
-        if not isinstance(scale, list):
-            scale = [[scale] * transformer.config.num_layers]
-        elif isinstance(scale, list) and isinstance(scale[0], int) or isinstance(scale[0], float):
-            if len(scale) != transformer.config.num_layers:
-                raise ValueError(f"Expected list of {transformer.config.num_layers} scales, got {len(scale)}.")
+
+        scale_type = Union[int, float]
+        num_ip_adapters = self.transformer.encoder_hid_proj.num_ip_adapters
+        num_layers = self.transformer.config.num_layers
+
+        # Single value for all layers of all IP-Adapters
+        if isinstance(scale, scale_type):
+            scale = [scale for _ in range(num_ip_adapters)]
+        # List of per-layer scales for a single IP-Adapter
+        elif _is_valid_type(scale, List[scale_type]) and num_ip_adapters == 1:
             scale = [scale]
+        # Invalid scale type
+        elif not _is_valid_type(scale, List[Union[scale_type, List[scale_type]]]):
+            raise TypeError(f"Unexpected type {_get_detailed_type(scale)} for scale.")
 
-        scale_configs = scale
+        if len(scale) != num_ip_adapters:
+            raise ValueError(f"Cannot assign {len(scale)} scales to {num_ip_adapters} IP-Adapters.")
 
-        key_id = 0
-        for attn_name, attn_processor in transformer.attn_processors.items():
-            if isinstance(attn_processor, (FluxIPAdapterJointAttnProcessor2_0)):
-                if len(scale_configs) != len(attn_processor.scale):
-                    raise ValueError(
-                        f"Cannot assign {len(scale_configs)} scale_configs to "
-                        f"{len(attn_processor.scale)} IP-Adapter."
-                    )
-                elif len(scale_configs) == 1:
-                    scale_configs = scale_configs * len(attn_processor.scale)
-                for i, scale_config in enumerate(scale_configs):
-                    attn_processor.scale[i] = scale_config[key_id]
-                key_id += 1
+        if any(len(s) != num_layers for s in scale if isinstance(s, list)):
+            invalid_scale_sizes = {len(s) for s in scale if isinstance(s, list)} - {num_layers}
+            raise ValueError(
+                f"Expected list of {num_layers} scales, got {', '.join(str(x) for x in invalid_scale_sizes)}."
+            )
+
+        # Scalars are transformed to lists with length num_layers
+        scale_configs = [[s] * num_layers if isinstance(s, scale_type) else s for s in scale]
+
+        # Set scales. zip over scale_configs prevents going into single transformer layers
+        for attn_processor, *scale in zip(self.transformer.attn_processors.values(), *scale_configs):
+            attn_processor.scale = scale
 
     def unload_ip_adapter(self):
         """

src/diffusers/__init__.py

@@ -2780,9 +2780,8 @@ def __call__(
 
             # IP-adapter
             ip_query = hidden_states_query_proj
-            ip_attn_output = None
-            # for ip-adapter
-            # TODO: support for multiple adapters
+            ip_attn_output = torch.zeros_like(hidden_states)
+
             for current_ip_hidden_states, scale, to_k_ip, to_v_ip in zip(
                 ip_hidden_states, self.scale, self.to_k_ip, self.to_v_ip
             ):
@@ -2793,12 +2792,14 @@ def __call__(
                 ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
                 # the output of sdp = (batch, num_heads, seq_len, head_dim)
                 # TODO: add support for attn.scale when we move to Torch 2.1
-                ip_attn_output = F.scaled_dot_product_attention(
+                current_ip_hidden_states = F.scaled_dot_product_attention(
                     ip_query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
                 )
-                ip_attn_output = ip_attn_output.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
-                ip_attn_output = scale * ip_attn_output
-                ip_attn_output = ip_attn_output.to(ip_query.dtype)
+                current_ip_hidden_states = current_ip_hidden_states.transpose(1, 2).reshape(
+                    batch_size, -1, attn.heads * head_dim
+                )
+                current_ip_hidden_states = current_ip_hidden_states.to(ip_query.dtype)
+                ip_attn_output += scale * current_ip_hidden_states
 
             return hidden_states, encoder_hidden_states, ip_attn_output
         else:

src/diffusers/loaders/ip_adapter.py

@@ -605,12 +605,13 @@ def forward(
         controlnet_cond: List[torch.Tensor],
         control_type: torch.Tensor,
         control_type_idx: List[int],
-        conditioning_scale: float = 1.0,
+        conditioning_scale: Union[float, List[float]] = 1.0,
         class_labels: Optional[torch.Tensor] = None,
         timestep_cond: Optional[torch.Tensor] = None,
         attention_mask: Optional[torch.Tensor] = None,
         added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
         cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        from_multi: bool = False,
         guess_mode: bool = False,
         return_dict: bool = True,
     ) -> Union[ControlNetOutput, Tuple[Tuple[torch.Tensor, ...], torch.Tensor]]:
@@ -647,6 +648,8 @@ def forward(
                 Additional conditions for the Stable Diffusion XL UNet.
             cross_attention_kwargs (`dict[str]`, *optional*, defaults to `None`):
                 A kwargs dictionary that if specified is passed along to the `AttnProcessor`.
+            from_multi (`bool`, defaults to `False`):
+                Use standard scaling when called from `MultiControlNetUnionModel`.
             guess_mode (`bool`, defaults to `False`):
                 In this mode, the ControlNet encoder tries its best to recognize the input content of the input even if
                 you remove all prompts. A `guidance_scale` between 3.0 and 5.0 is recommended.
@@ -658,6 +661,9 @@ def forward(
                 If `return_dict` is `True`, a [`~models.controlnet.ControlNetOutput`] is returned, otherwise a tuple is
                 returned where the first element is the sample tensor.
         """
+        if isinstance(conditioning_scale, float):
+            conditioning_scale = [conditioning_scale] * len(controlnet_cond)
+
         # check channel order
         channel_order = self.config.controlnet_conditioning_channel_order
 
@@ -742,12 +748,16 @@ def forward(
         inputs = []
         condition_list = []
 
-        for cond, control_idx in zip(controlnet_cond, control_type_idx):
+        for cond, control_idx, scale in zip(controlnet_cond, control_type_idx, conditioning_scale):
             condition = self.controlnet_cond_embedding(cond)
             feat_seq = torch.mean(condition, dim=(2, 3))
             feat_seq = feat_seq + self.task_embedding[control_idx]
-            inputs.append(feat_seq.unsqueeze(1))
-            condition_list.append(condition)
+            if from_multi:
+                inputs.append(feat_seq.unsqueeze(1))
+                condition_list.append(condition)
+            else:
+                inputs.append(feat_seq.unsqueeze(1) * scale)
+                condition_list.append(condition * scale)
 
         condition = sample
         feat_seq = torch.mean(condition, dim=(2, 3))
@@ -759,10 +769,13 @@ def forward(
             x = layer(x)
 
         controlnet_cond_fuser = sample * 0.0
-        for idx, condition in enumerate(condition_list[:-1]):
+        for (idx, condition), scale in zip(enumerate(condition_list[:-1]), conditioning_scale):
             alpha = self.spatial_ch_projs(x[:, idx])
             alpha = alpha.unsqueeze(-1).unsqueeze(-1)
-            controlnet_cond_fuser += condition + alpha
+            if from_multi:
+                controlnet_cond_fuser += condition + alpha
+            else:
+                controlnet_cond_fuser += condition + alpha * scale
 
         sample = sample + controlnet_cond_fuser
 
@@ -806,12 +819,13 @@ def forward(
         # 6. scaling
         if guess_mode and not self.config.global_pool_conditions:
             scales = torch.logspace(-1, 0, len(down_block_res_samples) + 1, device=sample.device)  # 0.1 to 1.0
-            scales = scales * conditioning_scale
+            if from_multi:
+                scales = scales * conditioning_scale[0]
             down_block_res_samples = [sample * scale for sample, scale in zip(down_block_res_samples, scales)]
             mid_block_res_sample = mid_block_res_sample * scales[-1]  # last one
-        else:
-            down_block_res_samples = [sample * conditioning_scale for sample in down_block_res_samples]
-            mid_block_res_sample = mid_block_res_sample * conditioning_scale
+        elif from_multi:
+            down_block_res_samples = [sample * conditioning_scale[0] for sample in down_block_res_samples]
+            mid_block_res_sample = mid_block_res_sample * conditioning_scale[0]
 
         if self.config.global_pool_conditions:
             down_block_res_samples = [

src/diffusers/models/attention_processor.py

@@ -47,9 +47,12 @@ def forward(
         guess_mode: bool = False,
         return_dict: bool = True,
     ) -> Union[ControlNetOutput, Tuple]:
+        down_block_res_samples, mid_block_res_sample = None, None
         for i, (image, ctype, ctype_idx, scale, controlnet) in enumerate(
             zip(controlnet_cond, control_type, control_type_idx, conditioning_scale, self.nets)
         ):
+            if scale == 0.0:
+                continue
             down_samples, mid_sample = controlnet(
                 sample=sample,
                 timestep=timestep,
@@ -63,12 +66,13 @@ def forward(
                 attention_mask=attention_mask,
                 added_cond_kwargs=added_cond_kwargs,
                 cross_attention_kwargs=cross_attention_kwargs,
+                from_multi=True,
                 guess_mode=guess_mode,
                 return_dict=return_dict,
             )
 
             # merge samples
-            if i == 0:
+            if down_block_res_samples is None and mid_block_res_sample is None:
                 down_block_res_samples, mid_block_res_sample = down_samples, mid_sample
             else:
                 down_block_res_samples = [