LCM Add Tests (huggingface#5707)

* lcm add tests

* uP

* Fix all

* uP

* Add

* all

* uP

* uP

* uP

* uP

* uP

* uP

* uP

Author: patrickvonplaten

loaders.py

@@ -1411,6 +1411,11 @@ def _best_guess_weight_name(cls, pretrained_model_name_or_path_or_dict, file_ext
             filter(lambda x: all(substring not in x for substring in unallowed_substrings), targeted_files)
         )
 
+        if any(f.endswith(LORA_WEIGHT_NAME) for f in targeted_files):
+            targeted_files = list(filter(lambda x: x.endswith(LORA_WEIGHT_NAME), targeted_files))
+        elif any(f.endswith(LORA_WEIGHT_NAME_SAFE) for f in targeted_files):
+            targeted_files = list(filter(lambda x: x.endswith(LORA_WEIGHT_NAME_SAFE), targeted_files))
+
         if len(targeted_files) > 1:
             raise ValueError(
                 f"Provided path contains more than one weights file in the {file_extension} format. Either specify `weight_name` in `load_lora_weights` or make sure there's only one  `.safetensors` or `.bin` file in  {pretrained_model_name_or_path_or_dict}."

pipelines/alt_diffusion/pipeline_alt_diffusion.py

@@ -588,6 +588,34 @@ def disable_freeu(self):
         """Disables the FreeU mechanism if enabled."""
         self.unet.disable_freeu()
 
+    def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
+        """
+        See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+        Args:
+            timesteps (`torch.Tensor`):
+                generate embedding vectors at these timesteps
+            embedding_dim (`int`, *optional*, defaults to 512):
+                dimension of the embeddings to generate
+            dtype:
+                data type of the generated embeddings
+
+        Returns:
+            `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
+        """
+        assert len(w.shape) == 1
+        w = w * 1000.0
+
+        half_dim = embedding_dim // 2
+        emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+        emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+        emb = w.to(dtype)[:, None] * emb[None, :]
+        emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+        if embedding_dim % 2 == 1:  # zero pad
+            emb = torch.nn.functional.pad(emb, (0, 1))
+        assert emb.shape == (w.shape[0], embedding_dim)
+        return emb
+
     @property
     def guidance_scale(self):
         return self._guidance_scale
@@ -605,7 +633,7 @@ def clip_skip(self):
     # corresponds to doing no classifier free guidance.
     @property
     def do_classifier_free_guidance(self):
-        return self._guidance_scale > 1
+        return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
 
     @property
     def cross_attention_kwargs(self):
@@ -804,6 +832,14 @@ def __call__(
         # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
         extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
 
+        # 6.5 Optionally get Guidance Scale Embedding
+        timestep_cond = None
+        if self.unet.config.time_cond_proj_dim is not None:
+            guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+            timestep_cond = self.get_guidance_scale_embedding(
+                guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+            ).to(device=device, dtype=latents.dtype)
+
         # 7. Denoising loop
         num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
         self._num_timesteps = len(timesteps)
@@ -818,6 +854,7 @@ def __call__(
                     latent_model_input,
                     t,
                     encoder_hidden_states=prompt_embeds,
+                    timestep_cond=timestep_cond,
                     cross_attention_kwargs=self.cross_attention_kwargs,
                     return_dict=False,
                 )[0]

pipelines/alt_diffusion/pipeline_alt_diffusion_img2img.py

@@ -646,6 +646,34 @@ def disable_freeu(self):
         """Disables the FreeU mechanism if enabled."""
         self.unet.disable_freeu()
 
+    def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
+        """
+        See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+        Args:
+            timesteps (`torch.Tensor`):
+                generate embedding vectors at these timesteps
+            embedding_dim (`int`, *optional*, defaults to 512):
+                dimension of the embeddings to generate
+            dtype:
+                data type of the generated embeddings
+
+        Returns:
+            `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
+        """
+        assert len(w.shape) == 1
+        w = w * 1000.0
+
+        half_dim = embedding_dim // 2
+        emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+        emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+        emb = w.to(dtype)[:, None] * emb[None, :]
+        emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+        if embedding_dim % 2 == 1:  # zero pad
+            emb = torch.nn.functional.pad(emb, (0, 1))
+        assert emb.shape == (w.shape[0], embedding_dim)
+        return emb
+
     @property
     def guidance_scale(self):
         return self._guidance_scale
@@ -659,7 +687,7 @@ def clip_skip(self):
     # corresponds to doing no classifier free guidance.
     @property
     def do_classifier_free_guidance(self):
-        return self._guidance_scale > 1
+        return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
 
     @property
     def cross_attention_kwargs(self):
@@ -849,6 +877,14 @@ def __call__(
         # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
         extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
 
+        # 7.5 Optionally get Guidance Scale Embedding
+        timestep_cond = None
+        if self.unet.config.time_cond_proj_dim is not None:
+            guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+            timestep_cond = self.get_guidance_scale_embedding(
+                guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+            ).to(device=device, dtype=latents.dtype)
+
         # 8. Denoising loop
         num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
         self._num_timesteps = len(timesteps)
@@ -863,6 +899,7 @@ def __call__(
                     latent_model_input,
                     t,
                     encoder_hidden_states=prompt_embeds,
+                    timestep_cond=timestep_cond,
                     cross_attention_kwargs=self.cross_attention_kwargs,
                     return_dict=False,
                 )[0]

pipelines/stable_diffusion/pipeline_stable_diffusion.py

@@ -576,6 +576,35 @@ def disable_freeu(self):
         """Disables the FreeU mechanism if enabled."""
         self.unet.disable_freeu()
 
+    # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
+    def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
+        """
+        See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+        Args:
+            timesteps (`torch.Tensor`):
+                generate embedding vectors at these timesteps
+            embedding_dim (`int`, *optional*, defaults to 512):
+                dimension of the embeddings to generate
+            dtype:
+                data type of the generated embeddings
+
+        Returns:
+            `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
+        """
+        assert len(w.shape) == 1
+        w = w * 1000.0
+
+        half_dim = embedding_dim // 2
+        emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+        emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+        emb = w.to(dtype)[:, None] * emb[None, :]
+        emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+        if embedding_dim % 2 == 1:  # zero pad
+            emb = torch.nn.functional.pad(emb, (0, 1))
+        assert emb.shape == (w.shape[0], embedding_dim)
+        return emb
+
     @property
     def guidance_scale(self):
         return self._guidance_scale
@@ -593,7 +622,7 @@ def clip_skip(self):
     # corresponds to doing no classifier free guidance.
     @property
     def do_classifier_free_guidance(self):
-        return self._guidance_scale > 1
+        return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
 
     @property
     def cross_attention_kwargs(self):
@@ -790,6 +819,14 @@ def __call__(
         # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
         extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
 
+        # 6.5 Optionally get Guidance Scale Embedding
+        timestep_cond = None
+        if self.unet.config.time_cond_proj_dim is not None:
+            guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+            timestep_cond = self.get_guidance_scale_embedding(
+                guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+            ).to(device=device, dtype=latents.dtype)
+
         # 7. Denoising loop
         num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
         self._num_timesteps = len(timesteps)
@@ -804,6 +841,7 @@ def __call__(
                     latent_model_input,
                     t,
                     encoder_hidden_states=prompt_embeds,
+                    timestep_cond=timestep_cond,
                     cross_attention_kwargs=self.cross_attention_kwargs,
                     return_dict=False,
                 )[0]

pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py

@@ -640,6 +640,35 @@ def disable_freeu(self):
         """Disables the FreeU mechanism if enabled."""
         self.unet.disable_freeu()
 
+    # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
+    def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
+        """
+        See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+        Args:
+            timesteps (`torch.Tensor`):
+                generate embedding vectors at these timesteps
+            embedding_dim (`int`, *optional*, defaults to 512):
+                dimension of the embeddings to generate
+            dtype:
+                data type of the generated embeddings
+
+        Returns:
+            `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
+        """
+        assert len(w.shape) == 1
+        w = w * 1000.0
+
+        half_dim = embedding_dim // 2
+        emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+        emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+        emb = w.to(dtype)[:, None] * emb[None, :]
+        emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+        if embedding_dim % 2 == 1:  # zero pad
+            emb = torch.nn.functional.pad(emb, (0, 1))
+        assert emb.shape == (w.shape[0], embedding_dim)
+        return emb
+
     @property
     def guidance_scale(self):
         return self._guidance_scale
@@ -653,7 +682,7 @@ def clip_skip(self):
     # corresponds to doing no classifier free guidance.
     @property
     def do_classifier_free_guidance(self):
-        return self._guidance_scale > 1
+        return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
 
     @property
     def cross_attention_kwargs(self):
@@ -841,6 +870,14 @@ def __call__(
         # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
         extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
 
+        # 7.5 Optionally get Guidance Scale Embedding
+        timestep_cond = None
+        if self.unet.config.time_cond_proj_dim is not None:
+            guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+            timestep_cond = self.get_guidance_scale_embedding(
+                guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+            ).to(device=device, dtype=latents.dtype)
+
         # 8. Denoising loop
         num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
         self._num_timesteps = len(timesteps)
@@ -855,6 +892,7 @@ def __call__(
                     latent_model_input,
                     t,
                     encoder_hidden_states=prompt_embeds,
+                    timestep_cond=timestep_cond,
                     cross_attention_kwargs=self.cross_attention_kwargs,
                     return_dict=False,
                 )[0]

pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py

@@ -765,6 +765,35 @@ def disable_freeu(self):
         """Disables the FreeU mechanism if enabled."""
         self.unet.disable_freeu()
 
+    # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
+    def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
+        """
+        See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+        Args:
+            timesteps (`torch.Tensor`):
+                generate embedding vectors at these timesteps
+            embedding_dim (`int`, *optional*, defaults to 512):
+                dimension of the embeddings to generate
+            dtype:
+                data type of the generated embeddings
+
+        Returns:
+            `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
+        """
+        assert len(w.shape) == 1
+        w = w * 1000.0
+
+        half_dim = embedding_dim // 2
+        emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+        emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+        emb = w.to(dtype)[:, None] * emb[None, :]
+        emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+        if embedding_dim % 2 == 1:  # zero pad
+            emb = torch.nn.functional.pad(emb, (0, 1))
+        assert emb.shape == (w.shape[0], embedding_dim)
+        return emb
+
     @property
     def guidance_scale(self):
         return self._guidance_scale
@@ -778,7 +807,7 @@ def clip_skip(self):
     # corresponds to doing no classifier free guidance.
     @property
     def do_classifier_free_guidance(self):
-        return self._guidance_scale > 1
+        return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
 
     @property
     def cross_attention_kwargs(self):
@@ -1087,6 +1116,14 @@ def __call__(
         # 9. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
         extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
 
+        # 9.5 Optionally get Guidance Scale Embedding
+        timestep_cond = None
+        if self.unet.config.time_cond_proj_dim is not None:
+            guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+            timestep_cond = self.get_guidance_scale_embedding(
+                guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+            ).to(device=device, dtype=latents.dtype)
+
         # 10. Denoising loop
         num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
         self._num_timesteps = len(timesteps)
@@ -1106,6 +1143,7 @@ def __call__(
                     latent_model_input,
                     t,
                     encoder_hidden_states=prompt_embeds,
+                    timestep_cond=timestep_cond,
                     cross_attention_kwargs=self.cross_attention_kwargs,
                     return_dict=False,
                 )[0]