can call encode_prompt with out setting a text encoder instance variable (#4396)

* can call encode_prompt with out setting a text encoder instance variable

* fix

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_upscale.py

@@ -268,7 +268,14 @@ class StableDiffusionUpscalePipeline(DiffusionPipeline, TextualInversionLoaderMi
             )
             prompt_embeds = prompt_embeds[0]
 
-        prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+        if self.text_encoder is not None:
+            prompt_embeds_dtype = self.text_encoder.dtype
+        elif self.unet is not None:
+            prompt_embeds_dtype = self.unet.dtype
+        else:
+            prompt_embeds_dtype = prompt_embeds.dtype
+
+        prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
 
         bs_embed, seq_len, _ = prompt_embeds.shape
         # duplicate text embeddings for each generation per prompt, using mps friendly method
@@ -324,7 +331,7 @@ class StableDiffusionUpscalePipeline(DiffusionPipeline, TextualInversionLoaderMi
             # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
             seq_len = negative_prompt_embeds.shape[1]
 
-            negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+            negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
 
             negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
             negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)

src/diffusers/pipelines/stable_diffusion/pipeline_stable_unclip.py

@@ -371,7 +371,14 @@ class StableUnCLIPPipeline(DiffusionPipeline, TextualInversionLoaderMixin, LoraL
             )
             prompt_embeds = prompt_embeds[0]
 
-        prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+        if self.text_encoder is not None:
+            prompt_embeds_dtype = self.text_encoder.dtype
+        elif self.unet is not None:
+            prompt_embeds_dtype = self.unet.dtype
+        else:
+            prompt_embeds_dtype = prompt_embeds.dtype
+
+        prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
 
         bs_embed, seq_len, _ = prompt_embeds.shape
         # duplicate text embeddings for each generation per prompt, using mps friendly method
@@ -427,7 +434,7 @@ class StableUnCLIPPipeline(DiffusionPipeline, TextualInversionLoaderMixin, LoraL
             # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
             seq_len = negative_prompt_embeds.shape[1]
 
-            negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+            negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
 
             negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
             negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)

src/diffusers/pipelines/stable_diffusion/pipeline_stable_unclip_img2img.py

@@ -272,7 +272,14 @@ class StableUnCLIPImg2ImgPipeline(DiffusionPipeline, TextualInversionLoaderMixin
             )
             prompt_embeds = prompt_embeds[0]
 
-        prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+        if self.text_encoder is not None:
+            prompt_embeds_dtype = self.text_encoder.dtype
+        elif self.unet is not None:
+            prompt_embeds_dtype = self.unet.dtype
+        else:
+            prompt_embeds_dtype = prompt_embeds.dtype
+
+        prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
 
         bs_embed, seq_len, _ = prompt_embeds.shape
         # duplicate text embeddings for each generation per prompt, using mps friendly method
@@ -328,7 +335,7 @@ class StableUnCLIPImg2ImgPipeline(DiffusionPipeline, TextualInversionLoaderMixin
             # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
             seq_len = negative_prompt_embeds.shape[1]
 
-            negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+            negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
 
             negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
             negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)

src/diffusers/pipelines/t2i_adapter/pipeline_stable_diffusion_adapter.py

@@ -331,7 +331,14 @@ class StableDiffusionAdapterPipeline(DiffusionPipeline):
             )
             prompt_embeds = prompt_embeds[0]
 
-        prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+        if self.text_encoder is not None:
+            prompt_embeds_dtype = self.text_encoder.dtype
+        elif self.unet is not None:
+            prompt_embeds_dtype = self.unet.dtype
+        else:
+            prompt_embeds_dtype = prompt_embeds.dtype
+
+        prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
 
         bs_embed, seq_len, _ = prompt_embeds.shape
         # duplicate text embeddings for each generation per prompt, using mps friendly method
@@ -387,7 +394,7 @@ class StableDiffusionAdapterPipeline(DiffusionPipeline):
             # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
             seq_len = negative_prompt_embeds.shape[1]
 
-            negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+            negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
 
             negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
             negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)

src/diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_synth.py

@@ -258,7 +258,14 @@ class TextToVideoSDPipeline(DiffusionPipeline, TextualInversionLoaderMixin, Lora
             )
             prompt_embeds = prompt_embeds[0]
 
-        prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+        if self.text_encoder is not None:
+            prompt_embeds_dtype = self.text_encoder.dtype
+        elif self.unet is not None:
+            prompt_embeds_dtype = self.unet.dtype
+        else:
+            prompt_embeds_dtype = prompt_embeds.dtype
+
+        prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
 
         bs_embed, seq_len, _ = prompt_embeds.shape
         # duplicate text embeddings for each generation per prompt, using mps friendly method
@@ -314,7 +321,7 @@ class TextToVideoSDPipeline(DiffusionPipeline, TextualInversionLoaderMixin, Lora
             # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
             seq_len = negative_prompt_embeds.shape[1]
 
-            negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+            negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
 
             negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
             negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)

src/diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_synth_img2img.py

@@ -320,7 +320,14 @@ class VideoToVideoSDPipeline(DiffusionPipeline, TextualInversionLoaderMixin, Lor
             )
             prompt_embeds = prompt_embeds[0]
 
-        prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+        if self.text_encoder is not None:
+            prompt_embeds_dtype = self.text_encoder.dtype
+        elif self.unet is not None:
+            prompt_embeds_dtype = self.unet.dtype
+        else:
+            prompt_embeds_dtype = prompt_embeds.dtype
+
+        prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
 
         bs_embed, seq_len, _ = prompt_embeds.shape
         # duplicate text embeddings for each generation per prompt, using mps friendly method
@@ -376,7 +383,7 @@ class VideoToVideoSDPipeline(DiffusionPipeline, TextualInversionLoaderMixin, Lor
             # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
             seq_len = negative_prompt_embeds.shape[1]
 
-            negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+            negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
 
             negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
             negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)