Fix latents.dtype before vae.decode() at ROCm devices in StableDiffusionPipelines

src/diffusers/pipelines/controlnet/pipeline_controlnet.py

@@ -562,6 +562,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16

src/diffusers/pipelines/controlnet/pipeline_controlnet_img2img.py

@@ -540,6 +540,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16

src/diffusers/pipelines/controlnet/pipeline_controlnet_inpaint.py

@@ -542,6 +542,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16

src/diffusers/pipelines/controlnet_xs/pipeline_controlnet_xs.py

@@ -422,6 +422,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16

src/diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_cycle_diffusion.py

@@ -553,6 +553,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16

src/diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_inpaint_legacy.py

@@ -454,6 +454,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16

src/diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_model_editing.py

@@ -392,6 +392,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16

src/diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_pix2pix_zero.py

@@ -606,6 +606,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16

src/diffusers/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion_dual_guided.py

@@ -305,6 +305,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16

src/diffusers/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion_image_variation.py

@@ -165,6 +165,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16

src/diffusers/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion_text_to_image.py

@@ -222,6 +222,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16

src/diffusers/pipelines/ledits_pp/pipeline_leditspp_stable_diffusion.py

@@ -406,6 +406,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16

src/diffusers/pipelines/paint_by_example/pipeline_paint_by_example.py

@@ -252,6 +252,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16

src/diffusers/pipelines/semantic_stable_diffusion/pipeline_semantic_stable_diffusion.py

@@ -112,6 +112,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py

@@ -577,6 +577,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
@@ -1048,6 +1050,8 @@ def __call__(
                     xm.mark_step()
 
         if not output_type == "latent":
+            if torch.cuda.is_available() and torch.version.hip:
+                latents = latents.to(dtype=self.vae.dtype)
             image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[
                 0
             ]

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_depth2img.py

@@ -385,6 +385,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_image_variation.py

@@ -176,6 +176,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py

@@ -616,6 +616,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_instruct_pix2pix.py

@@ -751,6 +751,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_latent_upscale.py

@@ -281,6 +281,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_upscale.py

@@ -410,6 +410,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16

src/diffusers/pipelines/stable_diffusion/pipeline_stable_unclip.py

@@ -473,6 +473,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16

src/diffusers/pipelines/stable_diffusion/pipeline_stable_unclip_img2img.py

@@ -435,6 +435,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16

src/diffusers/pipelines/stable_diffusion_attend_and_excite/pipeline_stable_diffusion_attend_and_excite.py

@@ -483,6 +483,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16

src/diffusers/pipelines/stable_diffusion_diffedit/pipeline_stable_diffusion_diffedit.py

@@ -626,6 +626,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16

src/diffusers/pipelines/stable_diffusion_k_diffusion/pipeline_stable_diffusion_k_diffusion.py

@@ -383,6 +383,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16

src/diffusers/pipelines/stable_diffusion_panorama/pipeline_stable_diffusion_panorama.py

@@ -530,6 +530,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16

src/diffusers/pipelines/stable_diffusion_safe/pipeline_stable_diffusion_safe.py

@@ -337,6 +337,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16

src/diffusers/pipelines/stable_diffusion_sag/pipeline_stable_diffusion_sag.py

@@ -457,6 +457,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16

src/diffusers/pipelines/t2i_adapter/pipeline_stable_diffusion_adapter.py

@@ -500,6 +500,8 @@ def decode_latents(self, latents):
         deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
 
         latents = 1 / self.vae.config.scaling_factor * latents
+        if torch.cuda.is_available() and torch.version.hip:
+            latents = latents.to(dtype=self.vae.dtype)
         image = self.vae.decode(latents, return_dict=False)[0]
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16