Remove forced float64 from onnx stable diffusion pipelines (#11054)

* Update pipeline_onnx_stable_diffusion.py to remove float64

init_noise_sigma was being set as float64 before multiplying with latents, which changed latents into float64 too, which caused errors with onnxruntime since the latter wanted float16.

* Update pipeline_onnx_stable_diffusion_inpaint.py to remove float64

init_noise_sigma was being set as float64 before multiplying with latents, which changed latents into float64 too, which caused errors with onnxruntime since the latter wanted float16.

* Update pipeline_onnx_stable_diffusion_upscale.py to remove float64

init_noise_sigma was being set as float64 before multiplying with latents, which changed latents into float64 too, which caused errors with onnxruntime since the latter wanted float16.

* Update pipeline_onnx_stable_diffusion.py with comment for previous commit

Added comment on purpose of init_noise_sigma.  This comment exists in related scripts that use the same line of code, but it was missing here.

---------
Co-authored-by: YiYi Xu <yixu310@gmail.com>

src/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion.py

@@ -383,7 +383,8 @@ class OnnxStableDiffusionPipeline(DiffusionPipeline):
         # set timesteps
         self.scheduler.set_timesteps(num_inference_steps)
 
-        latents = latents * np.float64(self.scheduler.init_noise_sigma)
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
 
         # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
         # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.

src/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_inpaint.py

@@ -483,7 +483,7 @@ class OnnxStableDiffusionInpaintPipeline(DiffusionPipeline):
         self.scheduler.set_timesteps(num_inference_steps)
 
         # scale the initial noise by the standard deviation required by the scheduler
-        latents = latents * np.float64(self.scheduler.init_noise_sigma)
+        latents = latents * self.scheduler.init_noise_sigma
 
         # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
         # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.

src/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_upscale.py

@@ -481,7 +481,7 @@ class OnnxStableDiffusionUpscalePipeline(DiffusionPipeline):
         timesteps = self.scheduler.timesteps
 
         # Scale the initial noise by the standard deviation required by the scheduler
-        latents = latents * np.float64(self.scheduler.init_noise_sigma)
+        latents = latents * self.scheduler.init_noise_sigma
 
         # 5. Add noise to image
         noise_level = np.array([noise_level]).astype(np.int64)