dont hardcode the generator device

app.py

@@ -181,8 +181,8 @@ def greet(audio, steps=25, scheduler="ddpm"):
         token_embeds = model.text_encoder.get_input_embeddings().weight.data
 
     token_embeds[model.placeholder_token_id] = audio_token.clone()
-    g_gpu = torch.Generator(device='cuda')
-    g_gpu.manual_seed(23229249375547) # no reason this can't be input by the user!
+    generator = torch.Generator(device=device)
+    generator.manual_seed(23229249375547) # no reason this can't be input by the user!
     pipeline = StableDiffusionPipeline.from_pretrained(
         pretrained_model_name_or_path=model.repo_id,
         tokenizer=model.tokenizer,
@@ -194,10 +194,8 @@ def greet(audio, steps=25, scheduler="ddpm"):
     ).to(device)
     pipeline.enable_xformers_memory_efficient_attention()
 
-    # pipeline.scheduler = EulerAncestralDiscreteScheduler.from_config(pipeline.scheduler.config)
-    # pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config)
-    print(f"taking {steps} steps using the {scheduler} scheduler")
-    image = pipeline(prompt, num_inference_steps=steps, guidance_scale=8.5, generator=g_gpu).images[0]
+    # print(f"taking {steps} steps using the {scheduler} scheduler")
+    image = pipeline(prompt, num_inference_steps=steps, guidance_scale=8.5, generator=generator).images[0]
     return image