[Speculative Decoding] Add speculative decoding

run_bug.py

@@ -0,0 +1,32 @@
+#!/usr/bin/env python3
+import torch
+
+from diffusers import AutoencoderKL, StableDiffusionXLPipeline, StableDiffusionXLImg2ImgPipeline, KDPM2AncestralDiscreteScheduler
+
+vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix",
+    torch_dtype=torch.float16
+)
+
+base = StableDiffusionXLPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0",
+    vae=vae,
+    torch_dtype=torch.float16,
+    variant="fp16",
+    use_safetensors=True
+)
+
+scheduler = KDPM2AncestralDiscreteScheduler.from_config(base.scheduler.config, use_karras_sigmas=True)
+base.scheduler = scheduler
+
+base.to("cuda")
+
+def print_step(s, t, latents):
+    print(s)
+
+generator=torch.manual_seed(1111)
+
+images = base(
+    prompt="LOVE",
+    num_inference_steps=10,
+    generator=generator,
+    callback=print_step
+).images

run_lora_on_off.py

@@ -0,0 +1,35 @@
+#!/usr/bin/env python3
+#@title Fuse/unfuse LoRAs sequentially leading to trouble
+import torch
+from diffusers import StableDiffusionXLPipeline
+
+pipe = StableDiffusionXLPipeline.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0",
+    torch_dtype=torch.float16,
+    variant="fp16",
+).to("cuda")
+
+pipe.load_lora_weights("Pclanglais/TintinIA")
+pipe.fuse_lora()
+images = pipe("a mecha robot", num_inference_steps=2)
+pipe.unfuse_lora()
+pipe.unload_lora_weights()
+
+pipe.load_lora_weights("ProomptEngineer/pe-balloon-diffusion-style")
+pipe.fuse_lora()
+images = pipe("a mecha robot", num_inference_steps=2)
+pipe.unfuse_lora()
+pipe.unload_lora_weights()
+
+pipe.load_lora_weights("ostris/crayon_style_lora_sdxl")
+pipe.fuse_lora()
+images = pipe("a mecha robot", num_inference_steps=2)
+pipe.unfuse_lora()
+pipe.unload_lora_weights()
+
+
+pipe.load_lora_weights("joachimsallstrom/aether-cloud-lora-for-sdxl")
+pipe.fuse_lora()
+images = pipe("a mecha robot", num_inference_steps=2)
+pipe.unfuse_lora()
+pipe.unload_lora_weights()

run_sd_lora.py

@@ -0,0 +1,20 @@
+#!/usr/bin/env python3
+from diffusers import DiffusionPipeline, AutoencoderKL
+import hf_image_uploader as hiu
+import torch
+
+vae = AutoencoderKL.from_pretrained(
+    "madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16
+)
+pipe = DiffusionPipeline.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0",
+    variant="fp16",
+    vae=vae,
+    torch_dtype=torch.float16,
+).to("cuda")
+
+pipe.load_lora_weights("rvorias/m_test")
+# pipe.enable_xformers_memory_efficient_attention()
+
+image = pipe("a photo of a pikachu pixel art", generator=torch.manual_seed(66)).images[0]
+hiu.upload(image, "patrickvonplaten/images")

run_sd_sde.py

@@ -0,0 +1,22 @@
+#!/usr/bin/env python3
+from diffusers import DPMSolverMultistepScheduler, StableDiffusionXLPipeline, DPMSolverSDEScheduler
+
+path = "hf-internal-testing/tiny-stable-diffusion-xl-pipe"
+
+pipe = StableDiffusionXLPipeline.from_pretrained(path)
+pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config, algorithm_type="sde-dpmsolver++")
+
+prompt = "An astronaut riding a green horse on Mars"
+steps = 20
+
+for i in range(2):
+    width = 32 * (i + 1)
+    height = 32 * (i + 1)
+    image = pipe(prompt=prompt, width=width, height=height, num_inference_steps=steps).images[0]
+
+pipe.scheduler = DPMSolverSDEScheduler.from_config(pipe.scheduler.config, algorithm_type="sde-dpmsolver++")
+
+for i in range(2):
+    width = 32 * (i + 1)
+    height = 32 * (i + 1)
+    image = pipe(prompt=prompt, width=width, height=height, num_inference_steps=steps).images[0]

run_sdxl_cpu.py

@@ -0,0 +1,25 @@
+#!/usr/bin/env python3
+from diffusers import DPMSolverMultistepScheduler, StableDiffusionXLPipeline, DPMSolverSDEScheduler
+import hf_image_uploader as hiu
+
+path = "hf-internal-testing/tiny-stable-diffusion-xl-pipe"
+
+pipe = StableDiffusionXLPipeline.from_pretrained(path)
+pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config, algorithm_type="sde-dpmsolver++")
+
+prompt = "An astronaut riding a green horse on Mars"
+steps = 20
+
+for i in range(2):
+    width = 32 * (i + 1)
+    height = 32 * (i + 1)
+    image = pipe(prompt=prompt, width=width, height=height, num_inference_steps=steps).images[0]
+    hiu.upload(image, "patrickvonplaten/images")
+
+pipe.scheduler = DPMSolverSDEScheduler.from_config(pipe.scheduler.config, algorithm_type="sde-dpmsolver++")
+
+for i in range(2):
+    width = 32 * (i + 1)
+    height = 32 * (i + 1)
+    image = pipe(prompt=prompt, width=width, height=height, num_inference_steps=steps).images[0]
+    hiu.upload(image, "patrickvonplaten/images")