Merge branch 'main' of https://huggingface.co/diffusers/tools into main

plot_bench.py

@@ -0,0 +1,72 @@
+#!/usr/bin/env python3
+import matplotlib.pyplot as plt
+import numpy as np
+import seaborn as sns
+
+sns.set(font_scale=1.1)
+
+batch_sizes = {
+    "fp16_4": {
+       "A100": [4.75, 3.26, 3.24, 3.10],  # those values are made up
+       "A10": [13.94, 9.81, 10.01, 9.35],
+       "T4": [38.81, 30.09, 29.74, 27.55],
+       "V100": [9.84, 8.16, 8.09, 7.65],
+       "3090": [10.04, 7.82, 7.89, 7.47],
+       "3090TI": [9.07, 7.14, 7.15, 6.81],
+    },
+    "fp16_16": {
+       "A100": [18.95, 13.57, 13.67, 12.25],
+       "A10": [0, 37.55, 38.31, 36.81],
+       "T4": [0, 111.47, 113.26, 106.93],
+       "V100": [0, 30.29, 29.84, 28.22],
+       "3090": [0, 29.06, 29.06, 28.2],
+       "3090TI": [0, 26.1, 26.28, 25.46],
+    },
+    "fp32_4": {
+       "A100": [16.56, 12.42, 12.2, 11.84],
+       "A10": [34.77, 27.63, 22.77, 22.07],
+       "T4": [0, 85.72, 85.78, 84.48],
+       "V100": [0, 25.73, 25.31, 24.7],
+       "3090": [22.69, 21.45, 18.67, 18.09],
+       "3090TI": [20.32, 19.31, 16.9, 16.37],
+    },
+    "fp32_16": {
+       "A100": [0, 47.08, 46.27, 44.8],
+       "A10": [0, 116.49, 88.56, 86.64],
+       "T4": [0, 276.47, 280.26, 270.93],  # numbers are made up
+       "V100": [0, 84.99, 84.73, 82.55],
+       "3090": [0, 85.35, 72.37, 70.25],
+       "3090TI": [0, 75.37, 65.25, 64.32],
+    },
+}
+
+batch_size = 16
+dtype = "fp32"
+
+key = f"{dtype}_{batch_size}"
+
+methods = {
+    "Vanilla Attention": [x[0] for x in batch_sizes[key].values()],
+    "xFormers": [x[1] for x in batch_sizes[key].values()],
+    "PyTorch2.0 SDPA": [x[2] for x in batch_sizes[key].values()],
+    "SDPA + torch.compile": [x[3] for x in batch_sizes[key].values()],
+}
+
+x = np.arange(len(batch_sizes[key]))  # the label locations
+width = 0.1  # the width of the bars
+multiplier = 0
+
+fig, ax = plt.subplots(constrained_layout=True)
+
+for attribute, measurement in methods.items():
+    offset = width * multiplier
+    rects = ax.bar(x + offset, measurement, width, label=attribute)
+    multiplier += 1
+
+# Add some text for labels, title and custom x-axis tick labels, etc.
+ax.set_ylabel('Time (s)')
+ax.set_title(f'Inference Speed at Batch Size={batch_size} for {dtype}')
+ax.set_xticks(x + width, batch_sizes[key])
+ax.legend(loc='upper left')
+
+plt.show()

run_local.py

@@ -4,6 +4,7 @@ from diffusers.schedulers.scheduling_unipc_multistep import UniPCMultistepSchedu
 import time
 import os
 from huggingface_hub import HfApi
+from compel import Compel
 import torch
 import sys
 from pathlib import Path
@@ -16,21 +17,31 @@ start_time = time.time()
 #pipe.scheduler = HeunDiscreteScheduler.from_config(pipe.scheduler.config)
 pipe = StableDiffusionPipeline.from_pretrained(path, torch_dtype=torch.float16)
 pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
+
+compel = Compel(tokenizer=pipe.tokenizer, text_encoder=pipe.text_encoder)
+
+
 pipe = pipe.to("cuda")
 
 prompt = "a highly realistic photo of green turtle"
 
-generator = torch.Generator(device="cuda").manual_seed(0)
-image = pipe(prompt, generator=generator, num_inference_steps=15).images[0]
+prompts = ["a cat playing with a ball++ in the forest", "a cat playing with a ball++ in the forest", "a cat playing with a ball-- in the forest"]
+
+prompt_embeds = torch.cat([compel.build_conditioning_tensor(prompt) for prompt in prompts])
+
+generator = [torch.Generator(device="cuda").manual_seed(0) for _ in range(prompt_embeds.shape[0])]
+images = pipe(prompt_embeds=prompt_embeds, generator=generator, num_inference_steps=15).images
+
 print("Time", time.time() - start_time)
 
-path = os.path.join(Path.home(), "images", "aa.png")
-image.save(path)
+for i, image in enumerate(image):
+    path = os.path.join(Path.home(), "images", f"aa_{i}.png")
+    image.save(path)
 
-api.upload_file(
-    path_or_fileobj=path,
-    path_in_repo=path.split("/")[-1],
-    repo_id="patrickvonplaten/images",
-    repo_type="dataset",
-)
-print("https://huggingface.co/datasets/patrickvonplaten/images/blob/main/aa.png")
+    api.upload_file(
+        path_or_fileobj=path,
+        path_in_repo=path.split("/")[-1],
+        repo_id="patrickvonplaten/images",
+        repo_type="dataset",
+    )
+    print(f"https://huggingface.co/datasets/patrickvonplaten/images/blob/main/aa_{i}.png")