Add TPU example for Ray Serve with Stable Diffusion

kevin85421 · web-flow · commit d16752eef753 · 2024-06-24T15:47:43.000-07:00
diff --git a/stable_diffusion/diffusion_results.png b/stable_diffusion/diffusion_results.png
diff --git a/stable_diffusion/stable_diffusion_tpu.py b/stable_diffusion/stable_diffusion_tpu.py
@@ -0,0 +1,144 @@
+"""Ray Serve Stable Diffusion example."""
+from io import BytesIO
+from typing import List
+from fastapi import FastAPI
+from fastapi.responses import Response
+import logging
+import ray
+from ray import serve
+import time
+
+app = FastAPI()
+_MAX_BATCH_SIZE = 64
+
+logger = logging.getLogger("ray.serve")
+
+@serve.deployment(num_replicas=1)
+@serve.ingress(app)
+class APIIngress:
+    def __init__(self, diffusion_model_handle) -> None:
+        self.handle = diffusion_model_handle
+
+    @app.get(
+        "/imagine",
+        responses={200: {"content": {"image/png": {}}}},
+        response_class=Response,
+    )
+    async def generate(self, prompt: str):
+        assert len(prompt), "prompt parameter cannot be empty"
+
+        image = await self.handle.generate.remote(prompt)
+        return image
+
+
+@serve.deployment(
+    ray_actor_options={
+        "resources": {"TPU": 4},
+    },
+)
+class StableDiffusion:
+  """FLAX Stable Diffusion Ray Serve deployment running on TPUs.
+
+  Attributes:
+    run_with_profiler: Whether or not to run with the profiler. Note that
+      this saves the profile to the separate TPU VM.
+
+  """
+
+  def __init__(
+      self, run_with_profiler: bool = False, warmup: bool = False,
+      warmup_batch_size: int = _MAX_BATCH_SIZE):
+    from diffusers import FlaxStableDiffusionPipeline
+    from flax.jax_utils import replicate
+    import jax 
+    import jax.numpy as jnp
+    from jax import pmap
+
+    model_id = "CompVis/stable-diffusion-v1-4"
+
+    self._pipeline, params = FlaxStableDiffusionPipeline.from_pretrained(
+        model_id,
+        revision="bf16",
+        dtype=jnp.bfloat16)
+
+    self._p_params = replicate(params)
+    self._p_generate = pmap(self._pipeline._generate)
+    self._run_with_profiler = run_with_profiler
+    self._profiler_dir = "/tmp/tensorboard"
+
+    if warmup:
+      logger.info("Sending warmup requests.")
+      warmup_prompts = ["A warmup request"] * warmup_batch_size
+      self.generate_tpu(warmup_prompts)
+
+  def generate_tpu(self, prompts: List[str]):
+    """Generates a batch of images from Diffusion from a list of prompts.
+
+    Args:
+      prompts: a list of strings. Should be a factor of 4.
+
+    Returns:
+      A list of PIL Images.
+    """
+    from flax.training.common_utils import shard
+    import jax
+    import numpy as np
+
+    rng = jax.random.PRNGKey(0)
+    rng = jax.random.split(rng, jax.device_count())
+
+    assert prompts, "prompt parameter cannot be empty"
+    logger.info("Prompts: %s", prompts)
+    prompt_ids = self._pipeline.prepare_inputs(prompts)
+    prompt_ids = shard(prompt_ids)
+    logger.info("Sharded prompt ids has shape: %s", prompt_ids.shape)
+    if self._run_with_profiler:
+      jax.profiler.start_trace(self._profiler_dir)
+
+    time_start = time.time()
+    images = self._p_generate(prompt_ids, self._p_params, rng)
+    images = images.block_until_ready()
+    elapsed = time.time() - time_start
+    if self._run_with_profiler:
+      jax.profiler.stop_trace()
+
+    logger.info("Inference time (in seconds): %f", elapsed)
+    logger.info("Shape of the predictions: %s", images.shape)
+    images = images.reshape(
+        (images.shape[0] * images.shape[1],) + images.shape[-3:])
+    logger.info("Shape of images afterwards: %s", images.shape)
+    return self._pipeline.numpy_to_pil(np.array(images))
+
+  @serve.batch(batch_wait_timeout_s=10, max_batch_size=_MAX_BATCH_SIZE)
+  async def batched_generate_handler(self, prompts: List[str]):
+    """Sends a batch of prompts to the TPU model server.
+
+    This takes advantage of @serve.batch, Ray Serve's built-in batching
+    mechanism.
+
+    Args:
+      prompts: A list of input prompts
+
+    Returns:
+      A list of responses which contents are raw PNG.
+    """
+    logger.info("Number of input prompts: %d", len(prompts))
+    num_to_pad = _MAX_BATCH_SIZE - len(prompts)
+    prompts += ["Scratch request"] * num_to_pad
+
+    images = self.generate_tpu(prompts)
+    results = []
+    for image in images[: _MAX_BATCH_SIZE - num_to_pad]:
+      file_stream = BytesIO()
+      image.save(file_stream, "PNG")
+      results.append(
+          Response(content=file_stream.getvalue(), media_type="image/png")
+      )
+    return results
+
+  async def generate(self, prompt):
+    return await self.batched_generate_handler(prompt)
+
+
+diffusion_bound = StableDiffusion.bind()
+deployment = APIIngress.bind(diffusion_bound)
diff --git a/stable_diffusion/stable_diffusion_tpu_req.py b/stable_diffusion/stable_diffusion_tpu_req.py
@@ -0,0 +1,132 @@
+import argparse
+from concurrent import futures
+import functools
+from io import BytesIO
+import numpy as np
+from PIL import Image
+import requests
+from tqdm import tqdm
+
+
+_PROMPTS = [
+    "Labrador in the style of Hokusai",
+    "Painting of a squirrel skating in New York",
+    "HAL-9000 in the style of Van Gogh",
+    "Times Square under water, with fish and a dolphin swimming around",
+    "Ancient Roman fresco showing a man working on his laptop",
+    "Armchair in the shape of an avocado",
+    "Clown astronaut in space, with Earth in the background",
+    "A cat sitting on a windowsill",
+    "A dog playing fetch in a park",
+    "A city skyline at night",
+    "A field of flowers in bloom",
+    "A tropical beach with palm trees",
+    "A snowy mountain range",
+    "A waterfall cascading into a pool",
+    "A forest at sunset",
+    "A desert landscape with cacti",
+    "A volcano erupting",
+    "A lightning storm in the distance",
+    "A rainbow over a rainbow",
+    "A unicorn grazing in a meadow",
+    "A dragon flying through the sky",
+    "A mermaid swimming in the ocean",
+    "A robot walking down the street",
+    "A UFO landing in a field",
+    "A portal to another dimension",
+    "A time traveler from the future",
+    "A talking cat",
+    "A bowl of fruit on a table",
+    "A group of friends laughing",
+    "A family sitting down for dinner",
+    "A couple kissing in the rain",
+    "A child playing with a toy",
+    "A musician playing an instrument",
+    "A painter painting a picture",
+    "A writer writing a book",
+    "A scientist conducting an experiment",
+    "A construction worker building a house",
+    "A doctor operating on a patient",
+    "A teacher teaching a class",
+    "A police officer arresting a suspect",
+    "A firefighter putting out a fire",
+    "A soldier fighting in a war",
+    "A farmer working in a field",
+    "A pilot flying a plane",
+    "An astronaut in space",
+    "A unicorn eating a rainbow"
+]
+
+
+def send_request_and_receive_image(prompt: str, url: str) -> BytesIO:
+  """Sends a single prompt request and returns the Image."""
+  try:
+    inputs = "%20".join(prompt.split(" "))
+    resp = requests.get(f"{url}?prompt={inputs}")
+    resp.raise_for_status()
+    return BytesIO(resp.content)
+  except requests.RequestException as e:
+    print(f"An error occurred while sending the request: {e}")
+
+
+def image_grid(imgs, rows, cols):
+  w, h = imgs[0].size
+  grid = Image.new("RGB", size=(cols * w, rows * h))
+  for i, img in enumerate(imgs):
+    grid.paste(img, box=(i % cols * w, i // cols * h))
+  return grid
+
+
+def send_requests(num_requests: int, batch_size: int, save_pictures: bool,
+                  url: str = "http://localhost:8000/imagine"):
+  """Sends a list of requests and processes the responses."""
+  print("num_requests: ", num_requests)
+  print("batch_size: ", batch_size)
+  print("url: ", url)
+  print("save_pictures: ", save_pictures)
+
+  prompts = _PROMPTS
+  if num_requests > len(_PROMPTS):
+    # Repeat until larger than num_requests
+    prompts = _PROMPTS * int(np.ceil(num_requests / len(_PROMPTS)))
+
+  prompts = np.random.choice(
+      prompts, num_requests, replace=False)
+
+  with futures.ThreadPoolExecutor(max_workers=batch_size) as executor:
+    raw_images = list(
+        tqdm(
+            executor.map(
+                functools.partial(send_request_and_receive_image, url=url),
+                prompts,
+            ),
+            total=len(prompts),
+        )
+    )
+
+  if save_pictures:
+    print("Saving pictures to diffusion_results.png")
+    images = [Image.open(raw_image) for raw_image in raw_images]
+    grid = image_grid(images, 2, num_requests // 2)
+    grid.save("./diffusion_results.png")
+
+
+if __name__ == "__main__":
+  parser = argparse.ArgumentParser(description="Sends requests to Diffusion.")
+  parser.add_argument(
+      "--num_requests", help="Number of requests to send.",
+      default=8)
+  parser.add_argument(
+      "--batch_size", help="The number of requests to send at a time.",
+      default=8)
+  parser.add_argument(
+      "--save_pictures", default=False, action="store_true",
+      help="Whether to save the generated pictures to disk.")
+  parser.add_argument(
+      "--ip", help="The IP address to send the requests to.")
+
+  args = parser.parse_args()
+
+  send_requests(
+      num_requests=int(args.num_requests), batch_size=int(args.batch_size),
+      save_pictures=bool(args.save_pictures))
