File size: 2,701 Bytes
9b00d20 15dbd5a 9b00d20 15dbd5a ef06ce9 9b00d20 3d28980 15dbd5a 9b00d20 15dbd5a 9b00d20 15dbd5a 9b00d20 fbdce04 d07745d fbdce04 9b00d20 15dbd5a fbdce04 15dbd5a 9b00d20 d07745d |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 |
from typing import cast, Union
import torch
from diffusers import AutoencoderKLHunyuanVideo
from diffusers.video_processor import VideoProcessor
from diffusers.utils import export_to_video
class EndpointHandler:
def __init__(self, path=""):
self.device = "cuda"
self.dtype = torch.float16
self.vae = cast(AutoencoderKLHunyuanVideo, AutoencoderKLHunyuanVideo.from_pretrained(path, torch_dtype=self.dtype).to(self.device, self.dtype).eval())
self.vae.enable_tiling()
self.vae_scale_factor = self.vae_scale_factor_spatial = self.vae.spatial_compression_ratio
self.video_processor = VideoProcessor(
vae_scale_factor=self.vae_scale_factor_spatial
)
@torch.no_grad()
def __call__(self, data) -> Union[torch.Tensor, bytes]:
"""
Args:
data (:obj:):
includes the input data and the parameters for the inference.
"""
tensor = cast(torch.Tensor, data["inputs"])
parameters = cast(dict, data.get("parameters", {}))
do_scaling = cast(bool, parameters.get("do_scaling", True))
scaling_factor = cast(float, parameters.get("scaling_factor", None))
if scaling_factor is not None:
scaling_factor = float(scaling_factor)
if do_scaling and scaling_factor is None:
scaling_factor = self.vae.config.scaling_factor
output_type = cast(str, parameters.get("output_type", "pil"))
partial_postprocess = cast(bool, parameters.get("partial_postprocess", False))
if partial_postprocess and output_type != "pt":
output_type = "pt"
tensor = tensor.to(self.device, self.dtype)
if scaling_factor is not None:
tensor = tensor / scaling_factor
with torch.no_grad():
frames = cast(torch.Tensor, self.vae.decode(tensor, return_dict=False)[0])
if partial_postprocess:
frames = frames[0].permute(1, 0, 2, 3)
frames = torch.stack([(frame * 0.5 + 0.5).clamp(0, 1) for frame in frames])
frames = frames.permute(0, 2, 3, 1).contiguous().float()
frames = (frames * 255).round().to(torch.uint8)
elif output_type == "pil":
frames = cast(torch.Tensor, self.video_processor.postprocess_video(frames, output_type="pt")[0])
elif output_type == "mp4":
frames = cast(torch.Tensor, self.video_processor.postprocess_video(frames, output_type="pil")[0])
path = export_to_video(frames, fps=15)
with open(path, "rb") as f:
frames = f.read()
elif output_type == "pt":
frames = frames
return frames
|