File size: 8,831 Bytes
2df6ae7 af36d5a 8b4a69c af36d5a 2df6ae7 8b4a69c 8b8ed09 8b4a69c 8d2995d 2df6ae7 8b4a69c 8b8ed09 8d2995d 8b8ed09 8d2995d 2df6ae7 8b4a69c 2df6ae7 97a4a3d 2df6ae7 8d2995d 8b4a69c 8d2995d 2df6ae7 |
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 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 |
from dataclasses import dataclass
from typing import Dict, Any, Optional
import base64
import logging
import random
import torch
from diffusers import HunyuanVideoPipeline
from varnish import Varnish
from enhance_a_video import enable_enhance, inject_enhance_for_hunyuanvideo, set_enhance_weight
from teacache import enable_teacache, disable_teacache
# Configure logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
@dataclass
class GenerationConfig:
"""Configuration for video generation"""
# Content settings
prompt: str
negative_prompt: str = ""
# Model settings
num_frames: int = 49 # Should be 4k + 1 format
height: int = 320
width: int = 576
num_inference_steps: int = 50
guidance_scale: float = 7.0
# Reproducibility
seed: int = -1
# Varnish post-processing settings
fps: int = 30
double_num_frames: bool = False
super_resolution: bool = False
grain_amount: float = 0.0
quality: int = 18 # CRF scale (0-51, lower is better)
# Audio settings
enable_audio: bool = False
audio_prompt: str = ""
audio_negative_prompt: str = "voices, voice, talking, speaking, speech"
# TeaCache settings
enable_teacache: bool = True
teacache_threshold: float = 0.15 # values: 0 (original), 0.1 (1.6x speedup), 0.15 (2.1x speedup)
# Enhance-A-Video settings
enable_enhance_a_video: bool = True
enhance_a_video_weight: float = 4.0
def validate_and_adjust(self) -> 'GenerationConfig':
"""Validate and adjust parameters"""
# Ensure num_frames follows 4k + 1 format
k = (self.num_frames - 1) // 4
self.num_frames = (k * 4) + 1
# Set random seed if not specified
if self.seed == -1:
self.seed = random.randint(0, 2**32 - 1)
return self
class EndpointHandler:
"""Handles video generation requests using HunyuanVideo and Varnish"""
def __init__(self, path: str = ""):
"""Initialize handler with models
Args:
path: Path to model weights
"""
self.device = "cuda" if torch.cuda.is_available() else "cpu"
# Initialize HunyuanVideo pipeline
self.pipeline = HunyuanVideoPipeline.from_pretrained(
path,
torch_dtype=torch.float16,
).to(self.device)
# Initialize text encoders in float16
self.pipeline.text_encoder = self.pipeline.text_encoder.half()
self.pipeline.text_encoder_2 = self.pipeline.text_encoder_2.half()
# Initialize transformer in bfloat16
self.pipeline.transformer = self.pipeline.transformer.to(torch.bfloat16)
# Initialize VAE in float16
self.pipeline.vae = self.pipeline.vae.half()
# Initialize Varnish for post-processing
self.varnish = Varnish(
device=self.device,
model_base_dir="/repository/varnish"
)
def __call__(self, data: Dict[str, Any]) -> Dict[str, Any]:
"""Process video generation requests
Args:
data: Request data containing:
- inputs (str): Prompt for video generation
- parameters (dict): Generation parameters
Returns:
Dictionary containing:
- video: Base64 encoded MP4 data URI
- content-type: MIME type
- metadata: Generation metadata
"""
# Extract inputs
inputs = data.pop("inputs", data)
if isinstance(inputs, dict):
prompt = inputs.get("prompt", "")
else:
prompt = inputs
params = data.get("parameters", {})
# Create and validate config
config = GenerationConfig(
prompt=prompt,
negative_prompt=params.get("negative_prompt", ""),
num_frames=params.get("num_frames", 49),
height=params.get("height", 320),
width=params.get("width", 576),
num_inference_steps=params.get("num_inference_steps", 50),
guidance_scale=params.get("guidance_scale", 7.0),
seed=params.get("seed", -1),
fps=params.get("fps", 30),
double_num_frames=params.get("double_num_frames", False),
super_resolution=params.get("super_resolution", False),
grain_amount=params.get("grain_amount", 0.0),
quality=params.get("quality", 18),
enable_audio=params.get("enable_audio", False),
audio_prompt=params.get("audio_prompt", ""),
audio_negative_prompt=params.get("audio_negative_prompt", "voices, voice, talking, speaking, speech"),
enable_teacache=params.get("enable_teacache", True),
# values: 0 (original), 0.1 (1.6x speedup), 0.15 (2.1x speedup).
teacache_threshold=params.get("teacache_threshold", 0.15),
enable_enhance_a_video=params.get("enable_enhance_a_video", True),
enhance_a_video_weight=params.get("enhance_a_video_weight", 4.0)
).validate_and_adjust()
try:
# Set random seeds
if config.seed != -1:
torch.manual_seed(config.seed)
random.seed(config.seed)
generator = torch.Generator(device=self.device).manual_seed(config.seed)
else:
generator = None
# Configure TeaCache
if config.enable_teacache:
enable_teacache(
self.pipeline.transformer,
num_inference_steps=config.num_inference_steps,
rel_l1_thresh=config.teacache_threshold
)
else:
disable_teacache(self.pipeline.transformer)
# Generate video frames
with torch.inference_mode():
output = self.pipeline(
prompt=config.prompt,
# Failed to generate video: HunyuanVideoPipeline.__call__() got an unexpected keyword argument 'negative_prompt'
#negative_prompt=config.negative_prompt,
num_frames=config.num_frames,
height=config.height,
width=config.width,
num_inference_steps=config.num_inference_steps,
guidance_scale=config.guidance_scale,
generator=generator,
output_type="pt",
).frames
# Process with Varnish
import asyncio
try:
loop = asyncio.get_event_loop()
except RuntimeError:
loop = asyncio.new_event_loop()
asyncio.set_event_loop(loop)
result = loop.run_until_complete(
self.varnish(
input_data=output,
fps=config.fps,
double_num_frames=config.double_num_frames,
super_resolution=config.super_resolution,
grain_amount=config.grain_amount,
enable_audio=config.enable_audio,
audio_prompt=config.audio_prompt,
audio_negative_prompt=config.audio_negative_prompt,
)
)
# Get video data URI
video_uri = loop.run_until_complete(
result.write(
type="data-uri",
quality=config.quality
)
)
return {
"video": video_uri,
"content-type": "video/mp4",
"metadata": {
"width": result.metadata.width,
"height": result.metadata.height,
"num_frames": result.metadata.frame_count,
"fps": result.metadata.fps,
"duration": result.metadata.duration,
"seed": config.seed,
"enable_teacache": config.enable_teacache,
"teacache_threshold": config.teacache_threshold if config.enable_teacache else 0,
"enable_enhance_a_video": config.enable_enhance_a_video,
"enhance_a_video_weight": config.enhance_a_video_weight if config.enable_enhance_a_video else 0,
}
}
except Exception as e:
logger.error(f"Error generating video: {str(e)}")
raise RuntimeError(f"Failed to generate video: {str(e)}") |