Spaces:
Running
on
L40S
Running
on
L40S
File size: 5,391 Bytes
d516aa4 7ba1d45 d516aa4 7ba1d45 d516aa4 7ba1d45 d516aa4 7ba1d45 70f11df 7ba1d45 |
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 |
import torch
import os
import shutil
import subprocess
import gradio as gr
import json
import tempfile
from huggingface_hub import snapshot_download
# Download All Required Models using `snapshot_download`
def download_and_extract(repo_id, target_dir):
"""
Downloads a model repo (cached) and copies its contents to a local target directory.
If the target_dir exists, it will be updated (not re-downloaded if cache is present).
"""
print(f"Downloading {repo_id} into cache...")
snapshot_path = snapshot_download(repo_id)
print(f"Copying files to {target_dir}...")
os.makedirs(target_dir, exist_ok=True)
shutil.copytree(snapshot_path, target_dir, dirs_exist_ok=True)
print(f"Done: {repo_id} extracted to {target_dir}")
return target_dir
wan_model_path = download_and_extract("Wan-AI/Wan2.1-I2V-14B-480P", "./weights/Wan2.1-I2V-14B-480P")
wav2vec_path = download_and_extract("TencentGameMate/chinese-wav2vec2-base", "./weights/chinese-wav2vec2-base")
multitalk_path = download_and_extract("MeiGen-AI/MeiGen-MultiTalk", "./weights/MeiGen-MultiTalk")
# Define paths
base_model_dir = "./weights/Wan2.1-I2V-14B-480P"
multitalk_dir = "./weights/MeiGen-MultiTalk"
# File to rename
original_index = os.path.join(base_model_dir, "diffusion_pytorch_model.safetensors.index.json")
backup_index = os.path.join(base_model_dir, "diffusion_pytorch_model.safetensors.index.json_old")
# Rename the original index file
if os.path.exists(original_index):
os.rename(original_index, backup_index)
print("Renamed original index file to .json_old")
# Copy updated index file from MultiTalk
shutil.copy2(
os.path.join(multitalk_dir, "diffusion_pytorch_model.safetensors.index.json"),
base_model_dir
)
# Copy MultiTalk model weights
shutil.copy2(
os.path.join(multitalk_dir, "multitalk.safetensors"),
base_model_dir
)
print("Copied MultiTalk files into base model directory.")
# Check if CUDA-compatible GPU is available
if torch.cuda.is_available():
# Get current GPU name
gpu_name = torch.cuda.get_device_name(torch.cuda.current_device())
print(f"Current GPU: {gpu_name}")
# Enforce GPU requirement
if "A100" not in gpu_name and "L4" not in gpu_name:
raise RuntimeError(f"This notebook requires an A100 or L4 GPU. Found: {gpu_name}")
elif "L4" in gpu_name:
print("Warning: L4 is supported, but A100 is recommended for faster inference.")
else:
raise RuntimeError("No CUDA-compatible GPU found. An A100 or L4 GPU is required.")
GPU_TO_VRAM_PARAMS = {
"NVIDIA A100": 11000000000,
"NVIDIA A100-SXM4-40GB": 11000000000,
"NVIDIA A100-SXM4-80GB": 22000000000,
"NVIDIA L4": 5000000000
}
USED_VRAM_PARAMS = GPU_TO_VRAM_PARAMS[gpu_name]
print("Using", USED_VRAM_PARAMS, "for num_persistent_param_in_dit")
def create_temp_input_json(prompt: str, cond_image_path: str, cond_audio_path: str) -> str:
"""
Create a temporary JSON file with the user-provided prompt, image, and audio paths.
Returns the path to the temporary JSON file.
"""
# Structure based on your original JSON format
data = {
"prompt": prompt,
"cond_image": cond_image_path,
"cond_audio": {
"person1": cond_audio_path
}
}
# Create a temp file
temp_json = tempfile.NamedTemporaryFile(delete=False, suffix=".json", mode='w', encoding='utf-8')
json.dump(data, temp_json, indent=4)
temp_json_path = temp_json.name
temp_json.close()
print(f"Temporary input JSON saved to: {temp_json_path}")
return temp_json_path
def infer(prompt, cond_image_path, cond_audio_path):
# Example usage (from user input)
prompt = "A woman sings passionately in a dimly lit studio."
cond_image_path = "examples/single/single1.png" # Assume uploaded via Gradio
cond_audio_path = "examples/single/1.wav" # Assume uploaded via Gradio
input_json_path = create_temp_input_json(prompt, cond_image_path, cond_audio_path)
cmd = [
"python3", "generate_multitalk.py",
"--ckpt_dir", "weights/Wan2.1-I2V-14B-480P",
"--wav2vec_dir", "weights/chinese-wav2vec2-base",
"--input_json", "./examples/single_example_1.json",
"--sample_steps", "20",
"--num_persistent_param_in_dit", str(USED_VRAM_PARAMS),
"--mode", "streaming",
"--use_teacache",
"--save_file", "multi_long_mediumvram_exp"
]
subprocess.run(cmd, check=True)
return "multi_long_mediumvra_exp.mp4"
with gr.Blocks(title="MultiTalk Inference") as demo:
gr.Markdown("## 🎤 MultiTalk Inference Demo")
with gr.Row():
with gr.Column():
prompt_input = gr.Textbox(
label="Text Prompt",
placeholder="Describe the scene...",
lines=4
)
image_input = gr.Image(
type="filepath",
label="Conditioning Image"
)
audio_input = gr.Audio(
type="filepath",
label="Conditioning Audio (.wav)"
)
submit_btn = gr.Button("Generate")
with gr.Column():
output_video = gr.Video(label="Generated Video")
submit_btn.click(
fn=infer,
inputs=[prompt_input, image_input, audio_input],
outputs=output_video
)
demo.launch()
|