cuio commited on
Commit
cb8d625
·
verified ·
1 Parent(s): a6f7e75

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +60 -10
app.py CHANGED
@@ -16,6 +16,8 @@
16
 
17
  import argparse
18
  import os
 
 
19
 
20
  import gradio as gr
21
  import numpy as np
@@ -25,6 +27,49 @@ import torch
25
  from diffnext.pipelines import URSAPipeline
26
  from diffnext.utils import export_to_image, export_to_video
27
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
28
  # Fix tokenizer fork issue.
29
  os.environ["TOKENIZERS_PARALLELISM"] = "true"
30
  # Switch to the allocator optimized for dynamic shape.
@@ -34,7 +79,7 @@ os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
34
  def parse_args():
35
  """Parse arguments."""
36
  parser = argparse.ArgumentParser(description="Serve URSA TI2V application")
37
- parser.add_argument("--model", default="BAAI/URSA-0.6B-FSQ320", help="model path")
38
  parser.add_argument("--device", type=int, default=0, help="device index")
39
  parser.add_argument("--precision", default="float16", help="compute precision")
40
  return parser.parse_args()
@@ -111,10 +156,10 @@ video_presets = {
111
 
112
  prompts = [
113
  "a lone grizzly bear walks through a misty forest at dawn, sunlight catching its fur.",
114
- "Many spotted jellyfish pulsating under water. Their bodies are transparent and glowing in deep ocean.", # noqa
115
- "An intense close-up of a soldier’s face, covered in dirt and sweat, his eyes filled with determination as he surveys the battlefield.", # noqa
116
- "a close-up shot of a woman standing in a dimly lit room. she is wearing a traditional chinese outfit, which includes a red and gold dress with intricate designs and a matching headpiece. the woman has her hair styled in an updo, adorned with a gold accessory. her makeup is done in a way that accentuates her features, with red lipstick and dark eyeshadow. she is looking directly at the camera with a neutral expression. the room has a rustic feel, with wooden beams and a stone wall visible in the background. the lighting in the room is soft and warm, creating a contrast with the woman's vibrant attire. there are no texts or other objects in the video. the style of the video is a portrait, focusing on the woman and her attire.", # noqa
117
- "The camera slowly rotates around a massive stack of vintage televisions that are placed within a large New York museum gallery. Each of the televisions is showing a different program. There are 1950s sci-fi movies with their distinctive visuals, horror movies with their creepy scenes, news broadcasts with moving images and words, static on some screens, and a 1970s sitcom with its characteristic look. The televisions are of various sizes and designs, some with rounded edges and others with more angular shapes. The gallery is well-lit, with light falling on the stack of televisions and highlighting the different programs being shown. There are no people visible in the immediate vicinity, only the stack of televisions and the surrounding gallery space.", # noqa
118
  ]
119
  motion_scores = [9, 9, 9, 9, 9]
120
  videos = ["", "", "", "", ""]
@@ -125,7 +170,12 @@ if __name__ == "__main__":
125
  args = parse_args()
126
 
127
  device = torch.device("cuda" if torch.cuda.is_available() else "cpu", args.device)
128
- model_args = {"torch_dtype": getattr(torch, args.precision.lower()), "trust_remote_code": True}
 
 
 
 
 
129
  pipe = URSAPipeline.from_pretrained(args.model, **model_args).to(device)
130
 
131
  # Application.
@@ -138,13 +188,13 @@ if __name__ == "__main__":
138
  prompt = gr.Text(
139
  label="Prompt",
140
  placeholder="Describe the video you want to generate",
141
- value="A lone grizzly bear walks through a misty forest at dawn, sunlight catching its fur.", # noqa
142
  lines=5,
143
  )
144
  negative_prompt = gr.Text(
145
  label="Negative Prompt",
146
  placeholder="Describe what you don't want in the video",
147
- value="worst quality, low quality, inconsistent motion, static, still, blurry, jittery, distorted, ugly", # noqa
148
  lines=1,
149
  )
150
  with gr.Row():
@@ -158,7 +208,7 @@ if __name__ == "__main__":
158
  randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
159
  guidance_scale = gr.Slider(label="Guidance scale", minimum=1, maximum=10.0, step=0.1, value=7.0)
160
  with gr.Row():
161
- num_inference_steps = gr.Slider(label="Inference steps", minimum=1, maximum=100, step=1, value=50) # noqa
162
  options.__exit__(), input_col.__exit__()
163
 
164
  # Results.
@@ -201,4 +251,4 @@ if __name__ == "__main__":
201
  ],
202
  outputs=[result, seed],
203
  )
204
- app.__exit__(), app.launch(share=False, ssr_mode=False, max_threads=1)
 
16
 
17
  import argparse
18
  import os
19
+ import sys
20
+ import types
21
 
22
  import gradio as gr
23
  import numpy as np
 
27
  from diffnext.pipelines import URSAPipeline
28
  from diffnext.utils import export_to_image, export_to_video
29
 
30
+ # =========================
31
+ # 🔧 关键补丁:修复 Qwen3Config 缺失 rope_theta
32
+ # =========================
33
+ def _patch_diffnext_qwen3():
34
+ """
35
+ 兼容 transformers 新版本中 Qwen3Config 没有 rope_theta 的问题
36
+ """
37
+ try:
38
+ from diffnext.models.text_encoders import qwen3
39
+ except Exception:
40
+ return
41
+
42
+ orig_from_config = qwen3.Qwen3RotaryEmbedding.from_config
43
+
44
+ @classmethod
45
+ def patched_from_config(cls, config):
46
+ # ---- rope_theta ----
47
+ if not hasattr(config, "rope_theta"):
48
+ # 常见存放位置之一
49
+ rp = getattr(config, "rope_parameters", None)
50
+ if isinstance(rp, dict) and "rope_theta" in rp:
51
+ config.rope_theta = rp["rope_theta"]
52
+ else:
53
+ # URSA / Qwen3 常用默认值
54
+ config.rope_theta = float(1e6)
55
+
56
+ # ---- max_position_embeddings ----
57
+ if not hasattr(config, "max_position_embeddings"):
58
+ config.max_position_embeddings = int(
59
+ getattr(config, "max_seq_len", 2048)
60
+ )
61
+
62
+ return orig_from_config.__func__(cls, config)
63
+
64
+ qwen3.Qwen3RotaryEmbedding.from_config = patched_from_config
65
+
66
+
67
+ _patch_diffnext_qwen3()
68
+ # =========================
69
+ # 🔧 补丁结束
70
+ # =========================
71
+
72
+
73
  # Fix tokenizer fork issue.
74
  os.environ["TOKENIZERS_PARALLELISM"] = "true"
75
  # Switch to the allocator optimized for dynamic shape.
 
79
  def parse_args():
80
  """Parse arguments."""
81
  parser = argparse.ArgumentParser(description="Serve URSA TI2V application")
82
+ parser.add_argument("--model", default="BAAI/URSA-0.6K-FSQ320", help="model path")
83
  parser.add_argument("--device", type=int, default=0, help="device index")
84
  parser.add_argument("--precision", default="float16", help="compute precision")
85
  return parser.parse_args()
 
156
 
157
  prompts = [
158
  "a lone grizzly bear walks through a misty forest at dawn, sunlight catching its fur.",
159
+ "Many spotted jellyfish pulsating under water. Their bodies are transparent and glowing in deep ocean.",
160
+ "An intense close-up of a soldier’s face, covered in dirt and sweat, his eyes filled with determination as he surveys the battlefield.",
161
+ "a close-up shot of a woman standing in a dimly lit room. she is wearing a traditional chinese outfit, which includes a red and gold dress with intricate designs and a matching headpiece. the woman has her hair styled in an updo, adorned with a gold accessory. her makeup is done in a way that accentuates her features, with red lipstick and dark eyeshadow. she is looking directly at the camera with a neutral expression. the room has a rustic feel, with wooden beams and a stone wall visible in the background. the lighting in the room is soft and warm, creating a contrast with the woman's vibrant attire. there are no texts or other objects in the video. the style of the video is a portrait, focusing on the woman and her attire.",
162
+ "The camera slowly rotates around a massive stack of vintage televisions that are placed within a large New York museum gallery. Each of the televisions is showing a different program. There are 1950s sci-fi movies with their distinctive visuals, horror movies with their creepy scenes, news broadcasts with moving images and words, static on some screens, and a 1970s sitcom with its characteristic look. The televisions are of various sizes and designs, some with rounded edges and others with more angular shapes. The gallery is well-lit, with light falling on the stack of televisions and highlighting the different programs being shown. There are no people visible in the immediate vicinity, only the stack of televisions and the surrounding gallery space.",
163
  ]
164
  motion_scores = [9, 9, 9, 9, 9]
165
  videos = ["", "", "", "", ""]
 
170
  args = parse_args()
171
 
172
  device = torch.device("cuda" if torch.cuda.is_available() else "cpu", args.device)
173
+ model_args = {
174
+ "torch_dtype": getattr(torch, args.precision.lower()),
175
+ "trust_remote_code": True,
176
+ }
177
+
178
+ # ✅ 不再报 rope_theta 错
179
  pipe = URSAPipeline.from_pretrained(args.model, **model_args).to(device)
180
 
181
  # Application.
 
188
  prompt = gr.Text(
189
  label="Prompt",
190
  placeholder="Describe the video you want to generate",
191
+ value="A lone grizzly bear walks through a misty forest at dawn, sunlight catching its fur.",
192
  lines=5,
193
  )
194
  negative_prompt = gr.Text(
195
  label="Negative Prompt",
196
  placeholder="Describe what you don't want in the video",
197
+ value="worst quality, low quality, inconsistent motion, static, still, blurry, jittery, distorted, ugly",
198
  lines=1,
199
  )
200
  with gr.Row():
 
208
  randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
209
  guidance_scale = gr.Slider(label="Guidance scale", minimum=1, maximum=10.0, step=0.1, value=7.0)
210
  with gr.Row():
211
+ num_inference_steps = gr.Slider(label="Inference steps", minimum=1, maximum=100, step=1, value=50)
212
  options.__exit__(), input_col.__exit__()
213
 
214
  # Results.
 
251
  ],
252
  outputs=[result, seed],
253
  )
254
+ app.__exit__(), app.launch(share=False, ssr_mode=False, max_threads=1)