LoloSemper commited on
Commit
67e2ced
·
verified ·
1 Parent(s): e9af694

Upload 4 files

Browse files
Files changed (1) hide show
  1. app.py +45 -7
app.py CHANGED
@@ -11,9 +11,14 @@
11
  # // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
  # // See the License for the specific language governing permissions and
13
  # // limitations under the License.
 
 
 
 
 
 
14
  import spaces
15
 
16
- import os
17
  import torch
18
  import mediapy
19
  from einops import rearrange
@@ -222,7 +227,20 @@ def generation_step(runner, text_embeds_dict, cond_latents):
222
  return samples
223
 
224
  @spaces.GPU(duration=100)
225
- def generation_loop(video_path='./test_videos', seed=666, fps_out=12, batch_size=1, cfg_scale=1.0, cfg_rescale=0.0, sample_steps=1, res_h=1280, res_w=720, sp_size=1):
 
 
 
 
 
 
 
 
 
 
 
 
 
226
  runner = configure_runner(1)
227
 
228
  def _extract_text_embeds():
@@ -240,8 +258,8 @@ def generation_loop(video_path='./test_videos', seed=666, fps_out=12, batch_size
240
  return positive_prompts_embeds
241
 
242
  def cut_videos(videos, sp_size):
243
- if videos.size(1) > 121:
244
- videos = videos[:, :121]
245
  t = videos.size(1)
246
  if t <= 4 * sp_size:
247
  print(f"Cut input video size: {videos.size()}")
@@ -317,8 +335,8 @@ def generation_loop(video_path='./test_videos', seed=666, fps_out=12, batch_size
317
  )[0]
318
  / 255.0
319
  )
320
- if video.size(0) > 121:
321
- video = video[:121]
322
  print(f"Read video size: {video.size()}")
323
  output_dir = 'output/' + str(uuid.uuid4()) + '.mp4'
324
  else:
@@ -356,6 +374,11 @@ def generation_loop(video_path='./test_videos', seed=666, fps_out=12, batch_size
356
  print(f"Encoding videos: {list(map(lambda x: x.size(), cond_latents))}")
357
  # runner.vae.to(torch.device("cuda"))
358
  cond_latents = runner.vae_encode(cond_latents)
 
 
 
 
 
359
  # runner.vae.to("cpu")
360
  # runner.dit.to(torch.device("cuda"))
361
 
@@ -429,13 +452,28 @@ with gr.Blocks(title="SeedVR2: One-Step Video Restoration via Diffusion Adversar
429
  seed = gr.Number(label="Seeds", value=666)
430
  fps = gr.Number(label="fps", value=24)
431
 
 
 
 
 
 
 
 
 
 
 
 
432
  with gr.Row():
433
  output_video = gr.Video(label="Output_Video")
434
  output_image = gr.Image(label="Output_Image")
435
  download_link = gr.File(label="Download the output")
436
 
437
  run_button = gr.Button("Run")
438
- run_button.click(fn=generation_loop, inputs=[input_video, seed, fps], outputs=[output_image, output_video, download_link])
 
 
 
 
439
 
440
  # Examples
441
  gr.Examples(
 
11
  # // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
  # // See the License for the specific language governing permissions and
13
  # // limitations under the License.
14
+ import os
15
+
16
+ # ZeroGPU: evitar fragmentacion del asignador CUDA durante los picos de memoria
17
+ # de SeedVR2. Debe establecerse antes de importar spaces o torch.
18
+ os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "expandable_segments:True")
19
+
20
  import spaces
21
 
 
22
  import torch
23
  import mediapy
24
  from einops import rearrange
 
227
  return samples
228
 
229
  @spaces.GPU(duration=100)
230
+ def generation_loop(video_path='./test_videos', seed=666, fps_out=12,
231
+ max_frames=17, work_height=540, batch_size=1,
232
+ cfg_scale=1.0, cfg_rescale=0.0, sample_steps=1,
233
+ sp_size=1):
234
+ # Diagnostico controlado: primero validar 17 fotogramas a 540p; luego
235
+ # ampliar gradualmente. La salida se recorta a max_frames deliberadamente.
236
+ max_frames = int(max_frames)
237
+ work_height = int(work_height)
238
+ if max_frames not in (17, 33, 65, 121):
239
+ raise ValueError("Elige 17, 33, 65 o 121 fotogramas.")
240
+ if work_height not in (360, 540, 720):
241
+ raise ValueError("Elige una altura de 360, 540 o 720 pixeles.")
242
+ res_h, res_w = int(work_height * 16 / 9), work_height
243
+ print(f"DIAGNOSTICO: hasta {max_frames} fotogramas, objetivo {work_height}p.")
244
  runner = configure_runner(1)
245
 
246
  def _extract_text_embeds():
 
258
  return positive_prompts_embeds
259
 
260
  def cut_videos(videos, sp_size):
261
+ if videos.size(1) > max_frames:
262
+ videos = videos[:, :max_frames]
263
  t = videos.size(1)
264
  if t <= 4 * sp_size:
265
  print(f"Cut input video size: {videos.size()}")
 
335
  )[0]
336
  / 255.0
337
  )
338
+ if video.size(0) > max_frames:
339
+ video = video[:max_frames]
340
  print(f"Read video size: {video.size()}")
341
  output_dir = 'output/' + str(uuid.uuid4()) + '.mp4'
342
  else:
 
374
  print(f"Encoding videos: {list(map(lambda x: x.size(), cond_latents))}")
375
  # runner.vae.to(torch.device("cuda"))
376
  cond_latents = runner.vae_encode(cond_latents)
377
+ gc.collect()
378
+ torch.cuda.empty_cache()
379
+ print("Memoria CUDA antes del DiT: "
380
+ f"asignada={torch.cuda.memory_allocated() / 2**30:.2f} GiB, "
381
+ f"reservada={torch.cuda.memory_reserved() / 2**30:.2f} GiB")
382
  # runner.vae.to("cpu")
383
  # runner.dit.to(torch.device("cuda"))
384
 
 
452
  seed = gr.Number(label="Seeds", value=666)
453
  fps = gr.Number(label="fps", value=24)
454
 
455
+ gr.Markdown("**Prueba de memoria:** se exportan solo los primeros fotogramas "
456
+ "seleccionados. Empieza por 17 fotogramas / 540p; si funciona, "
457
+ "sube a 33 y después a 720p.")
458
+ with gr.Row():
459
+ max_frames_control = gr.Dropdown(
460
+ label="Fotogramas maximos (recorta la salida)",
461
+ choices=[17, 33, 65, 121], value=17)
462
+ work_height_control = gr.Dropdown(
463
+ label="Altura de procesamiento (aproximada)",
464
+ choices=[360, 540, 720], value=540)
465
+
466
  with gr.Row():
467
  output_video = gr.Video(label="Output_Video")
468
  output_image = gr.Image(label="Output_Image")
469
  download_link = gr.File(label="Download the output")
470
 
471
  run_button = gr.Button("Run")
472
+ run_button.click(
473
+ fn=generation_loop,
474
+ inputs=[input_video, seed, fps, max_frames_control, work_height_control],
475
+ outputs=[output_image, output_video, download_link],
476
+ )
477
 
478
  # Examples
479
  gr.Examples(