Python
Merge remote-tracking branch 'upstream/main' into queues
Chris Malonebb3f99012fa7
diff --cc gradio_server.pyindex 7848b39,674d6df..d7037e4--- a/gradio_server.py+++ b/gradio_server.py@@@ -1219,161 -1154,188 +1225,161 @@@ def generate_videoif seed == None or seed <0:seed = random.randint(0, 999999999)- file_list = []- state["file_list"] = file_list+ global file_list+ state["file_list"] = file_listglobal save_pathos.makedirs(save_path, exist_ok=True)- video_no = 0- total_video = repeat_generation * len(prompts)- state["total_video"] = total_video- extra_generation = 0abort = False- start_time = time.time()- state["prompts_max"] = len(prompts)- for no, prompt in enumerate(prompts):- state["prompt"] = prompt- repeat_no = 0- state["prompt_no"] = no- extra_generation = 0- yield f"Prompt No{no}"- while True:- extra_orders = state.get("extra_orders",0)- state["extra_orders"] = 0- extra_generation += extra_orders- state["total_generation"] = repeat_generation + extra_generation- # total_video += (len(prompts)- no) * extra_orders- total_video += extra_orders- if abort or repeat_no >= (repeat_generation + extra_generation):- break-- if trans.enable_teacache:- trans.teacache_counter = 0- trans.num_steps = num_inference_steps- trans.teacache_skipped_steps = 0- trans.previous_residual_uncond = None- trans.previous_residual_cond = None-- video_no += 1- status = f"Video {video_no}/{total_video}"- state["video_no"] = video_no- state["progress_status"] = status- state["progress_phase"] = (" - Encoding Prompt", -1 )- progress(0, desc=status + " - Encoding Prompt" )- callback = build_callback(state, trans, progress, status, num_inference_steps)- offload.shared_state["callback"] = callback--- gc.collect()- torch.cuda.empty_cache()- wan_model._interrupt = False- try:- if image2video:- samples = wan_model.generate(- prompt,- image_to_continue[no].convert('RGB'),- image_to_end[no].convert('RGB') if image_to_end != None else None,- frame_num=(video_length // 4)* 4 + 1,- max_area=MAX_AREA_CONFIGS[resolution],- shift=flow_shift,- sampling_steps=num_inference_steps,- guide_scale=guidance_scale,- n_prompt=negative_prompt,- seed=seed,- offload_model=False,- callback=callback,- enable_RIFLEx = enable_RIFLEx,- VAE_tile_size = VAE_tile_size,- joint_pass = joint_pass,- slg_layers = slg_layers,- slg_start = slg_start/100,- slg_end = slg_end/100,- )+ with tracker_lock:+ progress_tracker[task_id] = {+ 'current_step': 0,+ 'total_steps': num_inference_steps,+ 'start_time': time.time(),+ 'last_update': time.time()+ }+ if trans.enable_teacache:+ trans.teacache_counter = 0+ trans.num_steps = num_inference_steps+ trans.teacache_skipped_steps = 0+ trans.previous_residual_uncond = None+ trans.previous_residual_cond = None+ callback = build_callback(task_id, num_inference_steps)+ offload.shared_state["callback"] = callback+ gc.collect()+ torch.cuda.empty_cache()+ wan_model._interrupt = False+ state["progress_status"] = "Starting"+ try:+ if image2video:+ samples = wan_model.generate(+ prompt,+ image_to_continue.convert('RGB'),+ image_to_end.convert('RGB') if image_to_end != None else None,+ frame_num=(video_length // 4)* 4 + 1,+ max_area=MAX_AREA_CONFIGS[resolution],+ shift=flow_shift,+ sampling_steps=num_inference_steps,+ guide_scale=guidance_scale,+ n_prompt=negative_prompt,+ seed=seed,+ offload_model=False,+ callback=callback,+ enable_RIFLEx = enable_RIFLEx,+ VAE_tile_size = VAE_tile_size,+ joint_pass = joint_pass,+ slg_layers = slg_layers,+ slg_start = slg_start/100,+ slg_end = slg_end/100,+ )- else:- samples = wan_model.generate(- prompt,- frame_num=(video_length // 4)* 4 + 1,- size=(width, height),- shift=flow_shift,- sampling_steps=num_inference_steps,- guide_scale=guidance_scale,- n_prompt=negative_prompt,- seed=seed,- offload_model=False,- callback=callback,- enable_RIFLEx = enable_RIFLEx,- VAE_tile_size = VAE_tile_size,- joint_pass = joint_pass,- slg_layers = slg_layers,- slg_start = slg_start/100,- slg_end = slg_end/100,- )- except Exception as e:- gen_in_progress = False- if temp_filename!= None and os.path.isfile(temp_filename):- os.remove(temp_filename)- offload.last_offload_obj.unload_all()- offload.unload_loras_from_model(trans)- # if compile:- # cache_size = torch._dynamo.config.cache_size_limit- # torch.compiler.reset()- # torch._dynamo.config.cache_size_limit = cache_size-- gc.collect()- torch.cuda.empty_cache()- s = str(e)- keyword_list = ["vram", "VRAM", "memory","allocat"]- VRAM_crash= False- if any( keyword in s for keyword in keyword_list):+ else:+ samples = wan_model.generate(+ prompt,+ frame_num=(video_length // 4)* 4 + 1,+ size=(width, height),+ shift=flow_shift,+ sampling_steps=num_inference_steps,+ guide_scale=guidance_scale,+ n_prompt=negative_prompt,+ seed=seed,+ offload_model=False,+ callback=callback,+ enable_RIFLEx = enable_RIFLEx,+ VAE_tile_size = VAE_tile_size,+ joint_pass = joint_pass,+ slg_layers = slg_layers,+ slg_start = slg_start/100,+ slg_end = slg_end/100,+ )+ except Exception as e:+ gen_in_progress = False+ if temp_filename!= None and os.path.isfile(temp_filename):+ os.remove(temp_filename)+ offload.last_offload_obj.unload_all()+ offload.unload_loras_from_model(trans)+ # if compile:+ # cache_size = torch._dynamo.config.cache_size_limit+ # torch.compiler.reset()+ # torch._dynamo.config.cache_size_limit = cache_size++ gc.collect()+ torch.cuda.empty_cache()+ s = str(e)+ keyword_list = ["vram", "VRAM", "memory","allocat"]+ VRAM_crash= False+ if any( keyword in s for keyword in keyword_list):+ VRAM_crash = True+ else:+ stack = traceback.extract_stack(f=None, limit=5)+ for frame in stack:+ if any( keyword in frame.name for keyword in keyword_list):VRAM_crash = True- else:- stack = traceback.extract_stack(f=None, limit=5)- for frame in stack:- if any( keyword in frame.name for keyword in keyword_list):- VRAM_crash = True- break- state["prompt"] = ""- if VRAM_crash:- raise gr.Error("The generation of the video has encountered an error: it is likely that you have unsufficient VRAM and you should therefore reduce the video resolution or its number of frames.")- else:- raise gr.Error(f"The generation of the video has encountered an error, please check your terminal for more information. '{s}'")-- if trans.enable_teacache:- print(f"Teacache Skipped Steps:{trans.teacache_skipped_steps}/{num_inference_steps}" )- trans.previous_residual_uncond = None- trans.previous_residual_cond = None-- if samples != None:- samples = samples.to("cpu")- offload.last_offload_obj.unload_all()- gc.collect()- torch.cuda.empty_cache()-- if samples == None:- end_time = time.time()- abort = True- state["prompt"] = ""- yield f"Video generation was aborted. Total Generation Time: {end_time-start_time:.1f}s"- else:- sample = samples.cpu()- # video = rearrange(sample.cpu().numpy(), "c t h w -> t h w c")+ break+ state["prompt"] = ""+ if VRAM_crash:+ raise gr.Error("The generation of the video has encountered an error: it is likely that you have unsufficient VRAM and you should therefore reduce the video resolution or its number of frames.")+ else:+ raise gr.Error(f"The generation of the video has encountered an error, please check your terminal for more information. '{s}'")+ finally:+ with tracker_lock:+ if task_id in progress_tracker:+ del progress_tracker[task_id]- time_flag = datetime.fromtimestamp(time.time()).strftime("%Y-%m-%d-%Hh%Mm%Ss")- if os.name == 'nt':- file_name = f"{time_flag}_seed{seed}_{sanitize_file_name(prompt[:50]).strip()}.mp4"- else:- file_name = f"{time_flag}_seed{seed}_{sanitize_file_name(prompt[:100]).strip()}.mp4"- video_path = os.path.join(save_path, file_name)- cache_video(- tensor=sample[None],- save_file=video_path,- fps=16,- nrow=1,- normalize=True,- value_range=(-1, 1))-- configs = {- 'prompt': prompt,- 'negative_prompt': negative_prompt,- 'resolution': resolution,- 'video_length': video_length,- 'seed': seed,- 'num_inference_steps': num_inference_steps,- }-- metadata_choice = server_config.get("metadata_choice","metadata")- if metadata_choice == "json":- with open(video_path.replace('.mp4', '.json'), 'w') as f:- json.dump(configs, f, indent=4)- elif metadata_choice == "metadata":- from mutagen.mp4 import MP4- file = MP4(video_path)- file.tags['©cmt'] = [json.dumps(configs)]- file.save()-- print(f"New video saved to Path: "+video_path)- file_list.append(video_path)- if video_no < total_video:- yield status- else:- end_time = time.time()- state["prompt"] = ""- yield f"Total Generation Time: {end_time-start_time:.1f}s"- seed += 1- repeat_no += 1+ if trans.enable_teacache:+ print(f"Teacache Skipped Steps:{trans.teacache_skipped_steps}/{num_inference_steps}" )+ trans.previous_residual_uncond = None+ trans.previous_residual_cond = None++ if samples != None:+ samples = samples.to("cpu")+ offload.last_offload_obj.unload_all()+ gc.collect()+ torch.cuda.empty_cache()++ if samples == None:+ end_time = time.time()+ abort = True+ state["prompt"] = ""+ else:+ sample = samples.cpu()+ # video = rearrange(sample.cpu().numpy(), "c t h w -> t h w c")++ time_flag = datetime.fromtimestamp(time.time()).strftime("%Y-%m-%d-%Hh%Mm%Ss")+ if os.name == 'nt':+ file_name = f"{time_flag}_seed{seed}_{sanitize_file_name(prompt[:50]).strip()}.mp4"+ else:+ file_name = f"{time_flag}_seed{seed}_{sanitize_file_name(prompt[:100]).strip()}.mp4"+ video_path = os.path.join(save_path, file_name)+ cache_video(+ tensor=sample[None],+ save_file=video_path,+ fps=16,+ nrow=1,+ normalize=True,+ value_range=(-1, 1))++ configs = {+ 'prompt': prompt,+ 'negative_prompt': negative_prompt,+ 'resolution': resolution,+ 'video_length': video_length,+ 'seed': seed,+ 'num_inference_steps': num_inference_steps,+ }+- metadata_choice = server_config["metadata_choice"]++ metadata_choice = server_config.get("metadata_choice","metadata")+ if metadata_choice == "json":+ with open(video_path.replace('.mp4', '.json'), 'w') as f:+ json.dump(configs, f, indent=4)+ elif metadata_choice == "metadata":+ from mutagen.mp4 import MP4+ file = MP4(video_path)+ file.tags['©cmt'] = [json.dumps(configs)]+ file.save()++ print(f"New video saved to Path: "+video_path)+ file_list.append(video_path)+ seed += 1if temp_filename!= None and os.path.isfile(temp_filename):os.remove(temp_filename)