优化生成速度

simple_app.py

@@ -13,9 +13,10 @@ snapshot_download(
 
 SIZE_OPTIONS = ["480*832", "832*480"]  # 分辨率选项
 FRAME_NUM_OPTIONS = [10, 20, 30, 40, 50, 60, 81]  # 帧数选项
+SAMPLING_STEPS_OPTIONS = [5, 10, 15, 20, 25, 30, 40, 50] # 采样步数选项
 
 
-def infer(prompt, video_size, frame_num, progress=gr.Progress(track_tqdm=True)):  # 添加 video_size, frame_num 参数
+def infer(prompt, video_size, frame_num, sampling_steps, progress=gr.Progress(track_tqdm=True)):
     # Configuration:
     total_process_steps = 11  # Total INFO messages expected
     irrelevant_steps = 4  # First 4 INFO messages are ignored
@@ -51,10 +52,13 @@ def infer(prompt, video_size, frame_num, progress=gr.Progress(track_tqdm=True)):
         "generate",  # using -u for unbuffered output
         "--task",
         "t2v-1.3B",
+        "--fp16", # Enable FP16 for acceleration
         "--size",
-        video_size,  # 使用 WebUI 传递的分辨率参数
+        video_size,
         "--frame_num",
-        str(frame_num),  # 使用 WebUI 传递的帧数参数 (注意转为字符串)
+        str(frame_num),
+        "--sample_steps",
+        str(sampling_steps), # Add sampling steps
         "--ckpt_dir",
         "./Wan2.1-T2V-1.3B",
         "--t5_cpu",
@@ -206,23 +210,30 @@ with gr.Blocks() as demo:
 
         video_size_dropdown = gr.Dropdown(
             choices=SIZE_OPTIONS,
-            value=SIZE_OPTIONS[0],  # 默认选择第一个分辨率
+            value=SIZE_OPTIONS[0],
             label="Video Size (Resolution)",
         )
         frame_num_slider = gr.Slider(
-            minimum=FRAME_NUM_OPTIONS[0],  # 最小帧数
-            maximum=FRAME_NUM_OPTIONS[-1],  # 最大帧数
-            value=FRAME_NUM_OPTIONS[2],  # 默认帧数 (例如，选择列表的第三个)
-            step=1,  # 步长为 1
+            minimum=FRAME_NUM_OPTIONS[0],
+            maximum=FRAME_NUM_OPTIONS[-1],
+            value=FRAME_NUM_OPTIONS[2],
+            step=1,
             label="Frame Number (Video Length)",
         )
+        sampling_steps_slider = gr.Slider(
+            minimum=SAMPLING_STEPS_OPTIONS[0],
+            maximum=SAMPLING_STEPS_OPTIONS[-1],
+            value=SAMPLING_STEPS_OPTIONS[1], # Default to 10 steps
+            step=1,
+            label="Sampling Steps (Fewer steps = Faster, Lower quality)",
+        )
 
         submit_btn = gr.Button("Submit")
         video_res = gr.Video(label="Generated Video")
 
     submit_btn.click(
         fn=infer,
-        inputs=[prompt, video_size_dropdown, frame_num_slider],  # inputs 添加 video_size_dropdown, frame_num_slider
+        inputs=[prompt, video_size_dropdown, frame_num_slider, sampling_steps_slider],
         outputs=[video_res],
     )
 

wan/image2video.py

@@ -123,6 +123,10 @@ class WanI2V:
         else:
             if not init_on_cpu:
                 self.model.to(self.device)
+                try:
+                    self.model.enable_xformers_memory_efficient_attention()
+                except Exception as e:
+                    logging.warning(f"Could not enable xformers memory efficient attention: {e}")
 
         self.sample_neg_prompt = config.sample_neg_prompt
 

wan/text2video.py

@@ -104,6 +104,10 @@ class WanT2V:
             self.model = shard_fn(self.model)
         else:
             self.model.to(self.device)
+            try:
+                self.model.enable_xformers_memory_efficient_attention()
+            except Exception as e:
+                logging.warning(f"Could not enable xformers memory efficient attention: {e}")
 
         self.sample_neg_prompt = config.sample_neg_prompt