update

gradio_multi_gpu.sh

@@ -27,7 +27,6 @@ CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES /primus_xpfs_workspace_T04/huangshiji
     --convert_model_dtype \
     --infer_frames 48 \
     --load_lora \
-    --lora_path "/primus_xpfs_workspace_T04/huangyubo/Causvid/checkpoints/1005_s2v_causal_sft_1_4_0/ckpt/merged_model_step_step_25000/only_lora.pt" \
     --lora_path_dmd "/primus_xpfs_workspace_T04/huangyubo/LongLive/1027_s2v_selfforcing_1_13_2/ckptcopy/step_2500_generator_lora.pt" \
     --sample_steps 4 \
     --sample_guide_scale 0 \

gradio_single_gpu.sh

@@ -19,7 +19,6 @@ CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES /primus_xpfs_workspace_T04/huangshiji
     --convert_model_dtype \
     --infer_frames 48 \
     --load_lora \
-    --lora_path "/primus_xpfs_workspace_T04/huangyubo/Causvid/checkpoints/1005_s2v_causal_sft_1_4_0/ckpt/merged_model_step_step_25000/only_lora.pt" \
     --lora_path_dmd "/primus_xpfs_workspace_T04/huangyubo/LongLive/1027_s2v_selfforcing_1_13_2/ckptcopy/step_2500_generator_lora.pt" \
     --sample_steps 4 \
     --sample_guide_scale 0 \

infinite_inference_multi_gpu.sh

@@ -9,12 +9,11 @@ CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES /primus_xpfs_workspace_T04/huangshiji
      --training_config liveavatar/configs/s2v_causal_sft.yaml \
      --offload_model False \
      --convert_model_dtype \
-     --prompt "A close-up, black and white video of a 60-70 year old male digital human, resembling Albert Einstein. He has a surprised and curious expression with raised eyebrows and slightly parted lips, as if about to speak. He wears a dark suit, white shirt, and dark tie. In the dimly lit background to his right, a blackboard with a chalk circle and symbols like 'L' and 'K' is visible. The scene has a thoughtful, intelligent, and slightly whimsical atmosphere, rendered in a classic black and white film style."  \
-     --image "examples/demo/case1.jpg" \
-     --audio "examples/demo/case1_talker.wav" \
+     --prompt "A stout, cheerful dwarf with a magnificent braided beard adorned with metal rings, wearing a heavy leather apron. He's standing in his fiery, cluttered forge, laughing heartily as he explains the mastery of his craft, holding up a glowing hammer. Style of Blizzard Entertainment cinematics (like World of Warcraft), warm, dynamic lighting from the forge."  \
+     --image "examples/dwarven_blacksmith.jpg" \
+     --audio "examples/dwarven_blacksmith.wav" \
      --infer_frames 48 \
      --load_lora \
-     --lora_path "/primus_xpfs_workspace_T04/huangyubo/Causvid/checkpoints/1005_s2v_causal_sft_1_4_0/ckpt/merged_model_step_step_25000/only_lora.pt"    \
      --lora_path_dmd "/primus_xpfs_workspace_T04/huangyubo/LongLive/1027_s2v_selfforcing_1_13_2/ckptcopy/step_2500_generator_lora.pt" \
      --sample_steps 4 \
      --sample_guide_scale 0 \

infinite_inference_single_gpu.sh

@@ -9,12 +9,11 @@ CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES /primus_xpfs_workspace_T04/huangshiji
      --training_config liveavatar/configs/s2v_causal_sft.yaml \
      --offload_model True \
      --convert_model_dtype \
-     --prompt "A vibrant 3D anime style avatar of a young blonde woman with high pigtails, wearing a black Gothic Lolita dress with an off-shoulder neckline, lace choker, and gloves, speaking playfully to the camera with expressive, exaggerated hand gestures and a slight smirk. The camera is steady, and the background is a dark, simple studio."  \
-     --image "examples/demo/case1.jpg" \
-     --audio "examples/talk.wav" \
+     --prompt "A stout, cheerful dwarf with a magnificent braided beard adorned with metal rings, wearing a heavy leather apron. He's standing in his fiery, cluttered forge, laughing heartily as he explains the mastery of his craft, holding up a glowing hammer. Style of Blizzard Entertainment cinematics (like World of Warcraft), warm, dynamic lighting from the forge."  \
+     --image "examples/dwarven_blacksmith.jpg" \
+     --audio "examples/dwarven_blacksmith.wav" \
      --infer_frames 48 \
      --load_lora \
-     --lora_path "/primus_xpfs_workspace_T04/huangyubo/Causvid/checkpoints/1005_s2v_causal_sft_1_4_0/ckpt/merged_model_step_step_25000/only_lora.pt"    \
      --lora_path_dmd "/primus_xpfs_workspace_T04/huangyubo/LongLive/1027_s2v_selfforcing_1_13_2/ckptcopy/step_2500_generator_lora.pt" \
      --sample_steps 4 \
      --sample_guide_scale 0 \

minimal_inference/batch_eval.py

@@ -619,41 +619,22 @@ def generate(args, training_settings):
             single_gpu=args.single_gpu,
             offload_kv_cache=args.offload_kv_cache,
         )
-        if args.load_lora and args.lora_path is not None:
-            print(f'Use LoRA: lora path: {args.lora_path}, lora rank:', training_settings['lora_rank'] ,", lora alpha: ",training_settings['lora_alpha'])
+        if args.load_lora and args.lora_path_dmd is not None:
+            print(f'Use LoRA: lora path: {args.lora_path_dmd}, lora rank:', training_settings['lora_rank'] ,", lora alpha: ",training_settings['lora_alpha'])
             # for version <1.5.0, using ckpt not merged 还需要把ckpt path改了
-            if args.using_merged_ckpt is False:
-                wan_s2v.add_lora_to_model(
-                        wan_s2v.noise_model,
-                        lora_rank=training_settings['lora_rank'],
-                        lora_alpha=training_settings['lora_alpha'],
-                        lora_target_modules=training_settings['lora_target_modules'],
-                        init_lora_weights=training_settings['init_lora_weights'],
-                        pretrained_lora_path=args.lora_path,
-                        load_lora_weight_only=False,
-                    )
 
-                if args.lora_path_dmd is not None:
-                    wan_s2v.add_lora_to_model(
-                        wan_s2v.noise_model,
-                        lora_rank=training_settings['lora_rank'],
-                        lora_alpha=training_settings['lora_alpha'],
-                        lora_target_modules=training_settings['lora_target_modules'],
-                        init_lora_weights=training_settings['init_lora_weights'],
-                        pretrained_lora_path=args.lora_path_dmd,
-                        load_only=True,
-                    )
-            else:
-                # for version >=1.5.0, using merged ckpt,还需要把ckpt path改了
-                if args.lora_path_dmd is not None:
-                    wan_s2v.add_lora_to_model(
-                        wan_s2v.noise_model,
-                        lora_rank=256,
-                        lora_alpha=256,
-                        lora_target_modules=training_settings['lora_target_modules'],
-                        init_lora_weights=training_settings['init_lora_weights'],
-                        pretrained_lora_path=args.lora_path_dmd
-                    )
+
+            if args.lora_path_dmd is not None:
+                wan_s2v.add_lora_to_model(
+                    wan_s2v.noise_model,
+                    lora_rank=training_settings['lora_rank'],
+                    lora_alpha=training_settings['lora_alpha'],
+                    lora_target_modules=training_settings['lora_target_modules'],
+                    init_lora_weights=training_settings['init_lora_weights'],
+                    pretrained_lora_path=args.lora_path_dmd,
+                    load_lora_weight_only=False,
+                )
+
     else:
         assert False, "Only s2v is supported for now."
 

minimal_inference/gradio_app.py

@@ -365,40 +365,22 @@ def initialize_pipeline(args, training_settings):
         )
         
         # Load LoRA if specified
-        if args.load_lora and args.lora_path is not None:
-            logging.info(f'Loading LoRA: path={args.lora_path}, rank={training_settings["lora_rank"]}, alpha={training_settings["lora_alpha"]}')
+        if args.load_lora and args.lora_path_dmd is not None:
+            logging.info(f'Loading LoRA: path={args.lora_path_dmd}, rank={training_settings["lora_rank"]}, alpha={training_settings["lora_alpha"]}')
             
-            if args.using_merged_ckpt is False:
+
+                
+            if args.lora_path_dmd is not None:
                 wan_s2v.add_lora_to_model(
                     wan_s2v.noise_model,
                     lora_rank=training_settings['lora_rank'],
                     lora_alpha=training_settings['lora_alpha'],
                     lora_target_modules=training_settings['lora_target_modules'],
                     init_lora_weights=training_settings['init_lora_weights'],
-                    pretrained_lora_path=args.lora_path,
+                    pretrained_lora_path=args.lora_path_dmd,
                     load_lora_weight_only=False,
                 )
-                
-                if args.lora_path_dmd is not None:
-                    wan_s2v.add_lora_to_model(
-                        wan_s2v.noise_model,
-                        lora_rank=training_settings['lora_rank'],
-                        lora_alpha=training_settings['lora_alpha'],
-                        lora_target_modules=training_settings['lora_target_modules'],
-                        init_lora_weights=training_settings['init_lora_weights'],
-                        pretrained_lora_path=args.lora_path_dmd,
-                        load_only=True,
-                    )
-            else:
-                if args.lora_path_dmd is not None:
-                    wan_s2v.add_lora_to_model(
-                        wan_s2v.noise_model,
-                        lora_rank=256,
-                        lora_alpha=256,
-                        lora_target_modules=training_settings['lora_target_modules'],
-                        init_lora_weights=training_settings['init_lora_weights'],
-                        pretrained_lora_path=args.lora_path_dmd
-                    )
+
         
         wan_s2v_pipeline = wan_s2v
         global_args = args

minimal_inference/s2v_streaming_interact.py

@@ -456,41 +456,21 @@ def generate(args, training_settings):
             single_gpu=args.single_gpu,
             offload_kv_cache=args.offload_kv_cache,
         )
-        if args.load_lora and args.lora_path is not None:
-            print(f'Use LoRA: lora path: {args.lora_path}, lora rank:', training_settings['lora_rank'] ,", lora alpha: ",training_settings['lora_alpha'])
-            # for version <1.5.0, using ckpt not merged 还需要把ckpt path改了
-            if args.using_merged_ckpt is False:
-                wan_s2v.add_lora_to_model(
-                        wan_s2v.noise_model,
-                        lora_rank=training_settings['lora_rank'],
-                        lora_alpha=training_settings['lora_alpha'],
-                        lora_target_modules=training_settings['lora_target_modules'],
-                        init_lora_weights=training_settings['init_lora_weights'],
-                        pretrained_lora_path=args.lora_path,
-                        load_lora_weight_only=False,
-                    )
+        if args.load_lora and args.lora_path_dmd is not None:
+            print(f'Use LoRA: lora path: {args.lora_path_dmd}, lora rank:', training_settings['lora_rank'] ,", lora alpha: ",training_settings['lora_alpha'])
 
-                if args.lora_path_dmd is not None:
-                    wan_s2v.add_lora_to_model(
-                        wan_s2v.noise_model,
-                        lora_rank=training_settings['lora_rank'],
-                        lora_alpha=training_settings['lora_alpha'],
-                        lora_target_modules=training_settings['lora_target_modules'],
-                        init_lora_weights=training_settings['init_lora_weights'],
-                        pretrained_lora_path=args.lora_path_dmd,
-                        load_only=True,
-                    )
-            else:
-                # for version >=1.5.0, using merged ckpt,还需要把ckpt path改了
-                if args.lora_path_dmd is not None:
-                    wan_s2v.add_lora_to_model(
-                        wan_s2v.noise_model,
-                        lora_rank=256,
-                        lora_alpha=256,
-                        lora_target_modules=training_settings['lora_target_modules'],
-                        init_lora_weights=training_settings['init_lora_weights'],
-                        pretrained_lora_path=args.lora_path_dmd
-                    )
+
+            if args.lora_path_dmd is not None:
+                wan_s2v.add_lora_to_model(
+                    wan_s2v.noise_model,
+                    lora_rank=training_settings['lora_rank'],
+                    lora_alpha=training_settings['lora_alpha'],
+                    lora_target_modules=training_settings['lora_target_modules'],
+                    init_lora_weights=training_settings['init_lora_weights'],
+                    pretrained_lora_path=args.lora_path_dmd,
+                    load_lora_weight_only=False,
+                )
+            
 
         # Prepare video path for SAM2 processing (will be handled in pipeline)
         logging.info(f"Generating video ...")