update load path

Author: gushiqiao

lightx2v/models/input_encoders/hf/qwen25/qwen25_vlforconditionalgeneration.py

@@ -88,10 +88,12 @@ def load(self):
         if not self.cpu_offload:
             self.text_encoder = self.text_encoder.to(AI_DEVICE)
 
-        self.tokenizer = Qwen2Tokenizer.from_pretrained(os.path.join(self.config["model_path"], "tokenizer"))
+        qwen25vl_tokenizer_path = self.config.get("tokenizer_path", os.path.join(self.config["model_path"], "tokenizer"))
+        self.tokenizer = Qwen2Tokenizer.from_pretrained(qwen25vl_tokenizer_path)
         if self.config["task"] == "i2i":
             self.image_processor = VaeImageProcessor(vae_scale_factor=self.config["vae_scale_factor"] * 2)
-            self.processor = Qwen2VLProcessor.from_pretrained(os.path.join(self.config["model_path"], "processor"))
+            qwen25vl_processor_path = self.config.get("processor_path", os.path.join(self.config["model_path"], "processor"))
+            self.processor = Qwen2VLProcessor.from_pretrained(qwen25vl_processor_path)
 
     def _extract_masked_hidden(self, hidden_states: torch.Tensor, mask: torch.Tensor):
         bool_mask = mask.bool()

lightx2v/models/schedulers/qwen_image/scheduler.py

@@ -287,7 +287,8 @@ class QwenImageScheduler(BaseScheduler):
     def __init__(self, config):
         super().__init__(config)
         self.config = config
-        self.scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(os.path.join(config["model_path"], "scheduler"))
+        scheduler_path = config.get("scheduler_path", os.path.join(config["model_path"], "scheduler"))
+        self.scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(scheduler_path)
         with open(os.path.join(config["model_path"], "scheduler", "scheduler_config.json"), "r") as f:
             self.scheduler_config = json.load(f)
         self.dtype = torch.bfloat16

lightx2v/models/video_encoders/hf/qwen_image/vae.py

@@ -33,7 +33,8 @@ def __init__(self, config):
         self.load()
 
     def load(self):
-        self.model = AutoencoderKLQwenImage.from_pretrained(os.path.join(self.config["model_path"], "vae")).to(self.device).to(self.dtype)
+        vae_path = self.config.get("vae_path", os.path.join(self.config["model_path"], "vae"))
+        self.model = AutoencoderKLQwenImage.from_pretrained(vae_path).to(self.device).to(self.dtype)
         self.image_processor = VaeImageProcessor(vae_scale_factor=self.config["vae_scale_factor"] * 2)
         with open(os.path.join(self.config["model_path"], "vae", "config.json"), "r") as f:
             vae_config = json.load(f)