Fix CLAP from_pretrained daemonic process error in ZeroGPU worker

Monkey-patch multiprocessing.Process.start to run inline before calling
ClapTextModelWithProjection.from_pretrained(), restoring it immediately
after. Prevents transformers from forking a subprocess for safe-tensor
verification, which fails in ZeroGPU daemonic workers.

HunyuanVideo-Foley/hunyuanvideo_foley/utils/model_utils.py

@@ -319,9 +319,22 @@ def load_model(model_path, config_path, device, enable_offload=False, model_size
         logger.info("SigLIP2 model and preprocessing pipeline loaded successfully")
 
         # clap text-encoder
+        # Transformers' from_pretrained() spawns a subprocess for safe-tensor
+        # verification, which fails inside ZeroGPU's daemonic worker process.
+        # Monkey-patch multiprocessing.Process.start to run the target inline
+        # instead of forking, then restore it immediately after.
         logger.info("Loading CLAP text encoder...")
-        clap_tokenizer = AutoTokenizer.from_pretrained("laion/larger_clap_general")
-        clap_model = ClapTextModelWithProjection.from_pretrained("laion/larger_clap_general").to(device)
+        import multiprocessing as _mp
+        _orig_start = _mp.Process.start
+        def _inline_start(self):
+            if self._target:
+                self._target(*self._args, **self._kwargs)
+        _mp.Process.start = _inline_start
+        try:
+            clap_tokenizer = AutoTokenizer.from_pretrained("laion/larger_clap_general")
+            clap_model = ClapTextModelWithProjection.from_pretrained("laion/larger_clap_general").to(device)
+        finally:
+            _mp.Process.start = _orig_start
         logger.info("CLAP tokenizer and model loaded successfully")
 
         # syncformer