Fix endpoint startup: load weights from upstream diffusers model ID

The HF Inference Toolkit passes the local endpoint repo dir as path, but that
dir has no model_index.json (only handler.py + requirements.txt), causing
OSError on startup. Always load from the upstream model ID instead.

handler.py

@@ -16,15 +16,25 @@ if device.type != 'cuda':
 dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
 
 class EndpointHandler():
+    # Upstream model that actually holds the weights. The HF Inference Toolkit
+    # passes the local endpoint repo dir (/repository) as `path`, but that repo
+    # only contains handler.py + requirements.txt — no model_index.json. So we
+    # always load from the upstream model ID, ignoring the passed-in path.
+    #
+    # NOTE: use the diffusers-format community repo, NOT tencent/HunyuanVideo —
+    # the official repo ships raw .pt checkpoints with no model_index.json and
+    # is not loadable via HunyuanVideoPipeline.from_pretrained.
+    MODEL_ID = "hunyuanvideo-community/HunyuanVideo"
+
     def __init__(self, path=""):
         """
         Initialize HunyuanVideo pipeline for video generation.
         Args:
-            path: Path to model weights (will be tencent/HunyuanVideo)
+            path: Local endpoint dir from the toolkit (ignored — has no weights).
         """
-        # Load HunyuanVideo pipeline
+        # Load HunyuanVideo pipeline from the upstream model ID.
         self.pipe = HunyuanVideoPipeline.from_pretrained(
-            path if path else "tencent/HunyuanVideo",
+            self.MODEL_ID,
             torch_dtype=dtype
         )
         self.pipe.to(device)