[BugFix][HunyuanImage3] Set MRoPE dynamic_arg_dims so graph mode can compile (#3630)

Signed-off-by: TaffyOfficial <2324465096@qq.com>
Co-authored-by: TaffyOfficial <2324465096@qq.com>
Co-authored-by: Codex <codex@openai.com>

vllm_omni/model_executor/models/hunyuan_image3/hunyuan_image3.py

@@ -96,7 +96,16 @@ from vllm_omni.model_executor.models.hunyuan_image3.siglip2 import LightProjecto
 logger = init_logger(__name__)
 
 
-@support_torch_compile
+@support_torch_compile(
+    dynamic_arg_dims={
+        "input_ids": 0,
+        # positions is (3, num_tokens) for mRoPE, so the token dimension is
+        # the last dimension rather than dim 0.
+        "positions": -1,
+        "intermediate_tensors": 0,
+        "inputs_embeds": 0,
+    }
+)
 class HunyuanModel(HunYuanModel):
     def _split_qkv_weight(self, qkv: torch.Tensor):
         num_attention_heads = self.config.num_attention_heads