[Bugfix] Fix broadcasting logic for multi_modal_kwargs (#6836)

vllm/spec_decode/draft_model_runner.py

@@ -15,6 +15,7 @@ from vllm.config import (CacheConfig, DeviceConfig, LoadConfig, LoRAConfig,
                          ModelConfig, MultiModalConfig, ParallelConfig,
                          PromptAdapterConfig, SchedulerConfig)
 from vllm.logger import init_logger
+from vllm.multimodal import MultiModalInputs
 from vllm.sequence import (ExecuteModelRequest, IntermediateTensors,
                            SamplerOutput)
 from vllm.worker.model_runner import (ModelInputForGPUWithSamplingMetadata,
@@ -323,7 +324,8 @@ class TP1DraftModelRunner(ModelRunner):
                 kv_caches=kv_caches,
                 attn_metadata=model_input.attn_metadata,
                 intermediate_tensors=intermediate_tensors,
-                **multi_modal_kwargs,
+                **MultiModalInputs.as_kwargs(multi_modal_kwargs,
+                                             device=self.device),
             )
 
             # Compute the logits.