[None][feat] Support SharedTensor on MultimodalParams

tensorrt_llm/_torch/pyexecutor/model_engine.py

@@ -1227,11 +1227,13 @@ def _prepare_tp_inputs(
             multimodal_params = MultimodalParams(
                 multimodal_data=request.py_multimodal_data,
                 multimodal_runtime=py_multimodal_runtime)
-            multimodal_params.to_device("multimodal_data",
-                                        "cuda",
-                                        pin_memory=True)
 
             if multimodal_params.has_content():
+                multimodal_params.to_device("multimodal_data",
+                                            "cuda",
+                                            pin_memory=True)
+                #re-assign the multimodal_data to the request after to_device for generation requests
+                request.py_multimodal_data = multimodal_params.multimodal_data
                 multimodal_params_list.append(multimodal_params)
 
             request.py_batch_idx = request.py_seq_slot
@@ -1265,10 +1267,12 @@ def _prepare_tp_inputs(
             multimodal_params = MultimodalParams(
                 multimodal_data=request.py_multimodal_data)
             multimodal_params.strip_for_generation()
-            multimodal_params.to_device("multimodal_data",
-                                        "cuda",
-                                        pin_memory=True)
             if multimodal_params.has_content():
+                multimodal_params.to_device("multimodal_data",
+                                            "cuda",
+                                            pin_memory=True)
+                # re-assign the multimodal_data to the request after strip_for_generation for another generation request,
+                request.py_multimodal_data = multimodal_params.multimodal_data
                 multimodal_params_list.append(multimodal_params)
         extend_requests += extend_dummy_requests
 

tensorrt_llm/executor/worker.py

@@ -486,7 +486,7 @@ def _deduce_max_tokens(request: GenerationRequest,
                 lora_config=lora_config,
                 prompt_tuning_config=prompt_tuning_config,
                 multimodal_input=multimodal_input,
-                #NOTE: `multimodal_embedding` and `mrope_config` will be in MultimodalParams.multimodal_data. And this will be handled below by `py_multimodal_data`.
+                # NOTE: `multimodal_embedding` and `mrope_config` will be in MultimodalParams.multimodal_data. And this will be handled below by `py_multimodal_data`.
                 multimodal_embedding=None,
                 mrope_config=None,
                 logits_post_processor_name=(
@@ -502,17 +502,8 @@ def _deduce_max_tokens(request: GenerationRequest,
 
             if self._is_pytorch_backend and request.multimodal_params is not None:
                 if request.multimodal_params.multimodal_data is not None:
-                    # Convert back to tensor, as opposite to `to_handle` in `llm.generate_async`
-                    # for values with non-selected keys, it's no-op
-                    request.multimodal_params.to_tensor(
-                        "multimodal_data", key="multimodal_embedding")
-                    embedding = request.multimodal_params.multimodal_data.get(
-                        "multimodal_embedding")
-                    if embedding is not None and embedding.is_cuda:
-                        # make sure the embedding resides on the local device
-                        request.multimodal_params.multimodal_data[
-                            "multimodal_embedding"] = embedding.to("cuda")
-
+                    # NOTE: Deserialize SharedTensor handle to actual tensor
+                    request.multimodal_params.to_tensor("multimodal_data")
                     executor_request.py_multimodal_data = request.multimodal_params.multimodal_data
 
             if self._is_pytorch_backend and request.sampling_params.logits_processor: