[Core] Automatically cast multi-modal input dtype (#18756)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

vllm/model_executor/models/deepseek_vl2.py

@@ -210,9 +210,7 @@ class DeepseekVL2MultiModalProcessor(
                 dict(prompt=prompt, **mm_data),
                 mm_kwargs,
             )
-            target_dtype = self.info.ctx.model_config.dtype
+            pixel_values = processed_outputs["pixel_values"]
-            pixel_values = processed_outputs.pop("pixel_values").to(
-                target_dtype)
             # split pixel values into patches corresponding to each image
             images_spatial_crop = processed_outputs["images_spatial_crop"]
             patches_per_image = [

vllm/model_executor/models/gemma3_mm.py

@@ -263,11 +263,6 @@ class Gemma3MultiModalProcessor(BaseMultiModalProcessor[Gemma3ProcessingInfo]):
             mm_data,
             mm_kwargs,
         )
-        if "pixel_values" in processed_outputs:
-            # Cast pixel values to model dtype already here,
-            # so we need to transfer less data to the GPU
-            processed_outputs["pixel_values"] = processed_outputs[
-                "pixel_values"].to(self.info.ctx.model_config.dtype)
 
         # HF processor pops the `num_crops` kwarg, which is needed by vLLM
         if (images := mm_data.get("images")) is not None:

vllm/multimodal/inputs.py

@@ -746,11 +746,17 @@ class MultiModalKwargs(UserDict[str, NestedTensors]):
         batched_inputs: BatchedTensorInputs,
         *,
         device: torch.types.Device,
+        dtype: Optional[torch.dtype] = None,
     ) -> BatchedTensorInputs:
         json_inputs = cast(JSONTree[torch.Tensor], batched_inputs)
 
+        def maybe_cast_dtype(x: torch.Tensor):
+            # This mimics the behavior of transformers.BatchFeature
+            return x.to(dtype=dtype) if x.is_floating_point() else x
+
         json_mapped = json_map_leaves(
-            lambda x: x.to(device, non_blocking=True),
+            # NOTE: Cast the dtype before sending it to device
+            lambda x: maybe_cast_dtype(x).to(device=device, non_blocking=True),
             json_inputs,
         )
 

vllm/spec_decode/draft_model_runner.py

@@ -294,8 +294,11 @@ class TP1DraftModelRunner(ModelRunnerWrapperBase):
                     inputs_embeds=None,
                     positions=model_input.input_positions,
                     intermediate_tensors=intermediate_tensors,
-                    **MultiModalKwargs.as_kwargs(multi_modal_kwargs,
+                    **MultiModalKwargs.as_kwargs(
-                                                 device=self.device),
+                        multi_modal_kwargs,
+                        dtype=self.model_runner.model_config.dtype,
+                        device=self.device,
+                    ),
                     **model_execute_kwargs,
                 )
 

vllm/v1/worker/gpu_model_runner.py

@@ -929,8 +929,11 @@ class GPUModelRunner(LoRAModelRunnerMixin):
         encoder_outputs = []
         for grouped_mm_inputs in grouped_mm_inputs_list:
             batched_mm_inputs = MultiModalKwargs.batch(grouped_mm_inputs)
-            batched_mm_inputs = MultiModalKwargs.as_kwargs(batched_mm_inputs,
+            batched_mm_inputs = MultiModalKwargs.as_kwargs(
-                                                           device=self.device)
+                batched_mm_inputs,
+                dtype=self.model_config.dtype,
+                device=self.device,
+            )
 
             # Run the encoder.
             # `curr_group_outputs` is either of the following:
@@ -1874,7 +1877,10 @@ class GPUModelRunner(LoRAModelRunnerMixin):
             batched_dummy_mm_inputs = MultiModalKwargs.batch(
                 [dummy_mm_kwargs] * max_num_mm_items)
             batched_dummy_mm_inputs = MultiModalKwargs.as_kwargs(
-                batched_dummy_mm_inputs, device=self.device)
+                batched_dummy_mm_inputs,
+                dtype=self.model_config.dtype,
+                device=self.device,
+            )
 
             # Run multimodal encoder.
             dummy_encoder_outputs = self.model.get_multimodal_embeddings(

vllm/v1/worker/tpu_model_runner.py

@@ -652,8 +652,11 @@ class TPUModelRunner(LoRAModelRunnerMixin):
         encoder_outputs = []
         for grouped_mm_inputs in grouped_mm_inputs_list:
             batched_mm_inputs = MultiModalKwargs.batch(grouped_mm_inputs)
-            batched_mm_inputs = MultiModalKwargs.as_kwargs(batched_mm_inputs,
+            batched_mm_inputs = MultiModalKwargs.as_kwargs(
-                                                           device=self.device)
+                batched_mm_inputs,
+                dtype=self.model_config.dtype,
+                device=self.device,
+            )
 
             # Run the encoder.
             # `curr_group_outputs` is either of the following:
@@ -1435,8 +1438,11 @@ class TPUModelRunner(LoRAModelRunnerMixin):
 
         batched_dummy_mm_inputs = MultiModalKwargs.batch([dummy_mm_kwargs] *
                                                          batch_size)
-        return MultiModalKwargs.as_kwargs(batched_dummy_mm_inputs,
+        return MultiModalKwargs.as_kwargs(
-                                          device=self.device)
+            batched_dummy_mm_inputs,
+            dtype=self.model_config.dtype,
+            device=self.device,
+        )
 
 
 def _get_req_paddings(min_req_size: int, max_req_size: int) -> list[int]:

vllm/worker/cpu_enc_dec_model_runner.py

@@ -297,8 +297,11 @@ class CPUEncoderDecoderModelRunner(
             model_input.encoder_input_tokens,
             "encoder_positions":
             model_input.encoder_input_positions,
-            **MultiModalKwargs.as_kwargs(model_input.multi_modal_kwargs or {},
+            **MultiModalKwargs.as_kwargs(
-                                         device=self.device),
+                model_input.multi_modal_kwargs or {},
+                dtype=self.model_config.dtype,
+                device=self.device,
+            ),
             "intermediate_tensors":
             intermediate_tensors,
         }

vllm/worker/cpu_model_runner.py

@@ -628,7 +628,10 @@ class CPUModelRunner(CPUModelRunnerBase[ModelInputForCPUWithSamplingMetadata]):
         multimodal_kwargs = {}
         if model_input.multi_modal_kwargs is not None:
             multimodal_kwargs = MultiModalKwargs.as_kwargs(
-                model_input.multi_modal_kwargs, device=self.device)
+                model_input.multi_modal_kwargs,
+                dtype=self.model_config.dtype,
+                device=self.device,
+            )
         execute_model_kwargs = {}
         if previous_hidden_states is not None:
             execute_model_kwargs.update(

vllm/worker/cpu_pooling_model_runner.py

@@ -50,8 +50,11 @@ class CPUPoolingModelRunner(
             model_input.input_tokens,
             "positions":
             model_input.input_positions,
-            **MultiModalKwargs.as_kwargs(model_input.multi_modal_kwargs or {},
+            **MultiModalKwargs.as_kwargs(
-                                         device=self.device),
+                model_input.multi_modal_kwargs or {},
+                dtype=self.model_config.dtype,
+                device=self.device,
+            ),
             **cross_enc_kwargs,
             "intermediate_tensors":
             intermediate_tensors,

vllm/worker/enc_dec_model_runner.py

@@ -202,9 +202,13 @@ class EncoderDecoderModelRunner(GPUModelRunnerBase[EncoderDecoderModelInput]):
                 encoder_input_ids=model_input.encoder_input_tokens,
                 encoder_positions=model_input.encoder_input_positions,
                 intermediate_tensors=intermediate_tensors,
-                **MultiModalKwargs.as_kwargs(multi_modal_kwargs,
+                **MultiModalKwargs.as_kwargs(
-                                             device=self.device),
+                    multi_modal_kwargs,
-                **seqlen_agnostic_kwargs)
+                    dtype=self.model_config.dtype,
+                    device=self.device,
+                ),
+                **seqlen_agnostic_kwargs,
+            )
 
         logits = self.model.compute_logits(hidden_or_intermediate_states,
                                            model_input.sampling_metadata)

vllm/worker/model_runner.py

@@ -1845,8 +1845,11 @@ class ModelRunner(GPUModelRunnerBase[ModelInputForGPUWithSamplingMetadata]):
                     inputs_embeds=model_input.inputs_embeds,
                     positions=model_input.input_positions,
                     intermediate_tensors=intermediate_tensors,
-                    **MultiModalKwargs.as_kwargs(multi_modal_kwargs,
+                    **MultiModalKwargs.as_kwargs(
-                                                 device=self.device),
+                        multi_modal_kwargs,
+                        dtype=self.model_config.dtype,
+                        device=self.device,
+                    ),
                     **seqlen_agnostic_kwargs,
                     **model_kwargs,
                 )

vllm/worker/multi_step_neuron_model_runner.py

@@ -70,8 +70,11 @@ class MultiStepNeuronModelRunner(NeuronModelRunner):
             input_ids=model_input.input_tokens,
             positions=model_input.input_positions,
             input_block_ids=model_input.input_block_ids,
-            **MultiModalKwargs.as_kwargs(model_input.multi_modal_kwargs or {},
+            **MultiModalKwargs.as_kwargs(
-                                         device=self.device),
+                model_input.multi_modal_kwargs or {},
+                dtype=self.model_config.dtype,
+                device=self.device,
+            ),
         )
 
         output = self.model.sample(

vllm/worker/multi_step_neuronx_distributed_model_runner.py

@@ -49,8 +49,11 @@ class MultiStepNeuronxDistributedModelRunner(NeuronxDistributedModelRunner):
             positions=model_input.input_positions,
             input_block_ids=model_input.input_block_ids,
             sampling_params=sampling_params,
-            **MultiModalKwargs.as_kwargs(model_input.multi_modal_kwargs or {},
+            **MultiModalKwargs.as_kwargs(
-                                         device=self.device),
+                model_input.multi_modal_kwargs or {},
+                dtype=self.model_config.dtype,
+                device=self.device,
+            ),
         )
 
         output = self.model.sample(

vllm/worker/neuron_model_runner.py

@@ -378,9 +378,11 @@ class NeuronModelRunner(ModelRunnerBase[ModelInputForNeuron]):
                 positions=model_input.input_positions,
                 input_block_ids=model_input.input_block_ids,
                 sampling_params=sampling_params,
-                **MultiModalKwargs.as_kwargs(model_input.multi_modal_kwargs
+                **MultiModalKwargs.as_kwargs(
-                                             or {},
+                    model_input.multi_modal_kwargs or {},
-                                             device=self.device),
+                    dtype=self.model_config.dtype,
+                    device=self.device,
+                ),
             )
         elif current_platform.use_transformers_neuronx():
             # [TODO] validate on-device sampling
@@ -389,9 +391,11 @@ class NeuronModelRunner(ModelRunnerBase[ModelInputForNeuron]):
                 input_ids=model_input.input_tokens,
                 positions=model_input.input_positions,
                 input_block_ids=model_input.input_block_ids,
-                **MultiModalKwargs.as_kwargs(model_input.multi_modal_kwargs
+                **MultiModalKwargs.as_kwargs(
-                                             or {},
+                    model_input.multi_modal_kwargs or {},
-                                             device=self.device),
+                    dtype=self.model_config.dtype,
+                    device=self.device,
+                ),
             )
 
         # Compute the logits only if the on-device sampling is turned off as

vllm/worker/pooling_model_runner.py

@@ -119,10 +119,14 @@ class PoolingModelRunner(
                 input_ids=model_input.input_tokens,
                 positions=model_input.input_positions,
                 intermediate_tensors=intermediate_tensors,
-                **MultiModalKwargs.as_kwargs(multi_modal_kwargs,
+                **MultiModalKwargs.as_kwargs(
-                                             device=self.device),
+                    multi_modal_kwargs,
+                    dtype=self.model_config.dtype,
+                    device=self.device,
+                ),
                 **cross_enc_kwargs,
-                **seqlen_agnostic_kwargs)
+                **seqlen_agnostic_kwargs,
+            )
 
         if (self.observability_config is not None
                 and self.observability_config.collect_model_forward_time):

vllm/worker/xpu_model_runner.py

@@ -562,9 +562,12 @@ class XPUModelRunner(ModelRunnerBase[ModelInputForXPUWithSamplingMetadata]):
                 input_ids=model_input.input_tokens,
                 positions=model_input.input_positions,
                 intermediate_tensors=intermediate_tensors,
-                **MultiModalKwargs.as_kwargs(model_input.multi_modal_kwargs
+                **MultiModalKwargs.as_kwargs(
-                                             or {},
+                    model_input.multi_modal_kwargs or {},
-                                             device=self.device))
+                    dtype=self.model_config.dtype,
+                    device=self.device,
+                ),
+            )
         # Compute the logits in the last pipeline stage.
         if not get_pp_group().is_last_rank:
             return hidden_or_intermediate_states