[Core] Cast multimodal input in hf processor (vllm-project#18862)

Signed-off-by: Lukas Geiger <[email protected]>

Author: lgeiger

vllm/inputs/registry.py

@@ -4,9 +4,12 @@
 from dataclasses import dataclass
 from typing import TYPE_CHECKING, Any, NamedTuple, Optional, Union
 
+import torch
 from transformers import BatchFeature, PretrainedConfig, ProcessorMixin
 from typing_extensions import TypeVar
 
+from vllm.jsontree import JSONTree, json_map_leaves
+from vllm.logger import init_logger
 from vllm.transformers_utils.processor import cached_processor_from_config
 from vllm.transformers_utils.tokenizer import AnyTokenizer
 from vllm.utils import resolve_mm_processor_kwargs
@@ -21,6 +24,8 @@
 _C = TypeVar("_C", bound=PretrainedConfig, default=PretrainedConfig)
 _P = TypeVar("_P", bound=ProcessorMixin, default=ProcessorMixin)
 
+logger = init_logger(__name__)
+
 
 @dataclass(frozen=True)
 class InputContext:
@@ -134,7 +139,7 @@ def call_hf_processor(
         hf_processor: ProcessorMixin,
         data: Mapping[str, object],
         kwargs: Mapping[str, object] = {},
-    ) -> BatchFeature:
+    ) -> Union[BatchFeature, JSONTree]:
         """
         Call `hf_processor` on the prompt `data`
         (text, image, audio...) with configurable options `kwargs`.
@@ -154,8 +159,25 @@ def call_hf_processor(
             allow_var_kwargs=True,
         )
 
+        def maybe_cast_dtype(x):
+            # This mimics the behavior of transformers.BatchFeature
+            if isinstance(x, torch.Tensor) and x.is_floating_point():
+                return x.to(dtype=self.model_config.dtype)
+            return x
+
         try:
-            return hf_processor(**data, **merged_kwargs, return_tensors="pt")
+            output = hf_processor(**data, **merged_kwargs, return_tensors="pt")
+            # this emulates output.to(dtype=self.model_config.dtype)
+            cast_output = json_map_leaves(maybe_cast_dtype, output)
+            if isinstance(output, BatchFeature):
+                return BatchFeature(cast_output)
+
+            logger.warning_once(
+                f"{type(hf_processor).__name__} did not return `BatchFeature`. "
+                "Make sure to match the behaviour of `ProcessorMixin` when "
+                "implementing custom processors.")
+            return cast_output
+
         except Exception as exc:
             msg = (f"Failed to apply {type(hf_processor).__name__} "
                    f"on data={data} with kwargs={merged_kwargs}")

vllm/multimodal/inputs.py

@@ -747,17 +747,11 @@ def as_kwargs(
         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(
-            # NOTE: Cast the dtype before sending it to device
-            lambda x: maybe_cast_dtype(x).to(device=device, non_blocking=True),
+            lambda x: x.to(device=device, non_blocking=True),
             json_inputs,
         )
 

vllm/spec_decode/draft_model_runner.py

@@ -297,7 +297,6 @@ def execute_model(
                     intermediate_tensors=intermediate_tensors,
                     **MultiModalKwargs.as_kwargs(
                         multi_modal_kwargs,
-                        dtype=self.model_runner.model_config.dtype,
                         device=self.device,
                     ),
                     **model_execute_kwargs,

vllm/v1/worker/gpu_model_runner.py

@@ -957,7 +957,6 @@ def _execute_mm_encoder(self, scheduler_output: "SchedulerOutput"):
             batched_mm_inputs = MultiModalKwargs.batch(grouped_mm_inputs)
             batched_mm_inputs = MultiModalKwargs.as_kwargs(
                 batched_mm_inputs,
-                dtype=self.model_config.dtype,
                 device=self.device,
             )
 
@@ -1951,7 +1950,6 @@ def profile_run(self) -> None:
                 [dummy_mm_kwargs] * max_num_mm_items)
             batched_dummy_mm_inputs = MultiModalKwargs.as_kwargs(
                 batched_dummy_mm_inputs,
-                dtype=self.model_config.dtype,
                 device=self.device,
             )
 

vllm/v1/worker/tpu_model_runner.py

@@ -718,7 +718,6 @@ def _execute_mm_encoder(self, scheduler_output: "SchedulerOutput"):
             batched_mm_inputs = MultiModalKwargs.batch(grouped_mm_inputs)
             batched_mm_inputs = MultiModalKwargs.as_kwargs(
                 batched_mm_inputs,
-                dtype=self.model_config.dtype,
                 device=self.device,
             )
 
@@ -1560,7 +1559,6 @@ def _get_mm_dummy_batch(self, modality: str,
                                                          batch_size)
         return MultiModalKwargs.as_kwargs(
             batched_dummy_mm_inputs,
-            dtype=self.model_config.dtype,
             device=self.device,
         )
 

vllm/worker/cpu_enc_dec_model_runner.py

@@ -300,7 +300,6 @@ def execute_model(
             model_input.encoder_input_positions,
             **MultiModalKwargs.as_kwargs(
                 model_input.multi_modal_kwargs or {},
-                dtype=self.model_config.dtype,
                 device=self.device,
             ),
             "intermediate_tensors":

vllm/worker/cpu_model_runner.py

@@ -630,7 +630,6 @@ def execute_model(
         if model_input.multi_modal_kwargs is not None:
             multimodal_kwargs = MultiModalKwargs.as_kwargs(
                 model_input.multi_modal_kwargs,
-                dtype=self.model_config.dtype,
                 device=self.device,
             )
         execute_model_kwargs = {}

vllm/worker/cpu_pooling_model_runner.py

@@ -53,7 +53,6 @@ def execute_model(
             model_input.input_positions,
             **MultiModalKwargs.as_kwargs(
                 model_input.multi_modal_kwargs or {},
-                dtype=self.model_config.dtype,
                 device=self.device,
             ),
             **cross_enc_kwargs,

vllm/worker/enc_dec_model_runner.py

@@ -205,7 +205,6 @@ def execute_model(
                 intermediate_tensors=intermediate_tensors,
                 **MultiModalKwargs.as_kwargs(
                     multi_modal_kwargs,
-                    dtype=self.model_config.dtype,
                     device=self.device,
                 ),
                 **seqlen_agnostic_kwargs,

vllm/worker/model_runner.py

@@ -1848,7 +1848,6 @@ def execute_model(
                     intermediate_tensors=intermediate_tensors,
                     **MultiModalKwargs.as_kwargs(
                         multi_modal_kwargs,
-                        dtype=self.model_config.dtype,
                         device=self.device,
                     ),
                     **seqlen_agnostic_kwargs,