[Input] Remove unused prompt field (vllm-project#26097)

Signed-off-by: DarkLight1337 <[email protected]>

Author: DarkLight1337

tests/models/multimodal/processing/test_transformers.py

@@ -37,4 +37,5 @@ def test_multimodal_processor(model_id):
         hf_processor_mm_kwargs={},
     )
 
-    assert str_processed_inputs["prompt"] == ids_processed_inputs["prompt"]
+    assert (str_processed_inputs["prompt_token_ids"]
+            == ids_processed_inputs["prompt_token_ids"])

vllm/engine/protocol.py

@@ -94,10 +94,15 @@ async def beam_search(
         #    this happens again in generation, so the double expansion causes
         #    a mismatch.
         # TODO - would be ideal to handle this more gracefully.
-        prompt_token_ids = prompt.get("prompt_token_ids")
-        multi_modal_data = prompt.get("multi_modal_data")
+        if isinstance(prompt, str):
+            prompt_text = prompt
+            prompt_token_ids = []
+            multi_modal_data = None
+        else:
+            prompt_text = prompt.get("prompt")
+            prompt_token_ids = prompt.get("prompt_token_ids", [])
+            multi_modal_data = prompt.get("multi_modal_data")
 
-        prompt_text = processed_inputs.get("prompt")
         mm_processor_kwargs = processed_inputs.get("mm_processor_kwargs")
 
         tokenized_length = len(prompt_token_ids)

vllm/inputs/data.py

@@ -205,11 +205,6 @@ class TokenInputs(TypedDict):
     prompt_token_ids: list[int]
     """The token IDs of the prompt."""
 
-    prompt: NotRequired[str]
-    """
-    The original prompt text corresponding to the token IDs, if available.
-    """
-
     cache_salt: NotRequired[str]
     """
     Optional cache salt to be used for prefix caching.
@@ -218,15 +213,12 @@ class TokenInputs(TypedDict):
 
 def token_inputs(
     prompt_token_ids: list[int],
-    prompt: Optional[str] = None,
     cache_salt: Optional[str] = None,
 ) -> TokenInputs:
     """Construct [`TokenInputs`][vllm.inputs.data.TokenInputs] from optional
     values."""
     inputs = TokenInputs(type="token", prompt_token_ids=prompt_token_ids)
 
-    if prompt is not None:
-        inputs["prompt"] = prompt
     if cache_salt is not None:
         inputs["cache_salt"] = cache_salt
 

vllm/inputs/preprocess.py

@@ -16,9 +16,10 @@
 from vllm.transformers_utils.tokenizer import AnyTokenizer
 
 from .data import (DecoderOnlyInputs, EmbedsInputs, EmbedsPrompt,
-                   EncoderDecoderInputs, ProcessorInputs, PromptType,
-                   SingletonInputs, SingletonPrompt, TextPrompt, TokenInputs,
-                   TokensPrompt, embeds_inputs, token_inputs)
+                   EncoderDecoderInputs, ExplicitEncoderDecoderPrompt,
+                   ProcessorInputs, PromptType, SingletonInputs,
+                   SingletonPrompt, TextPrompt, TokenInputs, TokensPrompt,
+                   embeds_inputs, token_inputs)
 from .parse import is_explicit_encoder_decoder_prompt, parse_singleton_prompt
 
 logger = init_logger(__name__)
@@ -322,7 +323,7 @@ def _process_tokens(
                 mm_uuids=mm_uuids,
             )
         else:
-            inputs = token_inputs(prompt_token_ids=prompt_token_ids)
+            inputs = token_inputs(prompt_token_ids)
 
         if cache_salt := parsed_content.get("cache_salt"):
             inputs["cache_salt"] = cache_salt
@@ -352,10 +353,7 @@ def _process_text(
                 prompt_text,
                 tokenization_kwargs=tokenization_kwargs,
             )
-            inputs = token_inputs(
-                prompt=prompt_text,
-                prompt_token_ids=prompt_token_ids,
-            )
+            inputs = token_inputs(prompt_token_ids)
 
         if cache_salt := parsed_content.get("cache_salt"):
             inputs["cache_salt"] = cache_salt
@@ -473,22 +471,17 @@ def _split_enc_dec_mm_inputs(
         decoder_inputs: SingletonInputs
 
         if inputs["type"] == "multimodal":  # Multimodal data inputs
-            if not ("encoder_prompt" in inputs
-                    and "encoder_prompt_token_ids" in inputs):
+            if "encoder_prompt_token_ids" not in inputs:
                 raise RuntimeError("You should register an encoder-decoder "
                                    "multi-modal processor for encoder-decoder "
                                    "models.")
             inputs = cast(MultiModalEncDecInputs, inputs)
 
-            encoder_inputs = token_inputs(
-                prompt=inputs["encoder_prompt"],
-                prompt_token_ids=inputs["encoder_prompt_token_ids"],
-            )
+            encoder_inputs = token_inputs(inputs["encoder_prompt_token_ids"])
 
             decoder_prompt_inputs = decoder_inputs_to_override or inputs
             decoder_inputs = MultiModalInputs(
                 type="multimodal",
-                prompt=decoder_prompt_inputs.get("prompt", ""),
                 prompt_token_ids=decoder_prompt_inputs["prompt_token_ids"],
                 mm_kwargs=inputs["mm_kwargs"],
                 mm_hashes=inputs["mm_hashes"],
@@ -498,7 +491,7 @@ def _split_enc_dec_mm_inputs(
                 decoder_inputs["cache_salt"] = cache_salt
 
         elif inputs["type"] == "token":  # Text-only inputs
-            encoder_inputs = token_inputs(prompt="", prompt_token_ids=[])
+            encoder_inputs = token_inputs(prompt_token_ids=[])
             decoder_inputs = decoder_inputs_to_override or inputs
         else:
             assert_never(inputs)  # type: ignore[arg-type]
@@ -549,12 +542,14 @@ def _process_encoder_decoder_prompt(
         decoder_inputs: Optional[SingletonInputs]
 
         if is_explicit_encoder_decoder_prompt(prompt):
+            # `cast` is needed for mypy, but not pyright
+            prompt_ = cast(ExplicitEncoderDecoderPrompt, prompt)
             encoder_inputs = self._prompt_to_llm_inputs(
-                prompt["encoder_prompt"],
+                prompt_["encoder_prompt"],
                 tokenization_kwargs=tokenization_kwargs,
                 mm_uuids=mm_uuids,
             )
-            if (decoder_input := prompt["decoder_prompt"]) is None:
+            if (decoder_input := prompt_["decoder_prompt"]) is None:
                 decoder_inputs = None
             else:
                 decoder_inputs = self._prompt_to_llm_inputs(decoder_input)
@@ -565,8 +560,9 @@ def _process_encoder_decoder_prompt(
                     self._split_enc_dec_mm_inputs(encoder_inputs,
                                                   decoder_inputs))
         else:
+            # `cast` is needed for mypy, but not pyright
             inputs = self._prompt_to_llm_inputs(
-                prompt,
+                cast(SingletonPrompt, prompt),
                 tokenization_kwargs=tokenization_kwargs,
                 mm_uuids=mm_uuids,
             )
@@ -641,8 +637,9 @@ def preprocess(
                              "to decoder-only models")
 
         # Decoder-only operation
+        # `cast` is needed for mypy, but not pyright
         return self._process_decoder_only_prompt(
-            prompt,
+            cast(SingletonPrompt, prompt),
             tokenization_kwargs=tokenization_kwargs,
             mm_uuids=mm_uuids,
         )

vllm/model_executor/models/llava.py

@@ -778,7 +778,7 @@ def get_replacement_mantis(item_idx: int):
             )
         ], mm_item_counts)
 
-        prompt_ids, prompt, _ = self._apply_prompt_updates(
+        prompt_ids, _ = self._apply_prompt_updates(
             result["prompt_token_ids"],
             mantis_mm_repls,
         )
@@ -798,7 +798,6 @@ def get_replacement_mantis(item_idx: int):
 
         return MultiModalInputs(
             type="multimodal",
-            prompt=prompt,
             prompt_token_ids=prompt_ids,
             mm_kwargs=mm_kwargs,
             mm_hashes=mm_hashes,

vllm/model_executor/models/paligemma.py

@@ -219,7 +219,6 @@ def apply(
         if len(prompt_token_ids) and prompt_token_ids[-1] != newline_token_id:
             prompt_token_ids.append(newline_token_id)
             mm_inputs["prompt_token_ids"] = prompt_token_ids
-            mm_inputs["prompt"] += newline_prompt
 
         return mm_inputs
 

vllm/model_executor/models/phi3v.py

@@ -461,7 +461,7 @@ def _apply_prompt_updates(
         self,
         token_ids: list[int],
         mm_prompt_updates: MultiModalPromptUpdates,
-    ) -> tuple[list[int], str, Mapping[str, list[PlaceholderFeaturesInfo]]]:
+    ) -> tuple[list[int], Mapping[str, list[PlaceholderFeaturesInfo]]]:
         # align to hf behavior when there are images
         if len(mm_prompt_updates):
             tokenizer = self.info.get_tokenizer()
@@ -496,14 +496,14 @@ def _apply_prompt_updates(
                 for ele in sublist for e in ele
             ]
 
-        token_ids, text, placeholders = super()._apply_prompt_updates(
+        token_ids, placeholders = super()._apply_prompt_updates(
             token_ids=token_ids,
             mm_prompt_updates=mm_prompt_updates,
         )
 
         # Keep the behavior in line with HF processor
-        if text.startswith("<s> <|image|>"):
-            text = text.replace("<s> <|image|>", "<s><|image|>", 1)
+        if token_ids[:2] == tokenizer.encode("<s> <|image|>",
+                                             add_special_tokens=False):
             token_ids = [token_ids[0], *token_ids[2:]]
             placeholders = {
                 modality: [
@@ -518,7 +518,7 @@ def _apply_prompt_updates(
                 for modality, ps in placeholders.items()
             }
 
-        return token_ids, text, placeholders
+        return token_ids, placeholders
 
 
 @MULTIMODAL_REGISTRY.register_processor(Phi3VMultiModalProcessor,

vllm/model_executor/models/qwen2_5_omni_thinker.py

@@ -63,7 +63,7 @@
                                         PromptReplacement, PromptUpdate)
 from vllm.multimodal.profiling import BaseDummyInputsBuilder
 from vllm.sequence import IntermediateTensors
-from vllm.transformers_utils.tokenizer import decode_tokens, encode_tokens
+from vllm.transformers_utils.tokenizer import encode_tokens
 from vllm.utils.tensor_schema import TensorSchema, TensorShape
 
 from .interfaces import (MultiModalEmbeddings, SupportsLoRA,
@@ -316,7 +316,7 @@ def _maybe_apply_prompt_updates(
         mm_kwargs: MultiModalKwargsItems,
         mm_prompt_updates: MultiModalPromptUpdates,
         is_update_applied: bool,
-    ) -> tuple[list[int], str, Mapping[str, list[PlaceholderFeaturesInfo]]]:
+    ) -> tuple[list[int], Mapping[str, list[PlaceholderFeaturesInfo]]]:
         """
         Qwen2.5-Omni reimplements this function to handle `use_audio_in_video`.
         """
@@ -341,28 +341,20 @@ def _maybe_apply_prompt_updates(
             self._validate_mm_placeholders(
                 mm_placeholders,
                 mm_item_counts,
-                use_audio_in_video=use_audio_in_video)
-
-            tokenizer = self.info.get_tokenizer()
-            prompt = decode_tokens(tokenizer, prompt_ids)
+                use_audio_in_video=use_audio_in_video,
+            )
         else:
-            (
-                prompt_ids,
-                prompt,
-                mm_placeholders,
-            ) = self._apply_prompt_updates(
+            prompt_ids, mm_placeholders = self._apply_prompt_updates(
                 prompt_ids,
                 mm_prompt_updates,
             )
             self._validate_mm_placeholders(
                 mm_placeholders,
                 mm_item_counts,
-                use_audio_in_video=use_audio_in_video)
-
-        tokenizer = self.info.get_tokenizer()
-        prompt = decode_tokens(tokenizer, prompt_ids)
+                use_audio_in_video=use_audio_in_video,
+            )
 
-        return prompt_ids, prompt, mm_placeholders
+        return prompt_ids, mm_placeholders
 
     def _get_prompt_updates(
         self,

vllm/model_executor/models/terratorch.py

@@ -190,7 +190,6 @@ def apply(
 
         return MultiModalInputs(
             type="multimodal",
-            prompt=prompt,
             prompt_token_ids=[1],
             mm_kwargs=mm_kwargs,
             mm_hashes=mm_hashes,

vllm/model_executor/models/transformers.py

@@ -453,7 +453,6 @@ def apply(
 
         return MultiModalInputs(
             type="multimodal",
-            prompt=prompt,
             prompt_token_ids=prompt_ids,
             mm_kwargs=mm_kwargs,
             mm_hashes=mm_hashes,