class Processor:
def __init__(
self,
vllm_config: VllmConfig,
tokenizer: AnyTokenizer,
mm_registry: MultiModalRegistry = MULTIMODAL_REGISTRY,
):
self.vllm_config = vllm_config
self.model_config = vllm_config.model_config
self.cache_config = vllm_config.cache_config
self.lora_config = vllm_config.lora_config
self.structured_outputs_config = vllm_config.structured_outputs_config
self.tokenizer = tokenizer
self.generation_config_fields = (
self.model_config.try_get_generation_config())
self.mm_registry = mm_registry
self.mm_processor_cache = processor_cache_from_config(
vllm_config, mm_registry)
self.input_preprocessor = InputPreprocessor(
self.model_config,
self.tokenizer,
mm_registry,
mm_processor_cache=self.mm_processor_cache,
)
def _validate_logprobs(
self,
params: SamplingParams,
) -> None:
max_logprobs = self.model_config.max_logprobs
if max_logprobs == -1:
max_logprobs = self.model_config.get_vocab_size()
# Validate sample logprobs.
if params.logprobs:
num_logprobs = params.logprobs
if num_logprobs == -1:
num_logprobs = self.model_config.get_vocab_size()
if num_logprobs > max_logprobs:
raise ValueError(
f"Requested sample logprobs of {num_logprobs}, "
f"which is greater than max allowed: {max_logprobs}")
# Validate prompt logprobs.
if params.prompt_logprobs:
num_prompt_logprobs = params.prompt_logprobs
if num_prompt_logprobs == -1:
num_prompt_logprobs = self.model_config.get_vocab_size()
if num_prompt_logprobs > max_logprobs:
raise ValueError(
f"Requested prompt logprobs of {num_prompt_logprobs}, "
f"which is greater than max allowed: {max_logprobs}")
def _validate_sampling_params(
self,
params: SamplingParams,
) -> None:
self._validate_structured_output(params)
self._validate_logit_bias(params)
if params.allowed_token_ids is None:
return
if not params.allowed_token_ids:
raise ValueError("allowed_token_ids is not None and empty!")
if self.tokenizer is None:
# When skip_tokenizer_init=True, we can't validate token IDs
# Skip validation and let the model handle invalid tokens
return
vocab_size = len(self.tokenizer)
if not all(0 <= tid < vocab_size for tid in params.allowed_token_ids):
raise ValueError(
"allowed_token_ids contains out-of-vocab token id!")
def _validate_logit_bias(
self,
params: SamplingParams,
) -> None:
"""Validate logit_bias token IDs are within vocabulary range."""
if not params.logit_bias:
return
vocab_size = self.model_config.get_vocab_size()
invalid_token_ids = []
for token_id in params.logit_bias:
if token_id < 0 or token_id >= vocab_size:
invalid_token_ids.append(token_id)
if invalid_token_ids:
raise ValueError(
f"token_id(s) {invalid_token_ids} in logit_bias contain "
f"out-of-vocab token ids. Vocabulary size: {vocab_size}")
def _validate_supported_sampling_params(
self,
params: SamplingParams,
) -> None:
# Best of not yet supported.
if params.best_of is not None and params.best_of > 1:
raise ValueError("vLLM V1 does not yet support best_of.")
# Logits processors not supported.
if params.logits_processors:
raise ValueError("vLLM V1 does not support per request "
"user provided logits processors.")
def _validate_params(
self,
params: Union[SamplingParams, PoolingParams],
):
"""
Validate supported SamplingParam.
Should raise ValueError if unsupported for API Server.
"""
if isinstance(params, PoolingParams):
return
self._validate_logprobs(params)
self._validate_sampling_params(params)
self._validate_supported_sampling_params(params)
def _validate_multi_modal_uuids(self, prompt: PromptType) -> None:
"""
Validate that user-provided multi_modal_uuids align with
multi_modal_data in the incoming request prompt(s).
Only checks lengths; `None` entries are allowed and will be
auto-hashed downstream.
"""
def _validate_single_prompt(single_prompt: Union[dict, str]) -> None:
if not isinstance(single_prompt, dict):
return
mm_data = single_prompt.get("multi_modal_data")
mm_uuids = single_prompt.get("multi_modal_uuids")
if not mm_data or not mm_uuids:
return
for modality, items in mm_data.items():
if modality in mm_uuids:
data_len = len(items) if isinstance(items, list) else 1
uuid_len = len(mm_uuids[modality]) if isinstance(
mm_uuids[modality], list) else 1
if uuid_len != data_len:
raise ValueError(
f"multi_modal_uuids for modality '{modality}' "
"must have same length as data: got "
f"{uuid_len} uuids vs "
f"{data_len} items.")
else:
raise ValueError(
f"multi_modal_uuids for modality '{modality}' must "
"be provided if multi_modal_data is provided.")
# Handle explicit encoder/decoder prompts or singleton prompt
if isinstance(prompt, dict) and "encoder_prompt" in prompt:
enc = prompt.get("encoder_prompt")
dec = prompt.get("decoder_prompt")
if enc is not None:
_validate_single_prompt(enc)
if dec is not None:
_validate_single_prompt(dec)
else:
_validate_single_prompt(prompt) # type: ignore[arg-type]
def _validate_lora(self, lora_request: Optional[LoRARequest]) -> None:
if lora_request is None:
return
# LoRA request passed in while LoRA is not enabled
if not self.lora_config:
raise ValueError(f"Got lora_request {lora_request} but LoRA is "
"not enabled!")
if self.tokenizer is not None:
logger.warning_once(
"vLLM has deprecated support for supporting different "
"tokenizers for different LoRAs. By default, vLLM uses base "
"model's tokenizer. If you are using a LoRA "
"with its own tokenizer, consider specifying `--tokenizer "
"[lora_path]` to use the LoRA tokenizer.")
def _validate_structured_output(self, params: SamplingParams) -> None:
if not params.structured_outputs or not self.structured_outputs_config:
return
if self.model_config.skip_tokenizer_init and params.structured_outputs:
raise ValueError(
"Structured outputs requires a tokenizer so it can't be used with 'skip_tokenizer_init'" # noqa: E501
)
backend = self.structured_outputs_config.backend
if _backend := params.structured_outputs._backend:
# Request-level backend selection is not supported.
# The values may differ if `params` is reused and was set
# to a specific backend based on `auto` behavior in a previous
# request. We remember that it was set as a result of `auto`
# using the `_backend_was_auto` field set in the params.
if (backend != _backend
and not (backend == "auto"
and params.structured_outputs._backend_was_auto)):
raise ValueError(
"Request-level structured output backend selection is not "
f"supported. The request specified '{_backend}', but vLLM "
f"was initialised with '{backend}'. This error can be "
"resolved by removing '_backend' from the request.")
else:
params.structured_outputs._backend = backend
# Request content validation
if (isinstance(params.structured_outputs.choice, list)
and not params.structured_outputs.choice):
# It is invalid for choice to be an empty list
raise ValueError(
f"Choice '{params.structured_outputs.choice}' cannot be an empty list" # noqa: E501
)
if backend.startswith("xgrammar"):
# xgrammar with no fallback
validate_xgrammar_grammar(params)
elif backend.startswith("guidance"):
# TODO: ideally we would have the LLTokenizer here as Lark syntax
# allows <|special_token|> and similar, see
# https://github.com/guidance-ai/llguidance/blob/main/docs/syntax.md#special-tokens
# Without tokenizer these are disallowed in grammars.
validate_guidance_grammar(params, tokenizer=None)
elif backend == "outlines":
# outlines backend
validate_structured_output_request_outlines(params)
elif backend == "lm-format-enforcer":
# lm format enforcer backend
validate_structured_output_request_lm_format_enforcer(params)
else:
# NOTE: backend must be "auto" here, because we have
# checked supported_backends above.
# In this mode, we set opinionated defaults based on what we think
# will satisfy the most use cases without having to worry about
# this setting. We include fallback behavior here, but not with any
# other setting where a specific backend was specified.
try:
validate_xgrammar_grammar(params)
params.structured_outputs._backend = "xgrammar"
except ValueError:
# The request either failed validation
# or includes some jsonschema feature(s) that
# are not supported in xgrammar. Fall back to guidance.
validate_guidance_grammar(params, tokenizer=None)
params.structured_outputs._backend = "guidance"
# Remember that this backend was set automatically
params.structured_outputs._backend_was_auto = True
def _maybe_build_mm_uuids(
self,
request_id: str,
prompt: PromptType,
) -> Optional[MultiModalUUIDDict]:
"""Build per-item multimodal hash overrides when enabled. In this case,
multimodal data items are identified by their request id, modality and
index rather than their content.
Returns a dictionary of modality -> list[str] of overrides, or None if
disabled or no multimodal data is present.
"""
def _extract_mm_data(p: PromptType):
if isinstance(p, dict) and "encoder_prompt" in p:
enc = p.get("encoder_prompt")
if isinstance(enc, dict):
return enc.get("multi_modal_data")
return None
if isinstance(p, dict):
return p.get("multi_modal_data")
return None
mm_data = _extract_mm_data(prompt)
if not mm_data:
return None
mm_uuids: MultiModalUUIDDict = {}
for modality, data in mm_data.items():
n = len(data) if isinstance(data, list) else 1
mm_uuids[modality] = [
f"{request_id}-{modality}-{i}" for i in range(n)
]
return mm_uuids
def process_inputs(
self,
request_id: str,
prompt: PromptType,
params: Union[SamplingParams, PoolingParams],
arrival_time: Optional[float] = None,
lora_request: Optional[LoRARequest] = None,
tokenization_kwargs: Optional[dict[str, Any]] = None,
trace_headers: Optional[Mapping[str, str]] = None,
priority: int = 0,
data_parallel_rank: Optional[int] = None,
) -> tuple[Optional[str], EngineCoreRequest]:
# TODO(woosuk): Support pooling models.
self._validate_lora(lora_request)
self._validate_params(params)
data_parallel_size = self.vllm_config.parallel_config.data_parallel_size
if data_parallel_rank is not None and not (0 <= data_parallel_rank <
data_parallel_size):
raise ValueError(f"data_parallel_rank {data_parallel_rank} "
f"is out of range [0, {data_parallel_size}).")
if arrival_time is None:
arrival_time = time.time()
# Optionally generate multimodal hash overrides to avoid hashing
# multimodal data items by their content as their identifiers.
# NOTE: when users explicitly turn off BOTH prefix caching and input
# processing caching, no multimodal features or embeddings will be
# reused across requests, therefore identifying multimodal data items
# by their content is no longer necessary, and we create uuids with
# request id-modality-index as multimodal hash overrides.
if (self.model_config.multimodal_config and
self.model_config.multimodal_config.mm_processor_cache_gb == 0
and not self.cache_config.enable_prefix_caching):
mm_uuids = self._maybe_build_mm_uuids(request_id, prompt)
else:
# Otherwise, use user-provided uuids as multimodal hash overrides
# if provided.
self._validate_multi_modal_uuids(prompt)
if isinstance(prompt, dict):
mm_uuids = prompt.get("multi_modal_uuids")
else:
mm_uuids = None
# Process inputs, which includes:
# 1. Tokenize text prompt, with LoRA request if one exists.
# 2. For multimodal models with a merged preprocessor, preprocess
# multimodal data and expand prompt token ids accordingly.
processed_inputs: ProcessorInputs = self.input_preprocessor.preprocess(
prompt,
tokenization_kwargs=tokenization_kwargs,
mm_uuids=mm_uuids,
)
from vllm.platforms import current_platform
current_platform.validate_request(
prompt=prompt,
params=params,
processed_inputs=processed_inputs,
)
eos_token_id = self.input_preprocessor.get_eos_token_id()
encoder_inputs, decoder_inputs = split_enc_dec_inputs(processed_inputs)
self._validate_model_inputs(encoder_inputs, decoder_inputs)
# Mypy does not always properly infer the types of some elements of
# discriminated unions of TypedDicts, because of how it handles
# inheritance of TypedDict. If we explicitly extract the items we want
# we can avoid type errors from using `dict.get` later in the method.
prompt_str: Optional[str] = None if decoder_inputs[
"type"] == "embeds" else decoder_inputs.get("prompt")
prompt_token_ids = decoder_inputs[
"prompt_token_ids"] if decoder_inputs["type"] != "embeds" else None
prompt_embeds = decoder_inputs["prompt_embeds"] if decoder_inputs[
"type"] == "embeds" else None
sampling_params = None
pooling_params = None
if isinstance(params, SamplingParams):
# TODO: can we avoid cloning here in multiproc case?
sampling_params = params.clone()
# If unset max tokens, then generate up to the max_model_len.
if sampling_params.max_tokens is None:
seq_len = length_from_prompt_token_ids_or_embeds(
prompt_token_ids, prompt_embeds)
sampling_params.max_tokens = \
self.model_config.max_model_len - seq_len
sampling_params.update_from_generation_config(
self.generation_config_fields, eos_token_id)
if self.tokenizer is not None:
sampling_params.update_from_tokenizer(self.tokenizer)
else:
pooling_params = params.clone()
# Multimodal related.
mm_features: Optional[list[MultiModalFeatureSpec]] = None
if decoder_inputs["type"] == "multimodal":
decoder_mm_inputs = decoder_inputs["mm_kwargs"]
decoder_mm_positions = decoder_inputs["mm_placeholders"]
decoder_mm_hashes = decoder_inputs["mm_hashes"]
# Merge and flatten multimodal placeholders, hashes and inputs
# from dictionaries to lists, and sort them by each item's position
# in the input sequence.
sorted_mm_idxs = argsort_mm_positions(decoder_mm_positions)
mm_features = []
for modality, idx in sorted_mm_idxs:
mm_features.append(
MultiModalFeatureSpec(
data=decoder_mm_inputs[modality][idx],
modality=modality,
identifier=decoder_mm_hashes[modality][idx],
mm_position=decoder_mm_positions[modality][idx]))
return prompt_str, EngineCoreRequest(
request_id=request_id,
prompt_token_ids=prompt_token_ids,
prompt_embeds=prompt_embeds,
mm_features=mm_features,
sampling_params=sampling_params,
pooling_params=pooling_params,
eos_token_id=eos_token_id,
arrival_time=arrival_time,
lora_request=lora_request,
cache_salt=decoder_inputs.get("cache_salt"),
priority=priority,
data_parallel_rank=data_parallel_rank,
trace_headers=trace_headers,
)
def _validate_model_inputs(self, encoder_inputs: Optional[SingletonInputs],
decoder_inputs: SingletonInputs):
if encoder_inputs is not None:
self._validate_model_input(encoder_inputs, prompt_type="encoder")
self._validate_model_input(decoder_inputs, prompt_type="decoder")
def _validate_model_input(
self,
prompt_inputs: SingletonInputs,
*,
prompt_type: Literal["encoder", "decoder"],
):
model_config = self.model_config
prompt_ids = None if prompt_inputs[
"type"] == "embeds" else prompt_inputs["prompt_token_ids"]
prompt_embeds = prompt_inputs["prompt_embeds"] if prompt_inputs[
"type"] == "embeds" else None
prompt_len = length_from_prompt_token_ids_or_embeds(
prompt_ids, prompt_embeds)
if not prompt_ids:
if prompt_type == "encoder" and model_config.is_multimodal_model:
pass # Mllama may have empty encoder inputs for text-only data
elif prompt_inputs["type"] == "embeds":
pass # Prompt embeds should not have prompt_ids.
else:
raise ValueError(f"The {prompt_type} prompt cannot be empty")
if self.model_config.skip_tokenizer_init:
tokenizer = None
else:
tokenizer = self.tokenizer
max_input_id = max(prompt_ids or [], default=0)
# NOTE: tokenizer.max_token_id is the tokenizer’s vocab size while
# self.model_config.get_vocab_size() is the model’s vocab size.
# For Qwen3 models, the language model has extra tokens that do
# not exist in the tokenizer, and vice versa for multimodal
# placeholder tokens in some multimodal models.
# See https://github.com/QwenLM/Qwen3/issues/29#issuecomment-1933720399 # noqa: E501
# and https://github.com/vllm-project/vllm/pull/22471#discussion_r2312251421 # noqa: E501
# Here we take the max of the two to determine if a token id is
# truly out-of-vocabulary.
if max_input_id > max(tokenizer.max_token_id,
self.model_config.get_vocab_size() - 1):
raise ValueError(
f"Token id {max_input_id} is out of vocabulary")
max_prompt_len = self.model_config.max_model_len
if prompt_len > max_prompt_len:
if prompt_type == "encoder" and model_config.is_multimodal_model:
mm_registry = self.input_preprocessor.mm_registry
mm_processor = mm_registry.create_processor(
model_config,
tokenizer=tokenizer,
)
assert isinstance(mm_processor, EncDecMultiModalProcessor)
if mm_processor.pad_dummy_encoder_prompt:
return # Skip encoder length check for Whisper
if model_config.is_multimodal_model:
suggestion = (
"Make sure that `max_model_len` is no smaller than the "
"number of text tokens plus multimodal tokens. For image "
"inputs, the number of image tokens depends on the number "
"of images, and possibly their aspect ratios as well.")
else:
suggestion = (
"Make sure that `max_model_len` is no smaller than the "
"number of text tokens.")
raise ValueError(
f"The {prompt_type} prompt (length {prompt_len}) is "
f"longer than the maximum model length of {max_prompt_len}. "
f"{suggestion}")
# TODO: Find out how many placeholder tokens are there so we can
# check that chunked prefill does not truncate them
# max_batch_len = self.scheduler_config.max_num_batched_tokens
def clear_cache(self) -> None:
self.input_preprocessor.clear_cache()