[Bugfix] Fix SHM cache initialization (#26427)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2025-10-09 02:48:04 -07:00
committed by GitHub
parent dc7976dd9f
commit 4bdf7ac593
30 changed files with 357 additions and 417 deletions
+4 -4
View File
@@ -20,13 +20,13 @@ class TextPrompt(TypedDict):
prompt: str
"""The input text to be tokenized before passing to the model."""
multi_modal_data: NotRequired["MultiModalDataDict"]
multi_modal_data: NotRequired[Optional["MultiModalDataDict"]]
"""
Optional multi-modal data to pass to the model,
if the model supports it.
"""
mm_processor_kwargs: NotRequired[dict[str, Any]]
mm_processor_kwargs: NotRequired[Optional[dict[str, Any]]]
"""
Optional multi-modal processor kwargs to be forwarded to the
multimodal input mapper & processor. Note that if multiple modalities
@@ -61,13 +61,13 @@ class TokensPrompt(TypedDict):
token_type_ids: NotRequired[list[int]]
"""A list of token type IDs to pass to the cross encoder model."""
multi_modal_data: NotRequired["MultiModalDataDict"]
multi_modal_data: NotRequired[Optional["MultiModalDataDict"]]
"""
Optional multi-modal data to pass to the model,
if the model supports it.
"""
mm_processor_kwargs: NotRequired[dict[str, Any]]
mm_processor_kwargs: NotRequired[Optional[dict[str, Any]]]
"""
Optional multi-modal processor kwargs to be forwarded to the
multimodal input mapper & processor. Note that if multiple modalities
+7 -10
View File
@@ -17,7 +17,7 @@ from vllm.multimodal.inputs import (
MultiModalUUIDDict,
)
from vllm.multimodal.processing import BaseMultiModalProcessor
from vllm.transformers_utils.tokenizer import AnyTokenizer, init_tokenizer_from_configs
from vllm.transformers_utils.tokenizer import AnyTokenizer
from vllm.utils.jsontree import json_iter_leaves
from .data import (
@@ -45,20 +45,17 @@ class InputPreprocessor:
def __init__(
self,
model_config: ModelConfig,
tokenizer: Optional[AnyTokenizer],
mm_registry: MultiModalRegistry = MULTIMODAL_REGISTRY,
mm_processor_cache: Optional[BaseMultiModalProcessorCache] = None,
) -> None:
super().__init__()
self.model_config = model_config
self.tokenizer = tokenizer
self.mm_registry = mm_registry
self.mm_processor_cache = mm_processor_cache
if model_config.skip_tokenizer_init:
self.tokenizer = None
else:
self.tokenizer = init_tokenizer_from_configs(model_config)
def get_tokenizer(self) -> AnyTokenizer:
if self.tokenizer is None:
raise ValueError(
@@ -351,8 +348,8 @@ class InputPreprocessor:
if self.model_config.is_multimodal_model:
inputs = self._process_multimodal(
prompt_token_ids,
parsed_content.get("multi_modal_data", {}),
parsed_content.get("mm_processor_kwargs"),
parsed_content.get("multi_modal_data") or {},
parsed_content.get("mm_processor_kwargs") or {},
tokenization_kwargs=tokenization_kwargs,
mm_uuids=mm_uuids,
)
@@ -380,8 +377,8 @@ class InputPreprocessor:
if self.model_config.is_multimodal_model:
inputs = self._process_multimodal(
prompt_text,
parsed_content.get("multi_modal_data", {}),
parsed_content.get("mm_processor_kwargs"),
parsed_content.get("multi_modal_data") or {},
parsed_content.get("mm_processor_kwargs") or {},
tokenization_kwargs=tokenization_kwargs,
mm_uuids=mm_uuids,
)