diff --git a/vllm/config/cache.py b/vllm/config/cache.py index 4fa7e1cfcb..352ccec320 100644 --- a/vllm/config/cache.py +++ b/vllm/config/cache.py @@ -1,10 +1,11 @@ # SPDX-License-Identifier: Apache-2.0 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project +from collections.abc import Callable from dataclasses import field -from typing import ClassVar, Literal +from typing import Any, ClassVar, Literal -from pydantic import Field, SkipValidation, field_validator, model_validator +from pydantic import Field, field_validator, model_validator from vllm.config.utils import config from vllm.logger import init_logger @@ -44,14 +45,14 @@ class CacheConfig: DEFAULT_BLOCK_SIZE: ClassVar[int] = 16 - block_size: SkipValidation[int] = None # type: ignore[assignment] + block_size: int = Field(default=None, gt=0) # type: ignore[assignment] """Size of a contiguous cache block in number of tokens. Accepts None (meaning "use default"). After construction, always int.""" user_specified_block_size: bool = field(default=False, init=False) """Whether block_size was explicitly provided. Derived automatically.""" user_specified_mamba_block_size: bool = field(default=False, init=False) """Whether mamba_block_size was explicitly provided. Derived automatically.""" - hash_block_size: SkipValidation[int] | None = None # type: ignore + hash_block_size: int | None = Field(default=None, gt=0) """Block size (in tokens) used for computing Request's block_hashes. This can be set to a finer granularity than the physical KV cache block @@ -220,19 +221,26 @@ class CacheConfig: _block_size_resolved: bool = field(default=False, init=False) """Guard against pydantic re-running _apply_block_size_default.""" + @field_validator("block_size", mode="wrap") + @classmethod + def _skip_none_validation(cls, value: Any, handler: Callable) -> Any: + if value is None: + return value + return handler(value) + @model_validator(mode="after") def _apply_block_size_default(self) -> "CacheConfig": # Pydantic re-runs validators when CacheConfig is nested inside # another pydantic model (e.g. VllmConfig). Guard against that. if self._block_size_resolved: return self - object.__setattr__(self, "_block_size_resolved", True) + self._block_size_resolved = True if self.block_size is None: - object.__setattr__(self, "block_size", self.DEFAULT_BLOCK_SIZE) + self.block_size = self.DEFAULT_BLOCK_SIZE else: - object.__setattr__(self, "user_specified_block_size", True) + self.user_specified_block_size = True if self.mamba_block_size is not None: - object.__setattr__(self, "user_specified_mamba_block_size", True) + self.user_specified_mamba_block_size = True return self @field_validator("calculate_kv_scales", mode="after") diff --git a/vllm/config/model.py b/vllm/config/model.py index 811798859b..491bd1a415 100644 --- a/vllm/config/model.py +++ b/vllm/config/model.py @@ -791,7 +791,7 @@ class ModelConfig: f"{type(self.tokenizer).__name__}: {self.tokenizer!r}. " "Please provide a valid tokenizer path or HuggingFace model ID." ) - if not isinstance(self.max_model_len, int): + if not isinstance(self.max_model_len, int) or self.max_model_len < 1: raise ValueError( f"max_model_len must be a positive integer, " f"got {type(self.max_model_len).__name__}: {self.max_model_len!r}. "