vllm-npu 0.4.2__py3-none-any.whl

vllm/lora/punica.py

@@ -0,0 +1,213 @@
+# Based on code from https://github.com/punica-ai/punica
+
+from typing import Optional
+
+import torch
+
+
+def _raise_import_error(e):
+    if torch.cuda.get_device_capability() < (8, 0):
+        raise ImportError(
+            "punica LoRA kernels require compute capability >= 8.0") from e
+    else:
+        raise ImportError(
+            "punica LoRA kernels could not be imported. If you built vLLM "
+            "from source, make sure VLLM_INSTALL_PUNICA_KERNELS=1 env var "
+            "was set.") from e
+
+
+def bgmv(
+    y: torch.Tensor,
+    x: torch.Tensor,
+    w_t_all: torch.Tensor,
+    indicies: torch.LongTensor,
+    layer_idx: int,
+    scale: float,
+):
+    """
+    Semantics:
+      y[i] += (
+          x[i].unsqueeze(0)
+          @ w_t_all[indices[i], layer_idx, :, :].transpose(-1, -2)
+          * scale
+        ).squeeze(0)
+
+    Args:
+      y: Shape: `[B, H2]`. Output vectors. Will be changed in-place.
+      x: Shape: `[B, H1]`. Input vectors.
+      w_t_all: Shape: `[None, L, H2, H1]`. All of the transposed weight
+        matrices.
+      indicies: Shape: `[B]`. Indices of the weight matrices.
+      layer_idx: Layer index of the weight matrices.
+      scale: Scaling factor.
+    """
+    try:
+        import vllm._punica_C as punica_kernels
+    except ImportError as e:
+        _raise_import_error(e)
+
+    punica_kernels.dispatch_bgmv(y, x, w_t_all, indicies, layer_idx, scale)
+
+
+def dispatch_bgmv_low_level(y: torch.Tensor, x: torch.Tensor,
+                            w_t_all: torch.Tensor, indicies: torch.LongTensor,
+                            layer_idx: int, scale: float, y_offset: int,
+                            y_slice_size: int):
+    """
+    Same as `bgmv` but you can operate on slices of y.
+    Pass whole y, define y_offset and y_slice_size.
+
+    Semantics:
+      y[i] += (
+          x[i].unsqueeze(0)
+          @ w_t_all[indices[i], layer_idx, :, :].transpose(-1, -2)
+          * scale
+        ).squeeze(0)
+
+    Args:
+      y: Shape: `[B, H2]`. Output vectors. Will be changed in-place.
+      x: Shape: `[B, H1]`. Input vectors.
+      w_t_all: Shape: `[None, L, y_slice_size, H1]`. Column partition of
+        all of the transposed LoRA matrices.
+      indicies: Shape: `[B]`. Indices of the LoRA weights.
+      layer_idx: Layer index of LoRA weights.
+      scale: Scaling factor.
+      y_offset: Offset to apply to the starting column of y.
+      y_slice_size: Size of the y column slice.
+    """
+    try:
+        import vllm._punica_C as punica_kernels
+    except ImportError as e:
+        _raise_import_error(e)
+    punica_kernels.dispatch_bgmv_low_level(
+        y,
+        x,
+        w_t_all,
+        indicies,
+        layer_idx,
+        scale,
+        x.size(1),
+        y_slice_size,
+        y_offset,
+    )
+
+
+def add_lora(y: torch.Tensor,
+             x: torch.Tensor,
+             wa_t_all: torch.Tensor,
+             wb_t_all: torch.Tensor,
+             indicies: torch.LongTensor,
+             layer_idx: int,
+             scale: float,
+             *,
+             buffer: Optional[torch.Tensor] = None):
+    """
+    Semantics:
+      y[i] += (
+          x[i].unsqueeze(0)
+          @ wa_t_all[indices[i], layer_idx, :, :].transpose(-1, -2)
+          @ wb_t_all[indices[i], layer_idx, :, :].transpose(-1, -2)
+          * scale
+        ).squeeze(0)
+
+    Args:
+      y: Shape: `[B, H2]`. Output vectors. Will be changed in-place.
+      x: Shape: `[B, H1]`. Input vectors.
+      wa_t_all: Shape: `[None, L, R, H1]`. All of the transposed
+        LoRA A matrices.
+      wb_t_all: Shape: `[None, L, H2, R]`. All of the transposed
+        LoRA B matrices.
+      indicies: Shape: `[B]`. Indices of the LoRA weights.
+      layer_idx: Layer index of LoRA weights.
+      scale: Scaling factor.
+      buffer: Optional. Shape: `[B, R]`. Temporary buffer.
+    """
+    try:
+        import vllm._punica_C as punica_kernels
+    except ImportError as e:
+        _raise_import_error(e)
+
+    r = wb_t_all.size(-1)
+    if buffer is None:
+        # We set the buffer to be float32 by default to avoid
+        # numerical inaccuracies that would otherwise happen
+        # due to downcasting.
+        buffer = torch.zeros((x.size(0), r),
+                             dtype=torch.float32,
+                             device=x.device)
+    punica_kernels.dispatch_bgmv(buffer, x, wa_t_all, indicies, layer_idx, 1.0)
+    punica_kernels.dispatch_bgmv(y, buffer, wb_t_all, indicies, layer_idx,
+                                 scale)
+
+
+def add_lora_slice(y: torch.Tensor,
+                   x: torch.Tensor,
+                   wa_t_all: torch.Tensor,
+                   wb_t_all: torch.Tensor,
+                   indicies: torch.LongTensor,
+                   layer_idx: int,
+                   scale: float,
+                   y_offset: int,
+                   y_slice_size: int,
+                   *,
+                   buffer: Optional[torch.Tensor] = None):
+    """
+    Same as `add_lora` but you can operate on slices of y.
+    Pass whole y, define y_offset and y_slice_size.
+
+    Semantics:
+      y[i] += (
+          x[i].unsqueeze(0)
+          @ wa_t_all[indices[i], layer_idx, :, :].transpose(-1, -2)
+          @ wb_t_all[indices[i], layer_idx, :, :].transpose(-1, -2)
+          * scale
+        ).squeeze(0)
+
+    Args:
+      y: Shape: `[B, H2]`. Output vectors. Will be changed in-place.
+      x: Shape: `[B, H1]`. Input vectors.
+      wa_t_all: Shape: `[None, L, R, H1]`. All of the transposed
+        LoRA A matrices.
+      wb_t_all: Shape: `[None, L, H2, R]`. All of the transposed
+        LoRA B matrices.
+      indicies: Shape: `[B]`. Indices of the LoRA weights.
+      layer_idx: Layer index of LoRA weights.
+      scale: Scaling factor.
+      y_offset: Offset to apply to the starting column of y.
+      y_slice_size: Size of the y column slice.
+    """
+    try:
+        import vllm._punica_C as punica_kernels
+    except ImportError as e:
+        _raise_import_error(e)
+
+    r = wb_t_all.size(-1)
+    if buffer is None:
+        # We set the buffer to be float32 by default to avoid
+        # numerical inaccuracies that would otherwise happen
+        # due to downcasting.
+        buffer = torch.zeros((x.size(0), r),
+                             dtype=torch.float32,
+                             device=x.device)
+    punica_kernels.dispatch_bgmv_low_level(
+        buffer,
+        x,
+        wa_t_all,
+        indicies,
+        layer_idx,
+        1.0,
+        x.size(1),
+        buffer.size(1),
+        0,
+    )
+    punica_kernels.dispatch_bgmv_low_level(
+        y,
+        buffer,
+        wb_t_all,
+        indicies,
+        layer_idx,
+        scale,
+        buffer.size(1),
+        y_slice_size,
+        y_offset,
+    )

vllm/lora/request.py

@@ -0,0 +1,32 @@
+from dataclasses import dataclass
+
+
+@dataclass
+class LoRARequest:
+    """
+    Request for a LoRA adapter.
+
+    Note that this class should be be used internally. For online
+    serving, it is recommended to not allow users to use this class but
+    instead provide another layer of abstraction to prevent users from
+    accessing unauthorized LoRA adapters.
+
+    lora_int_id must be globally unique for a given adapter.
+    This is currently not enforced in vLLM.
+    """
+
+    lora_name: str
+    lora_int_id: int
+    lora_local_path: str
+
+    def __post_init__(self):
+        if self.lora_int_id < 1:
+            raise ValueError(
+                f"lora_int_id must be > 0, got {self.lora_int_id}")
+
+    def __eq__(self, value: object) -> bool:
+        return isinstance(
+            value, LoRARequest) and self.lora_int_id == value.lora_int_id
+
+    def __hash__(self) -> int:
+        return self.lora_int_id

vllm/lora/utils.py

@@ -0,0 +1,98 @@
+from typing import List, Optional, Set, Tuple, Type
+
+from torch import nn
+from transformers import PretrainedConfig
+
+from vllm.config import LoRAConfig
+from vllm.logger import init_logger
+from vllm.lora.fully_sharded_layers import (
+    ColumnParallelLinearWithShardedLoRA,
+    MergedColumnParallelLinearWithShardedLoRA,
+    MergedQKVParallelLinearWithShardedLora, RowParallelLinearWithShardedLoRA)
+# being imported for _all_lora_classes below
+# yapf conflicts with isort for this block
+# yapf: disable
+from vllm.lora.layers import (BaseLayerWithLoRA, ColumnParallelLinearWithLoRA,
+                              LogitsProcessorWithLoRA,
+                              MergedColumnParallelLinearWithLoRA,
+                              MergedQKVParallelLinearWithLora,
+                              QKVParallelLinearWithLora,
+                              RowParallelLinearWithLoRA,
+                              VocabParallelEmbeddingWithLoRA)
+# yapf: enable
+from vllm.model_executor.layers.logits_processor import LogitsProcessor
+from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
+
+logger = init_logger(__name__)
+
+_all_lora_classes: Set[Type[BaseLayerWithLoRA]] = {
+    VocabParallelEmbeddingWithLoRA, ColumnParallelLinearWithLoRA,
+    MergedColumnParallelLinearWithLoRA, QKVParallelLinearWithLora,
+    MergedQKVParallelLinearWithLora, RowParallelLinearWithLoRA,
+    LogitsProcessorWithLoRA, ColumnParallelLinearWithShardedLoRA,
+    MergedColumnParallelLinearWithShardedLoRA,
+    MergedQKVParallelLinearWithShardedLora, RowParallelLinearWithShardedLoRA
+}
+
+
+def from_layer(layer: nn.Module,
+               max_loras: int,
+               lora_config: LoRAConfig,
+               packed_modules_list: List,
+               model_config: Optional[PretrainedConfig] = None) -> nn.Module:
+    for lora_cls in _all_lora_classes:
+        # specifying kwargs so they can be easily accessed in decorator
+        if lora_cls.can_replace_layer(source_layer=layer,
+                                      lora_config=lora_config,
+                                      packed_modules_list=packed_modules_list,
+                                      model_config=model_config):
+            ret = lora_cls(layer)
+            ret.create_lora_weights(max_loras, lora_config, model_config)
+            return ret
+    return layer
+
+
+def from_layer_logits_processor(
+    layer: LogitsProcessor,
+    lm_head: ParallelLMHead,
+    max_loras: int,
+    lora_config: LoRAConfig,
+    model_config: Optional[PretrainedConfig] = None,
+) -> LogitsProcessorWithLoRA:
+    ret = LogitsProcessorWithLoRA(layer, lm_head.embedding_dim,
+                                  lm_head.weight.dtype, lm_head.weight.device)
+    ret.create_lora_weights(max_loras, lora_config, model_config)
+    return ret
+
+
+def replace_submodule(model: nn.Module, module_name: str,
+                      new_module: nn.Module) -> nn.Module:
+    """Replace a submodule in a model with a new module."""
+    parent = model.get_submodule(".".join(module_name.split(".")[:-1]))
+    target_name = module_name.split(".")[-1]
+    setattr(parent, target_name, new_module)
+    return new_module
+
+
+def parse_fine_tuned_lora_name(name: str) -> Tuple[str, bool]:
+    """Parse the name of lora weights.
+
+    args:
+        name: the name of the fine-tuned LoRA, e.g.
+            base_model.model.dense1.weight
+    return:
+        Tuple(module_name, is_lora_a):
+            module_name: the name of the module, e.g. model.dense1,
+            is_lora_a whether the tensor is lora_a or lora_b.
+    """
+    parts = name.split(".")
+    assert parts[0] == "base_model"
+    assert parts[1] == "model"
+    if parts[-1] == "weight":
+        assert parts[-2] == "lora_A" or parts[-2] == "lora_B"
+        return ".".join(parts[2:-2]), parts[-2] == "lora_A"
+
+    if parts[-1] == "lora_embedding_A" or parts[-1] == "lora_embedding_B":
+        return ".".join(parts[2:-1]), parts[-1] == "lora_embedding_A"
+
+    raise ValueError(f"{name} is unsupported format")

vllm/lora/worker_manager.py

@@ -0,0 +1,251 @@
+from abc import ABC, abstractmethod, abstractproperty
+from typing import Any, Dict, List, Set, Type
+
+import torch
+
+from vllm.config import LoRAConfig
+from vllm.logger import init_logger
+from vllm.lora.layers import LoRAMapping
+from vllm.lora.models import (LoRAModel, LoRAModelManager,
+                              LRUCacheLoRAModelManager, create_lora_manager)
+from vllm.lora.request import LoRARequest
+
+logger = init_logger(__name__)
+
+
+class AbstractWorkerLoRAManager(ABC):
+    """Abstract class for managing LoRA models on the worker side."""
+
+    def __init__(self, max_num_seqs: int, max_num_batched_tokens: int,
+                 vocab_size: int, lora_config: LoRAConfig,
+                 device: torch.device):
+        self.max_num_seqs = max_num_seqs
+        self.max_num_batched_tokens = max_num_batched_tokens
+        self.vocab_size = vocab_size
+        self.device = device
+        self.lora_config = lora_config
+
+    @abstractproperty
+    def is_enabled(self) -> bool:
+        ...
+
+    @abstractmethod
+    def create_lora_manager(
+        self,
+        model: torch.nn.Module,
+    ) -> Any:
+        ...
+
+    @abstractmethod
+    def set_active_loras(self, lora_requests: Set[LoRARequest],
+                         lora_mapping: LoRAMapping) -> None:
+        ...
+
+    @abstractmethod
+    def add_lora(self, lora_request: LoRARequest) -> bool:
+        ...
+
+    @abstractmethod
+    def add_dummy_lora(self, lora_request: LoRARequest, rank: int) -> bool:
+        ...
+
+    @abstractmethod
+    def remove_lora(self, lora_id: int) -> bool:
+        ...
+
+    @abstractmethod
+    def remove_all_loras(self):
+        ...
+
+    @abstractmethod
+    def list_loras(self) -> Set[int]:
+        ...
+
+
+class WorkerLoRAManager(AbstractWorkerLoRAManager):
+    """WorkerLoRAManager that manages LoRA models on the worker side.
+
+    Every request, the requested LoRAs will be loaded (unless they are already
+    loaded), and every other LoRA will be unloaded."""
+
+    _lora_manager_cls: Type[LoRAModelManager] = LoRAModelManager
+
+    def __init__(
+        self,
+        max_num_seqs: int,
+        max_num_batched_tokens: int,
+        vocab_size: int,
+        lora_config: LoRAConfig,
+        device: torch.device,
+        embedding_modules: Dict[str, str],
+        embedding_padding_modules: List[str],
+        lora_model_cls: Type[LoRAModel] = LoRAModel,
+    ):
+        self._lora_model_cls = lora_model_cls
+        self.embedding_modules = embedding_modules
+        self.embedding_padding_modules = embedding_padding_modules
+        # Lazily initialized by create_lora_manager.
+        self._lora_manager: LoRAModelManager
+        super().__init__(max_num_seqs, max_num_batched_tokens, vocab_size,
+                         lora_config, device)
+
+    @property
+    def is_enabled(self) -> bool:
+        return True
+
+    def create_lora_manager(
+        self,
+        model: torch.nn.Module,
+    ) -> Any:
+        lora_manager = create_lora_manager(
+            model,
+            max_num_seqs=self.max_num_seqs,
+            max_num_batched_tokens=self.max_num_batched_tokens,
+            vocab_size=self.vocab_size,
+            lora_config=self.lora_config,
+            lora_manager_cls=self._lora_manager_cls,
+        )
+        self._lora_manager = lora_manager
+        return lora_manager.model
+
+    def set_active_loras(self, lora_requests: Set[LoRARequest],
+                         lora_mapping: LoRAMapping) -> None:
+        self._apply_loras(lora_requests)
+        self._lora_manager.set_lora_mapping(lora_mapping)
+
+    def _apply_loras(self, lora_requests: Set[LoRARequest]) -> None:
+        loras_that_exist = self.list_loras()
+        loras_map = {
+            lora_request.lora_int_id: lora_request
+            for lora_request in lora_requests if lora_request
+        }
+        if len(loras_map) > self._lora_manager.lora_slots:
+            raise RuntimeError(
+                f"Number of requested LoRAs ({len(loras_map)}) is greater "
+                "than the number of GPU LoRA slots "
+                f"({self._lora_manager.lora_slots}).")
+
+        new_loras = set(loras_map)
+        loras_to_add = new_loras - loras_that_exist
+        loras_to_remove = loras_that_exist - new_loras
+
+        for lora_id in loras_to_remove:
+            self.remove_lora(lora_id)
+
+        for lora_id in loras_to_add:
+            self.add_lora(loras_map[lora_id])
+
+    def _load_lora(self, lora_request: LoRARequest) -> LoRAModel:
+        try:
+            model = self._lora_manager.model
+            supported_lora_modules = model.supported_lora_modules
+            packed_modules_mapping = model.packed_modules_mapping
+            expected_lora_modules = []
+            for module in supported_lora_modules:
+                if module in packed_modules_mapping:
+                    expected_lora_modules.extend(
+                        packed_modules_mapping[module])
+                else:
+                    expected_lora_modules.append(module)
+            lora = self._lora_model_cls.from_local_checkpoint(
+                lora_request.lora_local_path,
+                expected_lora_modules,
+                lora_model_id=lora_request.lora_int_id,
+                device="cpu",
+                dtype=self.lora_config.lora_dtype,
+                target_embedding_padding=self.vocab_size +
+                self.lora_config.lora_extra_vocab_size,
+                embedding_modules=self.embedding_modules,
+                embedding_padding_modules=self.embedding_padding_modules,
+            )
+        except Exception as e:
+            raise RuntimeError(
+                f"Loading lora {lora_request.lora_local_path} failed") from e
+        if lora.rank > self.lora_config.max_lora_rank:
+            raise ValueError(
+                f"LoRA rank {lora.rank} is greater than max_lora_rank "
+                f"{self.lora_config.max_lora_rank}.")
+        if lora.extra_vocab_size > self.lora_config.lora_extra_vocab_size:
+            raise ValueError(f"LoRA added vocab size {lora.extra_vocab_size} "
+                             f"is greater than lora_extra_vocab_size "
+                             f"{self.lora_config.lora_extra_vocab_size}.")
+        return lora
+
+    def add_dummy_lora(self, lora_request: LoRARequest, rank: int) -> bool:
+        if lora_request.lora_int_id in self.list_loras():
+            return False
+        return self._lora_manager.add_lora(
+            self._lora_manager.create_dummy_lora(lora_request.lora_int_id,
+                                                 rank, self.embedding_modules))
+
+    def add_lora(self, lora_request: LoRARequest) -> bool:
+        if lora_request.lora_int_id in self.list_loras():
+            return False
+        lora = self._load_lora(lora_request)
+        loaded = self._lora_manager.add_lora(lora)
+        self._lora_manager.activate_lora(lora.id)
+        return loaded
+
+    def remove_lora(self, lora_id: int) -> bool:
+        return self._lora_manager.remove_lora(lora_id)
+
+    def remove_all_loras(self):
+        self._lora_manager.remove_all_loras()
+
+    def list_loras(self) -> Set[int]:
+        return set(self._lora_manager.list_loras())
+
+
+class LRUCacheWorkerLoRAManager(WorkerLoRAManager):
+    """WorkerLoRAManager that manages LoRA models on the worker side.
+
+    Uses an LRU Cache. Every request, the requested LoRAs will be loaded
+    (unless they are already loaded) and least recently used LoRAs will
+    be unloaded if the cache is above capacity."""
+
+    _lora_manager_cls: Type[
+        LRUCacheLoRAModelManager] = LRUCacheLoRAModelManager
+
+    def create_lora_manager(
+        self,
+        model: torch.nn.Module,
+    ) -> Any:
+        lora_manager = create_lora_manager(
+            model,
+            lora_manager_cls=self._lora_manager_cls,
+            max_num_seqs=self.max_num_seqs,
+            vocab_size=self.vocab_size,
+            lora_config=self.lora_config,
+            max_num_batched_tokens=self.max_num_batched_tokens,
+        )
+        self._lora_manager = lora_manager
+        return lora_manager.model
+
+    def _apply_loras(self, lora_requests: Set[LoRARequest]) -> None:
+        loras_map = {
+            lora_request.lora_int_id: lora_request
+            for lora_request in lora_requests if lora_request
+        }
+        if len(loras_map) > self._lora_manager.lora_slots:
+            raise RuntimeError(
+                f"Number of requested LoRAs ({len(loras_map)}) is greater "
+                "than the number of GPU LoRA slots "
+                f"({self._lora_manager.lora_slots}).")
+        for lora in loras_map.values():
+            self.add_lora(lora)
+
+    def add_lora(self, lora_request: LoRARequest) -> bool:
+        if lora_request.lora_int_id not in self.list_loras():
+            # Remove before we load the new lora to save memory
+            if len(self._lora_manager) + 1 > self._lora_manager.capacity:
+                assert isinstance(self._lora_manager, LRUCacheLoRAModelManager)
+                self._lora_manager.remove_oldest_lora()
+            lora = self._load_lora(lora_request)
+            loaded = self._lora_manager.add_lora(lora)
+        else:
+            # If the lora is already loaded, just touch it to
+            # update its position in the caches
+            loaded = self._lora_manager.get_lora(
+                lora_request.lora_int_id) is not None
+        self._lora_manager.activate_lora(lora_request.lora_int_id)
+        return loaded

vllm/model_executor/__init__.py

@@ -0,0 +1,7 @@
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.model_executor.utils import set_random_seed
+
+__all__ = [
+    "SamplingMetadata",
+    "set_random_seed",
+]

vllm/model_executor/guided_decoding/__init__.py

@@ -0,0 +1,25 @@
+from typing import Optional, Union
+
+from vllm.entrypoints.openai.protocol import (ChatCompletionRequest,
+                                              CompletionRequest)
+from vllm.model_executor.guided_decoding.lm_format_enforcer_decoding import (
+    get_lm_format_enforcer_guided_decoding_logits_processor)
+from vllm.model_executor.guided_decoding.outlines_decoding import (
+    get_outlines_guided_decoding_logits_processor)
+from vllm.sampling_params import LogitsProcessor
+
+
+async def get_guided_decoding_logits_processor(
+        guided_decoding_backend: str, request: Union[CompletionRequest,
+                                                     ChatCompletionRequest],
+        tokenizer) -> Optional[LogitsProcessor]:
+    if guided_decoding_backend == 'outlines':
+        return await get_outlines_guided_decoding_logits_processor(
+            request, tokenizer)
+    if guided_decoding_backend == 'lm-format-enforcer':
+        return await get_lm_format_enforcer_guided_decoding_logits_processor(
+            request, tokenizer)
+
+    raise ValueError(
+        f"Unknown guided decoding backend '{guided_decoding_backend}'. "
+        "Must be one of 'outlines, 'lm-format-enforcer'")

vllm/model_executor/guided_decoding/lm_format_enforcer_decoding.py

@@ -0,0 +1,70 @@
+from functools import lru_cache
+from json import loads as json_loads
+from typing import Optional, Union
+
+from lmformatenforcer import (CharacterLevelParser, JsonSchemaParser,
+                              RegexParser, StringParser,
+                              TokenEnforcerTokenizerData, UnionParser)
+from lmformatenforcer.integrations.vllm import (
+    build_vllm_logits_processor, build_vllm_token_enforcer_tokenizer_data)
+from pydantic import BaseModel
+from transformers import PreTrainedTokenizerBase
+
+from vllm.entrypoints.openai.protocol import (ChatCompletionRequest,
+                                              CompletionRequest)
+from vllm.model_executor.guided_decoding.outlines_decoding import (
+    get_outlines_guided_decoding_logits_processor)
+from vllm.sampling_params import LogitsProcessor
+
+
+async def get_lm_format_enforcer_guided_decoding_logits_processor(
+        request: Union[CompletionRequest, ChatCompletionRequest],
+        tokenizer) -> Optional[LogitsProcessor]:
+    """
+    Given an OpenAI-compatible request, check for guided decoding parameters
+    and get the necessary logits processor for the given guide.
+    We cache logit processors by (guide, tokenizer), and on cache hit
+    we make a shallow copy to reuse the same underlying FSM.
+    """
+
+    tokenizer_data = _cached_build_vllm_token_enforcer_tokenizer_data(
+        tokenizer)
+    character_level_parser: CharacterLevelParser
+    if request.guided_json:
+        schema = _normalize_json_schema_object(request.guided_json)
+        character_level_parser = JsonSchemaParser(schema)
+    elif request.guided_choice:
+        character_level_parser = UnionParser(
+            [StringParser(choice) for choice in request.guided_choice])
+    elif request.guided_regex:
+        character_level_parser = RegexParser(request.guided_regex)
+    elif request.guided_grammar:
+        # CFG grammar not supported by LMFE, revert to outlines
+        return await get_outlines_guided_decoding_logits_processor(
+            request, tokenizer)
+    elif (request.response_format is not None
+          and request.response_format.type == "json_object"):
+        character_level_parser = JsonSchemaParser(
+            None)  # None means any json object
+    else:
+        return None
+
+    logits_processor = build_vllm_logits_processor(tokenizer_data,
+                                                   character_level_parser)
+    return logits_processor
+
+
+def _normalize_json_schema_object(schema: Union[str, dict, BaseModel]) -> dict:
+    if isinstance(schema, str):
+        return json_loads(schema)
+    if isinstance(schema, dict):
+        return schema
+    if isinstance(schema, BaseModel):
+        return schema.model_json_schema()
+    raise AssertionError(f"Unsupported schema type {schema}")
+
+
+@lru_cache
+def _cached_build_vllm_token_enforcer_tokenizer_data(
+        tokenizer: PreTrainedTokenizerBase) -> TokenEnforcerTokenizerData:
+    return build_vllm_token_enforcer_tokenizer_data(tokenizer)