vllm-cpu-amxbf16 0.9.1__cp312-cp312-manylinux_2_17_x86_64.whl

vllm/lora/worker_manager.py

@@ -0,0 +1,259 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from contextlib import contextmanager
+from typing import Any, Literal, Optional, Union
+
+import torch
+
+from vllm.adapter_commons.utils import (add_adapter_worker,
+                                        apply_adapters_worker,
+                                        list_adapters_worker,
+                                        set_active_adapters_worker)
+from vllm.adapter_commons.worker_manager import AbstractWorkerManager
+from vllm.config import LoRAConfig
+from vllm.logger import init_logger
+from vllm.lora.models import (LoRAModel, LoRAModelManager,
+                              LRUCacheLoRAModelManager, create_lora_manager)
+from vllm.lora.peft_helper import PEFTHelper
+from vllm.lora.request import LoRARequest
+from vllm.lora.utils import get_adapter_absolute_path
+
+logger = init_logger(__name__)
+
+
+class WorkerLoRAManager(AbstractWorkerManager):
+    """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."""
+
+    _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,
+        max_position_embeddings: Optional[int] = None,
+    ):
+        self._lora_model_cls = lora_model_cls
+        self.embedding_modules = embedding_modules
+        self.embedding_padding_modules = embedding_padding_modules
+        self._cached_dummy_lora: Union[None, Literal[False], LoRAModel] = False
+        self.max_num_seqs = max_num_seqs
+        self.max_num_batched_tokens = max_num_batched_tokens
+        self.vocab_size = vocab_size
+        self.lora_config = lora_config
+        self.max_position_embeddings = max_position_embeddings
+        super().__init__(device)
+        # Lazily initialized by create_lora_manager.
+        self._adapter_manager: LoRAModelManager
+
+    @contextmanager
+    def dummy_lora_cache(self):
+        """Use this context manager to reuse the dummy lora model
+        to avoid creating it repeatedly."""
+        self._cached_dummy_lora = None
+        yield
+        self._cached_dummy_lora = False
+
+    @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,
+            device=self.device,
+            lora_manager_cls=self._manager_cls,
+        )
+        self._adapter_manager = lora_manager
+        return lora_manager.model
+
+    def _load_adapter(self, lora_request: LoRARequest) -> LoRAModel:
+        try:
+            supported_lora_modules = (
+                self._adapter_manager.supported_lora_modules)
+            packed_modules_mapping = (
+                self._adapter_manager.packed_modules_mapping)
+            expected_lora_modules: list[str] = []
+            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)
+
+            expected_lora_modules = list(set(expected_lora_modules))
+            lora_path = get_adapter_absolute_path(lora_request.lora_path)
+
+            peft_helper = PEFTHelper.from_local_dir(
+                lora_path, self.max_position_embeddings,
+                lora_request.tensorizer_config_dict)
+
+            # Validates the LoRA configuration against requirements before
+            # loading weights, throwing an exception if validation fails.
+            peft_helper.validate_legal(self.lora_config)
+
+            # For some models like Qwen2VL, we need to use hf_to_vllm_mapper
+            # to ensure correct loading of lora weights.
+            model = self._adapter_manager.model
+            hf_to_vllm_mapper = None
+            if (hasattr(model, "hf_to_vllm_mapper")
+                    and model.hf_to_vllm_mapper is not None):
+                hf_to_vllm_mapper = model.hf_to_vllm_mapper
+
+            lora = self._lora_model_cls.from_local_checkpoint(
+                lora_path,
+                expected_lora_modules,
+                peft_helper=peft_helper,
+                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,
+                tensorizer_config_dict=lora_request.tensorizer_config_dict,
+                weights_mapper=hf_to_vllm_mapper)
+
+        except FileNotFoundError as e:
+            # FileNotFoundError should be raised if both
+            # - No adapter found to download from huggingface (or in
+            #       offline mode)
+            # - No local adapter files found at `lora_request.lora_path`
+            # For NotFoundError
+            raise ValueError(
+                f"Loading lora {lora_request.lora_name} failed: No adapter "
+                f"found for {lora_request.lora_path}") from e
+        except Exception as e:
+            # For BadRequestError
+            raise e
+
+        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_adapters():
+            return False
+        if isinstance(self._cached_dummy_lora, LoRAModel):
+            dummy_lora = self._cached_dummy_lora.clone(
+                lora_request.lora_int_id)
+        else:
+            dummy_lora = self._adapter_manager.create_dummy_lora(
+                lora_request.lora_int_id, rank, 1, self.embedding_modules)
+            if self._cached_dummy_lora is None:
+                self._cached_dummy_lora = dummy_lora
+        return self._adapter_manager.add_adapter(dummy_lora)
+
+    def pin_adapter(self, adapter_id: int) -> bool:
+        return self._adapter_manager.pin_adapter(adapter_id)
+
+    def set_active_adapters(self, requests: set[Any],
+                            mapping: Optional[Any]) -> None:
+        set_active_adapters_worker(requests, mapping, self._apply_adapters,
+                                   self._adapter_manager.set_adapter_mapping)
+
+    def _apply_adapters(self, adapter_requests: set[Any]) -> None:
+        apply_adapters_worker(adapter_requests, self.list_adapters,
+                              self._adapter_manager.adapter_slots,
+                              self.remove_adapter, self.add_adapter)
+
+    def add_adapter(self, adapter_request: Any) -> bool:
+        return add_adapter_worker(adapter_request, self.list_adapters,
+                                  self._load_adapter,
+                                  self._adapter_manager.add_adapter,
+                                  self._adapter_manager.activate_adapter)
+
+    def remove_adapter(self, adapter_id: int) -> bool:
+        return self._adapter_manager.remove_adapter(adapter_id)
+
+    def remove_all_adapters(self):
+        self._adapter_manager.remove_all_adapters()
+
+    def list_adapters(self) -> set[int]:
+        return list_adapters_worker(self._adapter_manager.list_adapters)
+
+
+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."""
+
+    _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._manager_cls,
+            max_num_seqs=self.max_num_seqs,
+            vocab_size=self.vocab_size,
+            lora_config=self.lora_config,
+            device=self.device,
+            max_num_batched_tokens=self.max_num_batched_tokens,
+        )
+        self._adapter_manager = lora_manager
+        return lora_manager.model
+
+    def _apply_adapters(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._adapter_manager.lora_slots:
+            raise RuntimeError(
+                f"Number of requested LoRAs ({len(loras_map)}) is greater "
+                "than the number of GPU LoRA slots "
+                f"({self._adapter_manager.lora_slots}).")
+        for lora in loras_map.values():
+            self.add_adapter(lora)
+
+    def add_adapter(self, lora_request: LoRARequest) -> bool:
+        # Note that this method is not thread-safe. It may be invoked multiple
+        # times for the same adapter when using multiple API servers.
+        # This is ok because it's currently only called from
+        # the single-threaded core engine loop.
+
+        if lora_request.lora_int_id not in self.list_adapters():
+            # Load the new adapter first to ensure it is actually valid, before
+            # evicting any existing adapters.
+            # This may cause the # of loaded lora adapters to very temporarily
+            # exceed `--max-cpu-loras`.
+            lora = self._load_adapter(lora_request)
+
+            # Loading succeeded, now check if we will exceed cache capacity and
+            # evict if the oldest adapter if so
+            if len(self._adapter_manager) + 1 > self._adapter_manager.capacity:
+                assert isinstance(self._adapter_manager,
+                                  LRUCacheLoRAModelManager)
+                self._adapter_manager.remove_oldest_adapter()
+            # Then add the new adapter to the cache
+            loaded = self._adapter_manager.add_adapter(lora)
+        else:
+            # If the lora is already loaded, just touch it to
+            # update its position in the caches
+            loaded = self._adapter_manager.get_adapter(
+                lora_request.lora_int_id) is not None
+        self._adapter_manager.activate_adapter(lora_request.lora_int_id)
+        return loaded

vllm/model_executor/__init__.py

@@ -0,0 +1,16 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from vllm.model_executor.parameter import (BasevLLMParameter,
+                                           PackedvLLMParameter)
+from vllm.model_executor.sampling_metadata import (SamplingMetadata,
+                                                   SamplingMetadataCache)
+from vllm.model_executor.utils import set_random_seed
+
+__all__ = [
+    "SamplingMetadata",
+    "SamplingMetadataCache",
+    "set_random_seed",
+    "BasevLLMParameter",
+    "PackedvLLMParameter",
+]

vllm/model_executor/custom_op.py

@@ -0,0 +1,152 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import torch.nn as nn
+
+from vllm.config import get_current_vllm_config
+from vllm.logger import init_logger
+from vllm.platforms import current_platform
+
+logger = init_logger(__name__)
+
+
+class CustomOp(nn.Module):
+    """
+    Base class for custom ops.
+    Dispatches the forward method to the appropriate backend.
+    """
+
+    def __init__(self):
+        super().__init__()
+        self._forward_method = self.dispatch_forward()
+
+    def forward(self, *args, **kwargs):
+        return self._forward_method(*args, **kwargs)
+
+    def forward_native(self, *args, **kwargs):
+        """PyTorch-native implementation of the forward method.
+        This method is optional. If implemented, it can be used with compilers
+        such as torch.compile or PyTorch XLA. Also, it can be used for testing
+        purposes.
+        """
+        raise NotImplementedError
+
+    def forward_cuda(self, *args, **kwargs):
+        raise NotImplementedError
+
+    def forward_hip(self, *args, **kwargs):
+        # By default, we assume that HIP ops are compatible with CUDA ops.
+        return self.forward_cuda(*args, **kwargs)
+
+    def forward_xpu(self, *args, **kwargs):
+        # By default, we assume that XPU ops are compatible with the
+        # PyTorch-native implementation.
+        return self.forward_native(*args, **kwargs)
+
+    def forward_cpu(self, *args, **kwargs):
+        # By default, we assume that CPU ops are compatible with CUDA ops.
+        return self.forward_cuda(*args, **kwargs)
+
+    def forward_tpu(self, *args, **kwargs):
+        # By default, we assume that TPU ops are compatible with the
+        # PyTorch-native implementation.
+        # NOTE(woosuk): This is a placeholder for future extensions.
+        return self.forward_native(*args, **kwargs)
+
+    def forward_hpu(self, *args, **kwargs):
+        # By default, we assume that Gaudi ops are compatible with the
+        # PyTorch-native implementation.
+        return self.forward_native(*args, **kwargs)
+
+    def forward_neuron(self, *args, **kwargs):
+        # By default, we assume that Neuron ops are compatible with the
+        # PyTorch-native implementation.
+        return self.forward_native(*args, **kwargs)
+
+    def forward_oot(self, *args, **kwargs):
+        # By default, we assume that OOT ops are compatible with the
+        # PyTorch-native implementation.
+        return self.forward_native(*args, **kwargs)
+
+    def dispatch_forward(self):
+        # NOTE(woosuk): Here we assume that vLLM was built for only one
+        # specific backend. Currently, we do not support dynamic dispatching.
+        compilation_config = get_current_vllm_config().compilation_config
+        enabled = self.enabled()
+        if enabled:
+            compilation_config.enabled_custom_ops.update([self.__class__.name])
+        else:
+            compilation_config.disabled_custom_ops.update(
+                [self.__class__.name])
+
+        if not enabled:
+            return self.forward_native
+
+        if current_platform.is_rocm():
+            return self.forward_hip
+        elif current_platform.is_cpu():
+            return self.forward_cpu
+        elif current_platform.is_hpu():
+            return self.forward_hpu
+        elif current_platform.is_tpu():
+            return self.forward_tpu
+        elif current_platform.is_xpu():
+            return self.forward_xpu
+        elif current_platform.is_neuron():
+            return self.forward_neuron
+        elif current_platform.is_out_of_tree():
+            return self.forward_oot
+        else:
+            return self.forward_cuda
+
+    @classmethod
+    def enabled(cls) -> bool:
+        # if no name, then it was not registered
+        compilation_config = get_current_vllm_config().compilation_config
+        custom_ops = compilation_config.custom_ops
+        if not hasattr(cls, "name"):
+            logger.warning_once(
+                "Custom op %s was not registered, which means it won't appear in the op registry. It will be enabled/disabled based on the global settings.",  # noqa: E501
+                cls.__name__,
+            )
+            return CustomOp.default_on()
+
+        enabled = f"+{cls.name}" in custom_ops
+        disabled = f"-{cls.name}" in custom_ops
+        assert not (enabled
+                    and disabled), f"Cannot enable and disable {cls.name}"
+
+        return (CustomOp.default_on() or enabled) and not disabled
+
+    @staticmethod
+    def default_on() -> bool:
+        """
+        On by default if level < CompilationLevel.PIECEWISE
+        Specifying 'all' or 'none' in custom_op takes precedence.
+        """
+        from vllm.config import CompilationLevel
+        compilation_config = get_current_vllm_config().compilation_config
+        custom_ops = compilation_config.custom_ops
+        count_none = custom_ops.count("none")
+        count_all = custom_ops.count("all")
+        return compilation_config.level < CompilationLevel.PIECEWISE and \
+            not count_none > 0 or count_all > 0
+
+    # Dictionary of all custom ops (classes, indexed by registered name).
+    # To check if an op with a name is enabled, call .enabled() on the class.
+    # Examples:
+    # - MyOp.enabled()
+    # - op_registry["my_op"].enabled()
+    op_registry: dict[str, type['CustomOp']] = {}
+
+    # Decorator to register custom ops.
+    @classmethod
+    def register(cls, name: str):
+
+        def decorator(op_cls):
+            assert name not in cls.op_registry, f"Duplicate op name: {name}"
+            op_cls.name = name
+            cls.op_registry[name] = op_cls
+            return op_cls
+
+        return decorator

vllm/model_executor/guided_decoding/__init__.py

@@ -0,0 +1,181 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+from vllm.logger import init_logger
+from vllm.model_executor.guided_decoding.utils import (
+    convert_lark_to_gbnf, grammar_is_likely_lark,
+    has_lmf_unsupported_json_features, has_xgrammar_unsupported_json_features)
+from vllm.reasoning import ReasoningParserManager
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedTokenizer
+
+    from vllm.config import ModelConfig
+    from vllm.logits_process import LogitsProcessor
+    from vllm.sampling_params import GuidedDecodingParams
+
+logger = init_logger(__name__)
+
+
+def maybe_backend_fallback(
+        guided_params: GuidedDecodingParams) -> GuidedDecodingParams:
+
+    def fallback_or_error(guided_params: GuidedDecodingParams, message: str,
+                          fallback: str) -> None:
+        """Change the backend to the specified fallback with a warning log,
+        or raise a ValueError if the `disable_fallback` option is specified."""
+        if guided_params.disable_fallback:
+            raise ValueError(message)
+
+        logger.warning("%s Falling back to use %s instead.", message, fallback)
+        guided_params.backend = fallback
+
+    # `auto` was added for V1 to explicitly declare a mode that has fallbacks
+    # in place. If that is specified with V0, treat it as `xgrammar`, as we have
+    # fallbacks enabled for that and it is the V0 default.
+    if guided_params.backend == "auto":
+        guided_params.backend = "xgrammar"
+
+    # lm-format-enforce doesn't support grammar, fallback to xgrammar
+    if guided_params.backend == "lm-format-enforcer":
+        if guided_params.grammar is not None:
+            fallback_or_error(
+                guided_params,
+                "lm-format-enforcer does not support grammar guided decoding.",
+                "xgrammar")
+
+        # lm-format-enforcer doesn't support some JSON schema features
+        elif (guided_params.json is not None
+              and has_lmf_unsupported_json_features(guided_params.json)):
+            fallback_or_error(
+                guided_params,
+                "lm-format-enforcer does not support advanced JSON schema "
+                "features like patterns or numeric ranges.", "outlines")
+
+    if guided_params.backend == "xgrammar":
+        from vllm.model_executor.guided_decoding.xgrammar_decoding import (
+            xgr_installed)
+
+        # xgrammar doesn't support some JSON schema features
+        if (guided_params.json is not None and
+                has_xgrammar_unsupported_json_features(guided_params.json)):
+            fallback_or_error(
+                guided_params,
+                "xgrammar does not support advanced JSON schema features like "
+                "string length, item limits, or property bounds.", "outlines")
+
+        # xgrammar only supports GBNF grammars, so we must convert Lark.
+        # We must check if the grammar is likely Lark and if that
+        # grammar is convertible to GBNF
+        elif (guided_params.grammar is not None
+              and grammar_is_likely_lark(guided_params.grammar)):
+            try:
+                convert_lark_to_gbnf(guided_params.grammar)
+            except Exception:
+                fallback_or_error(
+                    guided_params,
+                    "xgrammar does not support Lark grammars and the "
+                    "grammar failed to convert to GBNF.", "outlines")
+
+        # If the xgrammar module cannot be imported successfully,
+        # we should still allow users to use guided decoding with a fallback.
+        elif not xgr_installed:
+            fallback_or_error(
+                guided_params,
+                "xgrammar module cannot be imported successfully.", "outlines")
+
+    if (guided_params.backend == "outlines"
+            and guided_params.json_object is not None):
+        # outlines doesn't support json_object, fallback to guidance
+        fallback_or_error(guided_params,
+                          "outlines does not support json_object.", "guidance")
+
+    return guided_params
+
+
+async def get_guided_decoding_logits_processor(
+        guided_params: GuidedDecodingParams,
+        tokenizer: PreTrainedTokenizer,
+        model_config: ModelConfig,
+        reasoning_backend: str | None = None) -> LogitsProcessor | None:
+
+    reasoner = None
+    if reasoning_backend:
+        reasoner_class = ReasoningParserManager.get_reasoning_parser(
+            reasoning_backend)
+        reasoner = reasoner_class(tokenizer)
+
+    guided_params = maybe_backend_fallback(guided_params)
+
+    # CFG grammar not supported by LMFE, so we use outlines instead
+    if guided_params.backend == 'outlines':
+        # NOTE: lazy import outlines to avoid https://github.com/vllm-project/vllm/issues/4193
+        from vllm.model_executor.guided_decoding.outlines_decoding import (  # noqa
+            get_outlines_guided_decoding_logits_processor)
+        return await get_outlines_guided_decoding_logits_processor(
+            guided_params, tokenizer, reasoner)
+    if guided_params.backend == 'lm-format-enforcer':
+        from vllm.model_executor.guided_decoding.lm_format_enforcer_decoding import (  # noqa
+            get_local_lm_format_enforcer_guided_decoding_logits_processor)
+        return get_local_lm_format_enforcer_guided_decoding_logits_processor(
+            guided_params, tokenizer)
+    if guided_params.backend == 'xgrammar':
+        from vllm.model_executor.guided_decoding.xgrammar_decoding import (  # noqa
+            get_local_xgrammar_guided_decoding_logits_processor)
+        return get_local_xgrammar_guided_decoding_logits_processor(
+            guided_params, tokenizer, model_config, reasoner)
+    if guided_params.backend == 'guidance':
+        from vllm.model_executor.guided_decoding.guidance_decoding import (
+            get_local_guidance_guided_decoding_logits_processor)
+        return get_local_guidance_guided_decoding_logits_processor(
+            guided_params, tokenizer)
+    raise ValueError(
+        f"Unknown guided decoding backend '{guided_params.backend}'. "
+        "Must be one of 'outlines, 'lm-format-enforcer', 'xgrammar', 'guidance'"
+    )
+
+
+def get_local_guided_decoding_logits_processor(
+        guided_params: GuidedDecodingParams,
+        tokenizer: PreTrainedTokenizer,
+        model_config: ModelConfig,
+        reasoning_backend: str | None = None) -> LogitsProcessor | None:
+    guided_params = maybe_backend_fallback(guided_params)
+
+    reasoner = None
+    if reasoning_backend:
+        reasoner_class = ReasoningParserManager.get_reasoning_parser(
+            reasoning_backend)
+        reasoner = reasoner_class(tokenizer)
+
+    # CFG grammar not supported by LMFE, so we use outlines instead
+    if guided_params.backend == 'outlines':
+        # NOTE: lazy import outlines to avoid https://github.com/vllm-project/vllm/issues/4193
+        from vllm.model_executor.guided_decoding.outlines_decoding import (  # noqa
+            get_local_outlines_guided_decoding_logits_processor)
+        return get_local_outlines_guided_decoding_logits_processor(
+            guided_params, tokenizer, reasoner)
+    if guided_params.backend == 'lm-format-enforcer':
+        from vllm.model_executor.guided_decoding.lm_format_enforcer_decoding import (  # noqa
+            get_local_lm_format_enforcer_guided_decoding_logits_processor)
+        return get_local_lm_format_enforcer_guided_decoding_logits_processor(
+            guided_params, tokenizer)
+    if guided_params.backend == 'xgrammar':
+        from vllm.model_executor.guided_decoding.xgrammar_decoding import (  # noqa
+            get_local_xgrammar_guided_decoding_logits_processor)
+        return get_local_xgrammar_guided_decoding_logits_processor(
+            guided_params, tokenizer, model_config, reasoner)
+    if guided_params.backend == 'guidance':
+        from vllm.model_executor.guided_decoding.guidance_decoding import (
+            get_local_guidance_guided_decoding_logits_processor)
+        return get_local_guidance_guided_decoding_logits_processor(
+            guided_params, tokenizer)
+
+    raise ValueError(
+        f"Unknown guided decoding backend '{guided_params.backend}'. "
+        "Must be one of 'outlines, 'lm-format-enforcer', 'xgrammar', 'guidance'"
+    )

vllm/model_executor/guided_decoding/guidance_decoding.py

@@ -0,0 +1,63 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import json
+
+import llguidance
+from regex import escape as regex_escape
+from transformers import PreTrainedTokenizerBase
+
+from vllm.model_executor.guided_decoding.guidance_logits_processors import (
+    GuidanceLogitsProcessor)
+from vllm.sampling_params import GuidedDecodingParams
+from vllm.v1.structured_output.backend_guidance import (
+    process_for_additional_properties)
+
+
+def get_local_guidance_guided_decoding_logits_processor(
+        guided_params: GuidedDecodingParams,
+        tokenizer: PreTrainedTokenizerBase) -> GuidanceLogitsProcessor:
+    """
+    Given an OpenAI-compatible request, check for guided decoding parameters
+    and get the necessary logits processor for the given guide.
+    """
+
+    grm = ""
+    any_whitespace = not guided_params.disable_any_whitespace
+    if (guide_json := guided_params.json) is not None:
+        # Optionally set additionalProperties to False at the top-level
+        # By default, other backends do not allow additional top-level
+        # properties, so this makes guidance more similar to other backends
+        if guided_params.disable_additional_properties:
+            if not isinstance(guide_json, str):
+                guide_json = json.dumps(guide_json)
+            guide_json = process_for_additional_properties(guide_json)
+
+        grm = llguidance.LLMatcher.grammar_from_json_schema(
+            guide_json,
+            overrides={"whitespace_pattern": guided_params.whitespace_pattern},
+            defaults={
+                "whitespace_flexible": any_whitespace,
+            })
+    elif guided_params.json_object:
+        grm = llguidance.LLMatcher.grammar_from_json_schema(
+            '{"type": "object"}',
+            overrides={"whitespace_pattern": guided_params.whitespace_pattern},
+            defaults={
+                "whitespace_flexible": any_whitespace,
+            })
+    elif guided_params.regex:
+        grm = llguidance.grammar_from("regex", guided_params.regex)
+    elif guided_params.choice:
+        # choice just uses regex
+        choices = (regex_escape(str(choice))
+                   for choice in guided_params.choice)
+        choices_regex = "(" + "|".join(choices) + ")"
+        grm = llguidance.grammar_from("regex", choices_regex)
+    elif guided_params.grammar:
+        # this supports Lark and GBNF
+        grm = llguidance.grammar_from("grammar", guided_params.grammar)
+
+    if grm:
+        return GuidanceLogitsProcessor(grm, tokenizer)
+
+    raise ValueError("Unknown guided decoding mode")