vllm-cpu 0.8.5.post2__cp310-cp310-manylinux_2_17_x86_64.whl

vllm/v1/structured_output/backend_guidance.py

@@ -0,0 +1,215 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import copy
+import json
+import os
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Any, Optional, Union
+
+import torch
+
+from vllm.config import VllmConfig
+from vllm.logger import init_logger
+from vllm.sampling_params import GuidedDecodingParams, SamplingParams
+from vllm.transformers_utils.tokenizer_group import init_tokenizer_from_configs
+from vllm.utils import LazyLoader
+from vllm.v1.structured_output.backend_types import (StructuredOutputBackend,
+                                                     StructuredOutputGrammar,
+                                                     StructuredOutputOptions)
+from vllm.v1.structured_output.request import get_structured_output_key
+
+if TYPE_CHECKING:
+    import llguidance
+    import llguidance.hf as llguidance_hf
+    import llguidance.torch as llguidance_torch
+else:
+    llguidance = LazyLoader("llguidance", globals(), "llguidance")
+    llguidance_hf = LazyLoader("llguidance.hf", globals(), "llguidance.hf")
+    llguidance_torch = LazyLoader("llguidance.torch", globals(),
+                                  "llguidance.torch")
+
+logger = init_logger(__name__)
+
+
+def _walk_json_for_additional_properties(data: object):
+    if isinstance(data, dict):
+        for value in data.values():
+            _walk_json_for_additional_properties(value)
+        if 'additionalProperties' not in data and \
+            ('properties' in data or 'patternProperties' in data):
+            data['additionalProperties'] = False
+    elif isinstance(data, list):
+        for item in data:
+            _walk_json_for_additional_properties(item)
+
+
+def process_for_additional_properties(
+        guide_json: Union[str, dict[str, Any]]) -> dict[str, Any]:
+    if isinstance(guide_json, str):
+        guide_json_obj = json.loads(guide_json)
+    else:
+        # copy for modifications
+        guide_json_obj = copy.deepcopy(guide_json)
+    _walk_json_for_additional_properties(guide_json_obj)
+    return guide_json_obj
+
+
+class GuidanceBackend(StructuredOutputBackend):
+
+    def __init__(self, vllm_config: VllmConfig):
+        self.vllm_config = vllm_config
+        tokenizer_group = init_tokenizer_from_configs(
+            model_config=vllm_config.model_config,
+            scheduler_config=vllm_config.scheduler_config,
+            lora_config=vllm_config.lora_config)  # type: ignore[arg-type]
+        self.vllm_config = vllm_config
+        self.vocab_size = vllm_config.model_config.get_vocab_size()
+
+        self.disable_any_whitespace = False
+        self.no_additional_properties = False
+        backend_options = GuidedDecodingParams(
+            backend=vllm_config.decoding_config.guided_decoding_backend
+        ).backend_options()
+        for option in backend_options:
+            if option == "disable-any-whitespace":
+                self.disable_any_whitespace = True
+            elif option == "no-additional-properties":
+                self.no_additional_properties = True
+            else:
+                raise ValueError(
+                    f"Unsupported option for the guidance backend: {option}")
+
+        tokenizer = tokenizer_group.get_lora_tokenizer(None)
+        self.ll_tokenizer = llguidance_hf.from_tokenizer(
+            tokenizer, self.vocab_size)
+
+    def compile_grammar(self, request_type: StructuredOutputOptions,
+                        grammar_spec: str) -> StructuredOutputGrammar:
+        self.serialized_grammar = serialize_guidance_grammar(
+            request_type, grammar_spec, self.disable_any_whitespace,
+            self.no_additional_properties)
+
+        ll_matcher = llguidance.LLMatcher(
+            self.ll_tokenizer,
+            self.serialized_grammar,
+            log_level=int(os.environ.get("LLGUIDANCE_LOG_LEVEL", "1")),
+        )
+
+        r = GuidanceGrammar(
+            ll_matcher=ll_matcher,
+            ll_tokenizer=self.ll_tokenizer,
+            vocab_size=self.vocab_size,
+        )
+
+        r.check_error()
+        return r
+
+    def allocate_token_bitmask(self, max_num_seqs: int):
+        return llguidance_torch.allocate_token_bitmask(
+            max_num_seqs, self.ll_tokenizer.vocab_size)
+
+    def destroy(self):
+        pass
+
+
+@dataclass
+class GuidanceGrammar(StructuredOutputGrammar):
+    ll_matcher: llguidance.LLMatcher
+    ll_tokenizer: llguidance.LLTokenizer
+    vocab_size: int
+    printed_error: bool = False
+    terminated: bool = False
+
+    def check_error(self):
+        if not self.printed_error:
+            err = self.ll_matcher.get_error()
+            if err:
+                self.printed_error = True
+                logger.warning("LLMatcher error: %s", err)
+
+    def accept_tokens(self, request_id: str, tokens: list[int]) -> bool:
+        """Accepts a list of tokens and advances the parser.
+
+        Returns True if the parser was advanced successfully.
+        Returns False if the parser failed to advance.
+        """
+
+        if self.ll_tokenizer.eos_token in tokens:
+            self.terminated = True
+
+        if self.ll_matcher.is_stopped():
+            return True
+
+        # TODO - Add jump decoding support in the future:
+        # self.ll_matcher.compute_ff_bytes() - this should always work
+        # self.ll_matcher.compute_ff_tokens() - this only works for
+        #   "canonical" tokenizers
+        # For conversion between the two, see
+        # https://github.com/guidance-ai/llguidance/blob/main/docs/fast_forward.md
+
+        r = self.ll_matcher.consume_tokens(tokens)
+
+        self.check_error()
+
+        return r
+
+    def fill_bitmask(self, bitmask: torch.Tensor, idx: int) -> None:
+        # this will automatically return [EOS] mask if the matcher is stopped
+        # or otherwise in an error state
+        llguidance_torch.fill_next_token_bitmask(self.ll_matcher, bitmask, idx)
+        self.check_error()
+
+    def is_terminated(self) -> bool:
+        return self.terminated
+
+    def reset(self):
+        # This method may be not needed anymore? TODO
+        self.ll_matcher.reset()
+
+
+def serialize_guidance_grammar(
+    request_type: StructuredOutputOptions,
+    grammar_spec: Union[str, dict[str, Any]],
+    disable_any_whitespace: bool = False,
+    no_additional_properties: bool = False,
+) -> str:
+    if request_type == StructuredOutputOptions.JSON:
+        if no_additional_properties:
+            grammar_spec = process_for_additional_properties(grammar_spec)
+        return llguidance.LLMatcher.grammar_from_json_schema(
+            grammar_spec,
+            defaults={
+                "whitespace_flexible": not disable_any_whitespace,
+            })
+    elif request_type == StructuredOutputOptions.JSON_OBJECT:
+        return llguidance.LLMatcher.grammar_from_json_schema(
+            '{"type": "object"}',
+            defaults={
+                "whitespace_flexible": not disable_any_whitespace,
+            })
+    else:
+        if request_type == StructuredOutputOptions.REGEX:
+            tp = "regex"
+        elif request_type == StructuredOutputOptions.GRAMMAR:
+            tp = "grammar"
+        elif request_type == StructuredOutputOptions.CHOICE:
+            tp = "choice"
+        elif request_type == StructuredOutputOptions.STRUCTURAL_TAG:
+            raise ValueError("Structural tag is not supported "
+                             "for guidance backend yet")
+        else:
+            logger.error("Validation should have already occurred. "
+                         "Please file an issue.")
+            raise ValueError("grammar is not of valid supported types. "
+                             f"({request_type!s})")
+        return llguidance.grammar_from(tp, grammar_spec)
+
+
+def validate_guidance_grammar(
+        sampling_params: SamplingParams,
+        tokenizer: Optional[llguidance.LLTokenizer] = None) -> None:
+    tp, grm = get_structured_output_key(sampling_params)
+    guidance_grm = serialize_guidance_grammar(tp, grm)
+    err = llguidance.LLMatcher.validate_grammar(guidance_grm, tokenizer)
+    if err:
+        raise ValueError(f"Grammar error: {err}")

vllm/v1/structured_output/backend_types.py

@@ -0,0 +1,96 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import enum
+from abc import ABC, abstractmethod
+
+import torch
+
+
+class StructuredOutputOptions(enum.Enum):
+    JSON = enum.auto()
+    JSON_OBJECT = enum.auto()
+    REGEX = enum.auto()
+    GRAMMAR = enum.auto()
+    CHOICE = enum.auto()
+    STRUCTURAL_TAG = enum.auto()
+
+
+StructuredOutputKey = tuple[StructuredOutputOptions, str]
+
+
+class StructuredOutputGrammar(ABC):
+    """Request-level backend for structured output requests."""
+
+    @abstractmethod
+    def accept_tokens(self, request_id: str, tokens: list[int]) -> bool:
+        """
+        Determines whether the provided tokens are accepted for the
+        given request.
+
+        Args:
+            request_id (str): The unique identifier for the request.
+            tokens (list[int]): A list of token IDs to evaluate.
+
+        Returns:
+            bool: True if the tokens are accepted, False otherwise.
+        """
+
+    @abstractmethod
+    def fill_bitmask(self, bitmask: torch.Tensor, batch_index: int) -> None:
+        """
+        Fills the bitmask for a specific batch index.
+
+        Args:
+            bitmask (torch.Tensor): The bitmask to fill
+            batch_index (int): The index in the bitmask to fill
+        """
+
+    @abstractmethod
+    def is_terminated(self) -> bool:
+        """
+        Checks whether the structured output process has terminated.
+
+        Returns:
+            bool: True if the process is terminated, False otherwise.
+        """
+
+    @abstractmethod
+    def reset(self):
+        """
+        Resets the state of the structured output grammar.
+        """
+
+
+class StructuredOutputBackend(ABC):
+    """Engine-level backend for structured output requests."""
+
+    @abstractmethod
+    def compile_grammar(self, request_type: StructuredOutputOptions,
+                        grammar_spec: str) -> StructuredOutputGrammar:
+        """
+        Compiles a grammar specification into a structured output grammar.
+
+        Args:
+            request_type (StructuredOutputOptions): The type of structured
+              output request.
+            grammar_spec (str): The grammar specification to compile.
+
+        Returns:
+            StructuredOutputGrammar: The compiled structured output grammar.
+        """
+
+    @abstractmethod
+    def allocate_token_bitmask(self, max_num_seqs: int):
+        """
+        Allocates a token bitmask for the specified maximum number of sequences.
+
+        Args:
+            max_num_seqs (int): The maximum number of sequences for which
+              to allocate the bitmask.
+        """
+
+    @abstractmethod
+    def destroy(self):
+        """
+        Backend-specific cleanup.
+        """

vllm/v1/structured_output/backend_xgrammar.py

@@ -0,0 +1,299 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import json
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING, Any
+
+import torch
+
+import vllm.envs
+from vllm.config import VllmConfig
+from vllm.logger import init_logger
+from vllm.sampling_params import GuidedDecodingParams, SamplingParams
+from vllm.transformers_utils.tokenizer_group import init_tokenizer_from_configs
+from vllm.transformers_utils.tokenizers.mistral import MistralTokenizer
+from vllm.utils import LazyLoader
+from vllm.v1.structured_output.backend_types import (StructuredOutputBackend,
+                                                     StructuredOutputGrammar,
+                                                     StructuredOutputOptions)
+from vllm.v1.structured_output.utils import (choice_as_grammar,
+                                             convert_lark_to_ebnf,
+                                             grammar_is_likely_lark)
+
+if TYPE_CHECKING:
+    import xgrammar as xgr
+else:
+    xgr = LazyLoader("xgr", globals(), "xgrammar")
+
+logger = init_logger(__name__)
+
+
+class XgrammarBackend(StructuredOutputBackend):
+
+    def __init__(self, vllm_config: VllmConfig):
+        self.vllm_config = vllm_config
+        tokenizer_group = init_tokenizer_from_configs(
+            model_config=vllm_config.model_config,
+            scheduler_config=vllm_config.scheduler_config,
+            lora_config=vllm_config.lora_config)  # type: ignore[arg-type]
+
+        self.disable_any_whitespace = False
+        backend_options = GuidedDecodingParams(
+            backend=vllm_config.decoding_config.guided_decoding_backend
+        ).backend_options()
+        for option in backend_options:
+            if option == "disable-any-whitespace":
+                self.disable_any_whitespace = True
+            else:
+                raise ValueError(
+                    f"Unsupported option for the xgrammar backend: {option}")
+
+        tokenizer = tokenizer_group.get_lora_tokenizer(None)
+        self.vocab_size = vllm_config.model_config.get_vocab_size()
+        if isinstance(tokenizer, MistralTokenizer):
+            # NOTE: ideally, xgrammar should handle this accordingly.
+            # refer to https://github.com/mlc-ai/xgrammar/blob/d77c0a0173ef14779c918e3be7966ba852f7910f/python/xgrammar/tokenizer_info.py#L98
+            try:
+                if tokenizer.is_tekken:
+                    encoded_vocab = tokenizer._vocab
+                else:
+                    encoded_vocab = [
+                        token for token, _ in sorted(
+                            tokenizer.get_vocab().items(),
+                            key=lambda x: x[1],
+                        )
+                    ]
+                stop_token_ids = None
+                if hasattr(
+                        tokenizer,
+                        "eos_token_id",
+                ) and tokenizer.eos_token_id is not None:
+                    stop_token_ids = [tokenizer.eos_token_id]
+            except AttributeError as e:
+                raise ValueError(
+                    f"Cannot get the vocabulary of the tokenizer "
+                    f"{type(tokenizer)}. The tokenizer should have a "
+                    "get_vocab method.") from e
+            tokenizer_info = xgr.TokenizerInfo(  # type: ignore
+                encoded_vocab=encoded_vocab,
+                # NOTE: https://github.com/mlc-ai/xgrammar/blob/5e141f6ff1ca02bc31f9e512e68b61f2a8ae88e5/tests/python/test_tokenizer_info.py#L43 # noqa: E501
+                vocab_type=xgr.VocabType.RAW
+                if tokenizer.is_tekken else xgr.VocabType.BYTE_FALLBACK,
+                vocab_size=self.vocab_size,
+                stop_token_ids=stop_token_ids,
+                add_prefix_space=True,
+            )
+        else:
+            tokenizer_info = xgr.TokenizerInfo.from_huggingface(
+                tokenizer,
+                vocab_size=self.vocab_size,
+            )
+        self.compiler = xgr.GrammarCompiler(
+            tokenizer_info,
+            max_threads=8,
+            cache_enabled=True,
+            cache_limit_bytes=vllm.envs.VLLM_XGRAMMAR_CACHE_MB * 1024 * 1024,
+        )
+
+    def compile_grammar(self, request_type: StructuredOutputOptions,
+                        grammar_spec: str) -> StructuredOutputGrammar:
+        if request_type == StructuredOutputOptions.JSON:
+            ctx = self.compiler.compile_json_schema(
+                grammar_spec, any_whitespace=not self.disable_any_whitespace)
+        elif request_type == StructuredOutputOptions.JSON_OBJECT:
+            ctx = self.compiler.compile_json_schema(
+                '{"type": "object"}',
+                any_whitespace=not self.disable_any_whitespace)
+        elif request_type == StructuredOutputOptions.GRAMMAR:
+            ctx = self.compiler.compile_grammar(grammar_spec)
+        elif request_type == StructuredOutputOptions.REGEX:
+            ctx = self.compiler.compile_regex(grammar_spec)
+        elif request_type == StructuredOutputOptions.STRUCTURAL_TAG:
+            s_tag = json.loads(grammar_spec)
+            tags = [
+                xgr.StructuralTagItem(
+                    begin=s["begin"],
+                    schema=json.dumps(s["schema"]),
+                    end=s["end"],
+                ) for s in s_tag["structures"]
+            ]
+            ctx = self.compiler.compile_structural_tag(tags, s_tag["triggers"])
+        else:
+            logger.error(
+                "Validation should have already occurred. Please file an issue."
+            )
+            raise ValueError(
+                f"grammar is not of valid supported types. ({request_type!s})")
+
+        return XgrammarGrammar(
+            matcher=xgr.GrammarMatcher(ctx),
+            vocab_size=self.vocab_size,
+            ctx=ctx,
+        )
+
+    def allocate_token_bitmask(self, max_num_seqs: int):
+        return xgr.allocate_token_bitmask(max_num_seqs, self.vocab_size)
+
+    def destroy(self):
+        del self.compiler
+
+
+@dataclass
+class XgrammarGrammar(StructuredOutputGrammar):
+    # NOTE: This would be a generic-enough class for
+    # supporting different backends, in the future.
+    # For now, just xgrammar.
+    #
+    # TODO: support max_rollback_tokens
+    # https://xgrammar.mlc.ai/docs/api/python/index.html#xgrammar.GrammarMatcher.find_jump_forward_string
+    # for jump-forward decoding
+
+    vocab_size: int
+    matcher: xgr.GrammarMatcher = field(hash=False)
+    ctx: xgr.CompiledGrammar = field(hash=False)
+    num_processed_tokens: int = field(default_factory=lambda: 0,
+                                      repr=False,
+                                      hash=False,
+                                      init=False)
+
+    def accept_tokens(self, request_id: str, tokens: list[int]) -> bool:
+        """Accepts a list of tokens and advances the FSM.
+
+        Returns True if the FSM was advanced successfully.
+        Returns False if the FSM failed to advance.
+        """
+        for token in tokens:
+            if not self.matcher.accept_token(token):
+                logger.error(
+                    "Failed to advance FSM for request %s "
+                    "for tokens %s. Please file an issue.", request_id, token)
+                return False
+            self.num_processed_tokens += 1
+        return True
+
+    def fill_bitmask(self, bitmask: torch.Tensor, idx: int) -> None:
+        self.matcher.fill_next_token_bitmask(bitmask, idx)
+
+    def is_terminated(self) -> bool:
+        return self.matcher.is_terminated()
+
+    def reset(self):
+        self.num_processed_tokens = 0
+        self.matcher.reset()
+
+
+def has_xgrammar_unsupported_json_features(schema: dict[str, Any]) -> bool:
+    """Check if JSON schema contains features unsupported by xgrammar."""
+
+    def check_object(obj: dict[str, Any]) -> bool:
+        if not isinstance(obj, dict):
+            return False
+
+        # Check for numeric ranges
+        if obj.get("type") in ("integer", "number") and ("multipleOf" in obj):
+            return True
+
+        # Check for array unsupported keywords
+        if obj.get("type") == "array" and any(
+                key in obj
+                for key in ("uniqueItems", "contains", "minContains",
+                            "maxContains", "minItems", "maxItems")):
+            return True
+
+        # Unsupported keywords for strings
+        if obj.get("type") == "string" and "format" in obj:
+            return True
+
+        # Unsupported keywords for objects
+        if obj.get("type") == "object" and any(
+                key in obj for key in ("minProperties", "maxProperties",
+                                       "propertyNames", "patternProperties")):
+            return True
+
+        # Recursively check all nested objects and arrays
+        for value in obj.values():
+            if isinstance(value, dict):
+                if check_object(value):
+                    return True
+            elif isinstance(value, list):
+                for item in value:
+                    if isinstance(item, dict) and check_object(item):
+                        return True
+
+        return False
+
+    return check_object(schema)
+
+
+def validate_xgrammar_grammar(sampling_params: SamplingParams) -> None:
+    """Validate that the request is supported by structured output.
+
+    Raises ValueError if the request is not supported.
+    """
+    if sampling_params.guided_decoding is None:
+        return
+
+    gd_params = sampling_params.guided_decoding
+
+    if gd_params.regex:
+        try:
+            xgr.Grammar.from_regex(gd_params.regex)
+        except Exception as err:
+            raise ValueError("Failed to transform regex into a grammar: "
+                             f"{err}") from err
+
+    if gd_params.choice:
+        choice_grammar = choice_as_grammar(gd_params.choice)
+        try:
+            xgr.Grammar.from_ebnf(choice_grammar)
+        except Exception as err:
+            raise ValueError("Failed to transform choices into a grammar: "
+                             "{err}") from err
+        gd_params.choice = None
+        gd_params.grammar = choice_grammar
+        return
+
+    if gd_params.json:
+        if isinstance(gd_params.json, str):
+            try:
+                schema = json.loads(gd_params.json)
+            except json.JSONDecodeError as e:
+                raise ValueError("Invalid JSON grammar specification.") from e
+        else:
+            schema = gd_params.json
+
+        if has_xgrammar_unsupported_json_features(schema):
+            raise ValueError("The provided JSON schema contains features not "
+                             "supported by xgrammar.")
+        return
+
+    if gd_params.grammar:
+        if grammar_is_likely_lark(gd_params.grammar):
+            # xgrammar supports EBNF grammars only
+            try:
+                gd_params.grammar = convert_lark_to_ebnf(gd_params.grammar)
+            except ValueError as e:
+                raise ValueError(
+                    "Failed to convert the grammar from Lark to EBNF. ") from e
+
+        # Test parsing EBNF grammar, possibly already converted from Lark
+        try:
+            # parse the grammar, but we aren't compiling it.
+            xgr.Grammar.from_ebnf(gd_params.grammar)
+        except Exception as e:
+            raise ValueError("Invalid grammar specification.") from e
+        return
+
+    if gd_params.structural_tag:
+        try:
+            s_tag = json.loads(gd_params.structural_tag)
+            tags = [
+                xgr.StructuralTagItem(
+                    begin=s["begin"],
+                    schema=json.dumps(s["schema"]),
+                    end=s["end"],
+                ) for s in s_tag["structures"]
+            ]
+            xgr.Grammar.from_structural_tag(tags, s_tag["triggers"])
+        except Exception as e:
+            raise ValueError("Invalid structural tag specification.") from e

vllm/v1/structured_output/request.py

@@ -0,0 +1,84 @@
+# SPDX-License-Identifier: Apache-2.0
+from __future__ import annotations
+
+import dataclasses
+import functools
+import json
+from concurrent.futures import Future
+from concurrent.futures._base import TimeoutError
+from typing import Optional, Union, cast
+
+from vllm.sampling_params import SamplingParams
+from vllm.v1.structured_output.backend_types import (StructuredOutputGrammar,
+                                                     StructuredOutputKey,
+                                                     StructuredOutputOptions)
+
+
+@dataclasses.dataclass
+class StructuredOutputRequest:
+
+    sampling_params: SamplingParams
+    _grammar: Optional[Union[Future[StructuredOutputGrammar],
+                             StructuredOutputGrammar]] = None
+
+    def _check_grammar_completion(self) -> bool:
+        # NOTE: We have to lazy import to gate circular imports
+        from vllm.v1.request import RequestStatus
+
+        if isinstance(self._grammar, Future):
+            try:
+                # We will check whether the future is ready within 100 us
+                self._grammar = self._grammar.result(timeout=0.0001)
+                self.status = RequestStatus.WAITING
+            except TimeoutError:
+                return False
+        return True
+
+    @property
+    def is_grammar_ready(self) -> bool:
+        return self._check_grammar_completion()
+
+    @property
+    def grammar(self) -> Optional[StructuredOutputGrammar]:
+        completed = self._check_grammar_completion()
+        return cast(Optional[StructuredOutputGrammar],
+                    self._grammar) if completed else None
+
+    @grammar.setter
+    def grammar(
+        self, grammar: Union[StructuredOutputGrammar,
+                             Future[StructuredOutputGrammar]]
+    ) -> None:
+        self._grammar = grammar
+
+    @functools.cached_property
+    def structured_output_key(self) -> StructuredOutputKey:
+        return get_structured_output_key(self.sampling_params)
+
+
+def get_structured_output_key(
+        sampling_params: SamplingParams) -> StructuredOutputKey:
+    params = sampling_params.guided_decoding
+    assert params is not None, "params can't be None."
+    if params.json is not None:
+        if not isinstance(params.json, str):
+            json_str = json.dumps(params.json)
+        else:
+            json_str = params.json
+        return (StructuredOutputOptions.JSON, json_str)
+    elif params.json_object:
+        return (StructuredOutputOptions.JSON_OBJECT, "")
+    elif params.regex is not None:
+        return (StructuredOutputOptions.REGEX, params.regex)
+    elif params.choice is not None:
+        if not isinstance(params.choice, str):
+            json_str = json.dumps(params.choice)
+        else:
+            json_str = params.choice
+        return (StructuredOutputOptions.CHOICE, json_str)
+    elif params.grammar is not None:
+        return (StructuredOutputOptions.GRAMMAR, params.grammar)
+    elif params.structural_tag is not None:
+        return (StructuredOutputOptions.STRUCTURAL_TAG, params.structural_tag)
+    else:
+        raise ValueError("No valid structured output parameter found")