[Deprecation] Deprecate LM Format Enforcer structured outputs backend (#63423)

Signed-off-by: Woosuk Kwon <woosuk@inferact.ai>
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

docs/features/structured_outputs.md

@@ -17,6 +17,27 @@ available to generate structured outputs.
     - `structural_tag` -> `{"structured_outputs": {"structural_tag": ...}}` or `StructuredOutputsParams(structural_tag=...)`
     - `guided_decoding_backend` -> Remove this field from your request
 
+## Migrating from LM Format Enforcer
+
+!!! warning
+    The `lm-format-enforcer` structured outputs backend is deprecated starting
+    in v0.33.0. It remains usable with a warning in v0.33.0, will require an
+    explicit opt-in in v0.34.0, and will be removed in v0.35.0.
+
+For online serving, replace
+`--structured-outputs-config.backend lm-format-enforcer` with
+`--structured-outputs-config.backend auto`, or remove the option to use the
+default. For offline inference, replace
+`structured_outputs_config={"backend": "lm-format-enforcer"}` with
+`structured_outputs_config={"backend": "auto"}` when constructing `LLM`.
+You can also explicitly select `xgrammar` or `guidance`.
+
+Validate your existing JSON schemas, regex patterns, and generated outputs with
+the replacement backend before migrating. Backend support and formatting can
+differ: in particular, LM Format Enforcer uses Python-style regex syntax, while
+XGrammar and Guidance use Rust-style regex syntax. Changing backends can change
+which tokens are allowed and therefore the generated output.
+
 ## Online Serving (OpenAI API)
 
 You can generate structured outputs using the OpenAI's [Completions](https://platform.openai.com/docs/api-reference/completions) and [Chat](https://platform.openai.com/docs/api-reference/chat) API.

vllm/config/structured_outputs.py

@@ -6,8 +6,11 @@ from typing import Any, Literal, Self
 from pydantic import model_validator
 
 from vllm.config.utils import config
+from vllm.logger import init_logger
 from vllm.utils.hashing import safe_hash
 
+logger = init_logger(__name__)
+
 StructuredOutputsBackend = Literal[
     "auto", "xgrammar", "guidance", "outlines", "lm-format-enforcer"
 ]
@@ -21,7 +24,9 @@ class StructuredOutputsConfig:
     """Which engine will be used for structured outputs (e.g. JSON schema,
     regex, etc) by default. With "auto", we will make opinionated choices
     based on request contents and what the backend libraries currently support,
-    so the behavior is subject to change in each release."""
+    so the behavior is subject to change in each release.
+    The "lm-format-enforcer" backend is deprecated and will be removed in
+    v0.35.0. Use "auto", "xgrammar", or "guidance" instead."""
     disable_any_whitespace: bool = False
     """If `True`, json output will always be compact without any whitespace.
     If `False`, the model may generate whitespace between JSON fields,
@@ -69,4 +74,12 @@ class StructuredOutputsConfig:
                 "disable_additional_properties is only supported "
                 "for the guidance backend."
             )
+        if self.backend == "lm-format-enforcer":
+            logger.warning_once(
+                "The lm-format-enforcer structured outputs backend is deprecated "
+                "and will be removed in v0.35.0. Use auto, xgrammar, or guidance "
+                "instead. See https://docs.vllm.ai/en/latest/features/"
+                "structured_outputs/#migrating-from-lm-format-enforcer "
+                "for migration guidance."
+            )
         return self