[V1] Add regex structured output support with xgrammar

This is built on top of vllm-project#14563, as the upgraded xgrammar is required for regex
support to work.

Signed-off-by: Russell Bryant <[email protected]>

Author: russellb

tests/v1/entrypoints/llm/test_struct_output_generate.py

@@ -1,6 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 
 import json
+import re
 
 import jsonschema
 import pytest
@@ -219,25 +220,24 @@ def test_guided_regex(monkeypatch, sample_regex, guided_decoding_backend: str):
                                      guided_decoding=GuidedDecodingParams(
                                          regex=sample_regex,
                                          backend=guided_decoding_backend))
-    with pytest.raises(ValueError,
-                       match="Regex guided decoding is not supported."):
-        llm.generate(prompts=[
+    outputs = llm.generate(
+        prompts=[
             f"Give an example IPv4 address with this regex: {sample_regex}"
         ] * 2,
-                     sampling_params=sampling_params,
-                     use_tqdm=True)
+        sampling_params=sampling_params,
+        use_tqdm=True,
+    )
 
-    # Once regex is supported --
-    #assert outputs is not None
-    #for output in outputs:
-    #    assert output is not None
-    #    assert isinstance(output, RequestOutput)
-    #    prompt = output.prompt
-    #    generated_text = output.outputs[0].text
-    #    print(generated_text)
-    #    assert generated_text is not None
-    #    assert re.fullmatch(sample_regex, generated_text) is not None
-    #    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
+    assert outputs is not None
+    for output in outputs:
+        assert output is not None
+        assert isinstance(output, RequestOutput)
+        prompt = output.prompt
+        generated_text = output.outputs[0].text
+        print(generated_text)
+        assert generated_text is not None
+        assert re.fullmatch(sample_regex, generated_text) is not None
+        print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
 
 
 @pytest.mark.skip_global_cleanup

vllm/v1/structured_output/__init__.py

@@ -112,6 +112,8 @@ def initialize_grammar(self, key: StructuredOutputKey) -> Grammar:
             ctx = self.compiler.compile_builtin_json_grammar()
         elif request_type == StructuredOutputOptions.GRAMMAR:
             ctx = self.compiler.compile_grammar(grammar_spec)
+        elif request_type == StructuredOutputOptions.REGEX:
+            ctx = self.compiler.compile_regex(grammar_spec)
         else:
             logger.error("Validation should have already occurred. "
                          "Please file an issue.")

vllm/v1/structured_output/utils.py

@@ -251,7 +251,11 @@ def validate_structured_output_request(
     gd_params = sampling_params.guided_decoding
 
     if gd_params.regex:
-        raise ValueError("Regex structured output is not supported.")
+        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)