[Bugfix] Reject structured outputs for diffusion decoders with a clear error (#45468)

Signed-off-by: Wayne Chiu <waynehacking8@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>

tests/v1/structured_output/test_validation.py

@@ -0,0 +1,50 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Request-time validation of structured output requests."""
+
+import pytest
+
+from vllm.config import StructuredOutputsConfig
+from vllm.sampling_params import SamplingParams, StructuredOutputsParams
+
+pytestmark = pytest.mark.cpu_test
+
+JSON_SCHEMA = {
+    "type": "object",
+    "properties": {
+        "invoice_id": {"type": "string"},
+        "customer": {"type": "string"},
+    },
+    "required": ["invoice_id", "customer"],
+    "additionalProperties": False,
+}
+
+
+class _StubModelConfig:
+    def __init__(self, is_diffusion: bool):
+        self.is_diffusion = is_diffusion
+
+
+def test_structured_outputs_rejected_for_diffusion_models():
+    """Diffusion LLMs denoise the canvas in parallel, which is incompatible
+    with the token-by-token grammar FSM. The request must fail with a clear
+    validation error instead of an FSM rejection mid-generation (#45436)."""
+    params = SamplingParams(
+        structured_outputs=StructuredOutputsParams(json=JSON_SCHEMA)
+    )
+    with pytest.raises(ValueError, match="not yet supported for diffusion"):
+        params._validate_structured_outputs(
+            _StubModelConfig(is_diffusion=True),
+            StructuredOutputsConfig(),
+            tokenizer=None,
+        )
+
+
+def test_plain_request_allowed_for_diffusion_models():
+    """Requests without structured outputs are unaffected by the guard."""
+    params = SamplingParams()
+    params._validate_structured_outputs(
+        _StubModelConfig(is_diffusion=True),
+        StructuredOutputsConfig(),
+        tokenizer=None,
+    )

vllm/sampling_params.py

@@ -720,7 +720,9 @@ class SamplingParams(
         self._validate_logits_processors(model_config)
         self._validate_allowed_token_ids(tokenizer)
         self._validate_spec_decode(speculative_config)
-        self._validate_structured_outputs(structured_outputs_config, tokenizer)
+        self._validate_structured_outputs(
+            model_config, structured_outputs_config, tokenizer
+        )
 
     def _validate_logprobs(self, model_config: ModelConfig) -> None:
         max_logprobs = model_config.max_logprobs
@@ -853,12 +855,25 @@ class SamplingParams(
 
     def _validate_structured_outputs(
         self,
+        model_config: ModelConfig,
         structured_outputs_config: StructuredOutputsConfig | None,
         tokenizer: TokenizerLike | None,
     ) -> None:
         if structured_outputs_config is None or self.structured_outputs is None:
             return
 
+        if model_config.is_diffusion:
+            # Diffusion LLMs denoise a whole canvas of tokens in parallel
+            # rather than sampling left-to-right, which the grammar FSM
+            # requires. Without this check, requests fail mid-generation
+            # with an FSM rejection (HTTP 500). See issue #45436.
+            raise ValueError(
+                "Structured outputs are not yet supported for diffusion "
+                "language models. Remove the structured output constraint "
+                "(e.g. `response_format`, `structured_outputs`) from the "
+                "request."
+            )
+
         if tokenizer is None:
             raise ValueError(
                 "Structured outputs requires a tokenizer so it can't be used with 'skip_tokenizer_init'"  # noqa: E501