From e9ba99f296e77038f59f59e33579f4011ed56411 Mon Sep 17 00:00:00 2001
From: Shanshan Shen <467638484@qq.com>
Date: Mon, 7 Apr 2025 19:06:24 +0800
Subject: [PATCH] [V1][Structured Output] Add `supports_structured_output()`
 method to Platform (#16148)

Signed-off-by: shen-shanshan <467638484@qq.com>
---
 vllm/platforms/cpu.py       | 4 ++++
 vllm/platforms/cuda.py      | 4 ++++
 vllm/platforms/hpu.py       | 4 ++++
 vllm/platforms/interface.py | 7 +++++++
 vllm/platforms/neuron.py    | 4 ++++
 vllm/platforms/rocm.py      | 4 ++++
 vllm/platforms/tpu.py       | 5 +++++
 vllm/platforms/xpu.py       | 4 ++++
 vllm/v1/engine/processor.py | 8 +++++---
 9 files changed, 41 insertions(+), 3 deletions(-)

diff --git a/vllm/platforms/cpu.py b/vllm/platforms/cpu.py
index 67466bdb98075..cfd7bc2a40571 100644
--- a/vllm/platforms/cpu.py
+++ b/vllm/platforms/cpu.py
@@ -180,3 +180,7 @@ class CpuPlatform(Platform):
         Get device specific communicator class for distributed communication.
         """
         return "vllm.distributed.device_communicators.cpu_communicator.CpuCommunicator"  # noqa
+
+    @classmethod
+    def supports_structured_output(cls) -> bool:
+        return True
diff --git a/vllm/platforms/cuda.py b/vllm/platforms/cuda.py
index 0576022be448b..053cf74ebceb0 100644
--- a/vllm/platforms/cuda.py
+++ b/vllm/platforms/cuda.py
@@ -308,6 +308,10 @@ class CudaPlatformBase(Platform):
     def supports_v1(cls, model_config: ModelConfig) -> bool:
         return True
 
+    @classmethod
+    def supports_structured_output(cls) -> bool:
+        return True
+
     @classmethod
     def use_custom_allreduce(cls) -> bool:
         return True
diff --git a/vllm/platforms/hpu.py b/vllm/platforms/hpu.py
index 4c842b5251105..f011f14029a39 100644
--- a/vllm/platforms/hpu.py
+++ b/vllm/platforms/hpu.py
@@ -92,3 +92,7 @@ class HpuPlatform(Platform):
     @classmethod
     def get_device_communicator_cls(cls) -> str:
         return "vllm.distributed.device_communicators.hpu_communicator.HpuCommunicator"  # noqa
+
+    @classmethod
+    def supports_structured_output(cls) -> bool:
+        return True
diff --git a/vllm/platforms/interface.py b/vllm/platforms/interface.py
index b6f6029de9c82..2bb543bd73f70 100644
--- a/vllm/platforms/interface.py
+++ b/vllm/platforms/interface.py
@@ -379,6 +379,13 @@ class Platform:
         """
         return False
 
+    @classmethod
+    def supports_structured_output(cls) -> bool:
+        """
+        Returns whether the current platform can support structured output.
+        """
+        return False
+
     @classmethod
     def use_custom_allreduce(cls) -> bool:
         """
diff --git a/vllm/platforms/neuron.py b/vllm/platforms/neuron.py
index c1f426e5b8801..93657881cbdd8 100644
--- a/vllm/platforms/neuron.py
+++ b/vllm/platforms/neuron.py
@@ -67,3 +67,7 @@ class NeuronPlatform(Platform):
     @classmethod
     def use_all_gather(cls) -> bool:
         return True
+
+    @classmethod
+    def supports_structured_output(cls) -> bool:
+        return True
diff --git a/vllm/platforms/rocm.py b/vllm/platforms/rocm.py
index d18b7c26f7ec5..a2fbf416ecf20 100644
--- a/vllm/platforms/rocm.py
+++ b/vllm/platforms/rocm.py
@@ -303,6 +303,10 @@ class RocmPlatform(Platform):
         # V1 support on AMD gpus is experimental
         return True
 
+    @classmethod
+    def supports_structured_output(cls) -> bool:
+        return True
+
     @classmethod
     def use_custom_allreduce(cls) -> bool:
         # We only enable custom allreduce for MI300 series
diff --git a/vllm/platforms/tpu.py b/vllm/platforms/tpu.py
index 43d3044cb93ee..eeadb4a71e5e7 100644
--- a/vllm/platforms/tpu.py
+++ b/vllm/platforms/tpu.py
@@ -133,3 +133,8 @@ class TpuPlatform(Platform):
     def supports_v1(cls, model_config: ModelConfig) -> bool:
         # V1 support on TPU is experimental
         return True
+
+    @classmethod
+    def supports_structured_output(cls) -> bool:
+        # Structured output is not supported on TPU.
+        return False
diff --git a/vllm/platforms/xpu.py b/vllm/platforms/xpu.py
index 225e756cd7ce8..c4bd639384a40 100644
--- a/vllm/platforms/xpu.py
+++ b/vllm/platforms/xpu.py
@@ -140,3 +140,7 @@ class XPUPlatform(Platform):
     @classmethod
     def get_device_communicator_cls(cls) -> str:
         return "vllm.distributed.device_communicators.xpu_communicator.XpuCommunicator"  # noqa
+
+    @classmethod
+    def supports_structured_output(cls) -> bool:
+        return True
diff --git a/vllm/v1/engine/processor.py b/vllm/v1/engine/processor.py
index 0d2892837eb28..403edddfcbee6 100644
--- a/vllm/v1/engine/processor.py
+++ b/vllm/v1/engine/processor.py
@@ -136,9 +136,11 @@ class Processor:
                                  f" != {engine_level_backend}")
         else:
             params.guided_decoding.backend = engine_level_backend
-        import vllm.platforms
-        if vllm.platforms.current_platform.is_tpu():
-            raise ValueError("Structured output is not supported on TPU.")
+
+        from vllm.platforms import current_platform
+        if not current_platform.supports_structured_output():
+            raise ValueError("Structured output is not supported on "
+                             f"{current_platform.device_name}.")
 
         # Request content validation
         if engine_level_backend.startswith("xgrammar"):