[Bugfix] Only require XGrammar on x86 (#10865)

Signed-off-by: mgoin <michael@neuralmagic.com>

vllm/platforms/__init__.py

@@ -1,5 +1,5 @@
 from .interface import _Backend  # noqa: F401
-from .interface import Platform, PlatformEnum, UnspecifiedPlatform
+from .interface import CpuArchEnum, Platform, PlatformEnum, UnspecifiedPlatform
 
 current_platform: Platform
 
@@ -120,4 +120,4 @@ elif is_openvino:
 else:
     current_platform = UnspecifiedPlatform()
 
-__all__ = ['Platform', 'PlatformEnum', 'current_platform']
+__all__ = ['Platform', 'PlatformEnum', 'current_platform', 'CpuArchEnum']

vllm/platforms/interface.py

@@ -1,4 +1,5 @@
 import enum
+import platform
 import random
 from typing import TYPE_CHECKING, NamedTuple, Optional, Tuple, Union
 
@@ -37,6 +38,14 @@ class PlatformEnum(enum.Enum):
     UNSPECIFIED = enum.auto()
 
 
+class CpuArchEnum(enum.Enum):
+    X86 = enum.auto()
+    ARM = enum.auto()
+    POWERPC = enum.auto()
+    OTHER = enum.auto()
+    UNKNOWN = enum.auto()
+
+
 class DeviceCapability(NamedTuple):
     major: int
     minor: int
@@ -184,6 +193,23 @@ class Platform:
                 f"{quant} quantization is currently not supported in "
                 f"{cls.device_name}.")
 
+    @classmethod
+    def get_cpu_architecture(cls) -> CpuArchEnum:
+        """
+        Determine the CPU architecture of the current system.
+        Returns CpuArchEnum indicating the architecture type.
+        """
+        machine = platform.machine().lower()
+
+        if machine in ("x86_64", "amd64", "i386", "i686"):
+            return CpuArchEnum.X86
+        elif machine.startswith("arm") or machine.startswith("aarch"):
+            return CpuArchEnum.ARM
+        elif machine.startswith("ppc"):
+            return CpuArchEnum.POWERPC
+
+        return CpuArchEnum.OTHER if machine else CpuArchEnum.UNKNOWN
+
 
 class UnspecifiedPlatform(Platform):
     _enum = PlatformEnum.UNSPECIFIED