[BugFix] Lazily import XgrammarBackend to avoid early cuda init

Importing xgrammar appears to initialize the cuda context, which we don't want to do in the front-end process. It also means that the server can't be started with the (default) multiproc context mode of fork.

I guess this is what LazyLoader is meant to help with, but it doesn't seem to be working as intended since vllm-project#14694 was merged.

Signed-off-by: Nick Hill <[email protected]>

Author: njhill

vllm/v1/structured_output/__init__.py

@@ -9,7 +9,6 @@
 from vllm.logger import init_logger
 from vllm.v1.structured_output.backend_types import (StructuredOutputBackend,
                                                      StructuredOutputGrammar)
-from vllm.v1.structured_output.backend_xgrammar import XgrammarBackend
 
 if TYPE_CHECKING:
     import numpy as np
@@ -47,6 +46,9 @@ def grammar_init(self, request: Request) -> None:
         if self.backend is None:
             backend_name = request.sampling_params.guided_decoding.backend_name
             if backend_name == "xgrammar":
+                from vllm.v1.structured_output.backend_xgrammar import (
+                    XgrammarBackend)
+
                 self.backend = XgrammarBackend(self.vllm_config)
             else:
                 raise ValueError(