Upgrade xgrammar to 0.1.23

requirements/common.txt

@@ -25,7 +25,7 @@ outlines == 0.1.11 ; platform_machine == "s390x"
 # required for outlines backend disk cache
 diskcache == 5.6.3
 lark == 1.2.2
-xgrammar == 0.1.21; platform_machine == "x86_64" or platform_machine == "aarch64" or platform_machine == "arm64"
+xgrammar == 0.1.23; platform_machine == "x86_64" or platform_machine == "aarch64" or platform_machine == "arm64"
 typing_extensions >= 4.10
 filelock >= 3.16.1 # need to contain https://github.com/tox-dev/filelock/pull/317
 partial-json-parser # used for parsing partial JSON outputs

vllm/v1/worker/gpu_model_runner.py

@@ -90,15 +90,11 @@
 
 if TYPE_CHECKING:
     import xgrammar as xgr
-    import xgrammar.kernels.apply_token_bitmask_inplace_torch_compile as xgr_torch_compile  # noqa: E501
 
     from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
     from vllm.v1.core.sched.output import SchedulerOutput
 else:
     xgr = LazyLoader("xgr", globals(), "xgrammar")
-    xgr_torch_compile = LazyLoader(
-        "xgr_torch_compile", globals(),
-        "xgrammar.kernels.apply_token_bitmask_inplace_torch_compile")
 
 logger = init_logger(__name__)
 
@@ -1333,10 +1329,7 @@ def apply_grammar_bitmask(
         # so we receive it in that format.
         grammar_bitmask = torch.from_numpy(grammar_bitmask).contiguous()
 
-        # Force use of the torch.compile implementation from xgrammar to work
-        # around issues with the Triton kernel in concurrent structured output
-        # scenarios. See PR #19565 and issues #19493, #18376 for details.
-        xgr_torch_compile.apply_token_bitmask_inplace_torch_compile(
+        xgr.apply_token_bitmask_inplace(
             logits,
             grammar_bitmask.to(self.device, non_blocking=True),
             indices=out_indices if not skip_out_indices else None,