[V1] Resolve failed concurrent structured output requests (vllm-project#19565)

Signed-off-by: Russell Bryant <rbryant@redhat.com>
Signed-off-by: minpeter <kali2005611@gmail.com>

Author: russellb

vllm/v1/worker/gpu_model_runner.py

@@ -66,11 +66,15 @@
 
 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__)
 
@@ -1103,7 +1107,10 @@ def apply_grammar_bitmask(
         # so we receive it in that format.
         grammar_bitmask = torch.from_numpy(grammar_bitmask)
 
-        xgr.apply_token_bitmask_inplace(
+        # 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(
             logits,
             grammar_bitmask.to(self.device, non_blocking=True),
             indices=out_indices,