Skip to content

[V1] Resolve failed concurrent structred output requests #19565

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
aarnphm merged 2 commits into from
Jun 12, 2025
Merged

[V1] Resolve failed concurrent structred output requests #19565

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
3df1171
[V1] Resolve failed concurrent structred output requests
russellb Jun 12, 2025
5288105
Update vllm/v1/worker/gpu_model_runner.py
russellb Jun 12, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 8 additions & 1 deletion vllm/v1/worker/gpu_model_runner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -65,11 +65,15 @@

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__)

Expand Down Expand Up @@ -1102,7 +1106,10 @@ 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(
Comment thread
russellb marked this conversation as resolved.
logits,
grammar_bitmask.to(self.device, non_blocking=True),
indices=out_indices,
Expand Down