xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-22 20:55:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

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

Browse Source 
Signed-off-by: Russell Bryant <rbryant@redhat.com>
This commit is contained in:
Russell Bryant 2025-06-12 19:30:09 -04:00 committed by GitHub
parent dba68f9159
commit c57bb199b3
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 8 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
vllm/v1/worker/gpu_model_runner.py
View File

@ -66,11 +66,15 @@ from .utils import (gather_mm_placeholders, initialize_kv_cache_for_kv_sharing,
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 @@ class GPUModelRunner(LoRAModelRunnerMixin):
# 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,