[Core] Skip JIT monitor when JIT warmup is disabled (#58590)

Signed-off-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: Kimi Code <noreply@moonshot.cn>
This commit is contained in:
Nick Hill
2026-09-24 12:08:16 -07:00
committed by GitHub
co-authored by Kimi Code
parent 074d57fce1
commit e30559b58f
2 changed files with 44 additions and 6 deletions
+15 -6
View File
@@ -966,12 +966,7 @@ class Worker(WorkerBase):
# All warmup is done — start monitoring for unexpected JIT
# compilations that would cause latency spikes during inference.
from vllm.utils.jit_monitor import activate as activate_jit_monitor
activate_jit_monitor(
mode=self.observability_config.jit_monitor_mode,
verbose=self.observability_config.jit_monitor_verbose,
)
self._maybe_activate_jit_monitor()
# Freeze the worker heap so the GC won't scan static objects
# (model weights, KV caches, CUDA graphs) during inference.
@@ -987,6 +982,20 @@ class Worker(WorkerBase):
encoder=self.compilation_config.encoder_compilation_time,
)
def _maybe_activate_jit_monitor(self) -> None:
# When JIT warmup is disabled (e.g. enforce_eager), runtime JIT
# compilation is expected, so monitoring would only produce noise
# (or spurious errors in "error" mode).
if not self.vllm_config.kernel_config.enable_jit_warmup:
return
from vllm.utils.jit_monitor import activate as activate_jit_monitor
activate_jit_monitor(
mode=self.observability_config.jit_monitor_mode,
verbose=self.observability_config.jit_monitor_verbose,
)
def _get_cudagraph_capture_context(self) -> AbstractContextManager[None]:
"""Let the configured profiler observe CUDA graph capture."""
if not self.use_v2_model_runner: