A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 13, 07:00 UTC
[BugFix] Ensure `EngineArgs.create_engine_config` is idempotent (#28515)
Signed-off-by: Nick Hill <nhill@redhat.com>
Nick Hill committed
Nov 13, 2025 at 09:14 UTC
327c0a9a23f2939923d02fbf882640753bf1e030
1 file changed
+9
-10
vllm/engine/arg_utils.py
+9
-10
@@ -1631,40 +1631,39 @@ class EngineArgs:
1631
)
1632
1633
observability_config = ObservabilityConfig(
1634
- show_hidden_metrics_for_version=(self.show_hidden_metrics_for_version),
1634
+ show_hidden_metrics_for_version=self.show_hidden_metrics_for_version,
1635
otlp_traces_endpoint=self.otlp_traces_endpoint,
1636
collect_detailed_traces=self.collect_detailed_traces,
1637
)
1638
1639
# Compilation config overrides
1640
+ compilation_config = copy.deepcopy(self.compilation_config)
1641
if self.cuda_graph_sizes is not None:
1642
logger.warning(
1643
"--cuda-graph-sizes is deprecated and will be removed in v0.13.0 or "
1644
"v1.0.0, whichever is soonest. Please use --cudagraph-capture-sizes "
1645
"instead."
1646
)
1646
- if self.compilation_config.cudagraph_capture_sizes is not None:
1647
+ if compilation_config.cudagraph_capture_sizes is not None:
1648
raise ValueError(
1649
"cuda_graph_sizes and compilation_config."
1650
"cudagraph_capture_sizes are mutually exclusive"
1651
)
1651
- self.compilation_config.cudagraph_capture_sizes = self.cuda_graph_sizes
1652
+ compilation_config.cudagraph_capture_sizes = self.cuda_graph_sizes
1653
if self.cudagraph_capture_sizes is not None:
1653
- if self.compilation_config.cudagraph_capture_sizes is not None:
1654
+ if compilation_config.cudagraph_capture_sizes is not None:
1655
raise ValueError(
1656
"cudagraph_capture_sizes and compilation_config."
1657
"cudagraph_capture_sizes are mutually exclusive"
1658
)
1658
- self.compilation_config.cudagraph_capture_sizes = (
1659
- self.cudagraph_capture_sizes
1660
- )
1659
+ compilation_config.cudagraph_capture_sizes = self.cudagraph_capture_sizes
1660
if self.max_cudagraph_capture_size is not None:
1662
- if self.compilation_config.max_cudagraph_capture_size is not None:
1661
+ if compilation_config.max_cudagraph_capture_size is not None:
1662
raise ValueError(
1663
"max_cudagraph_capture_size and compilation_config."
1664
"max_cudagraph_capture_size are mutually exclusive"
1665
)
1667
- self.compilation_config.max_cudagraph_capture_size = (
1666
+ compilation_config.max_cudagraph_capture_size = (
1667
self.max_cudagraph_capture_size
1668
)
1669
@@ -1679,7 +1678,7 @@ class EngineArgs:
1678
load_config=load_config,
1679
structured_outputs_config=self.structured_outputs_config,
1680
observability_config=observability_config,
1682
- compilation_config=self.compilation_config,
1681
+ compilation_config=compilation_config,
1682
kv_transfer_config=self.kv_transfer_config,
1683
kv_events_config=self.kv_events_config,
1684
ec_transfer_config=self.ec_transfer_config,