fix: /metrics endpoint always reports engine_type="unified" in PD disaggregation mode (#18552)
Co-authored-by: joo_yeon.lee <joo_yeon.lee@samsung.com>
This commit is contained in:
@@ -325,7 +325,7 @@ class SGLangSchedulerServicer(sglang_scheduler_pb2_grpc.SglangSchedulerServicer)
|
|||||||
token_ids_logprob=None,
|
token_ids_logprob=None,
|
||||||
)
|
)
|
||||||
# Set disaggregation params if needed
|
# Set disaggregation params if needed
|
||||||
if self.server_args.disaggregation_mode != DisaggregationMode.NULL:
|
if self.server_args.disaggregation_mode != DisaggregationMode.NULL.value:
|
||||||
health_req.bootstrap_host = FAKE_BOOTSTRAP_HOST
|
health_req.bootstrap_host = FAKE_BOOTSTRAP_HOST
|
||||||
health_req.bootstrap_room = 0
|
health_req.bootstrap_room = 0
|
||||||
else:
|
else:
|
||||||
@@ -1087,7 +1087,7 @@ def _execute_grpc_server_warmup(server_args: ServerArgs):
|
|||||||
}
|
}
|
||||||
|
|
||||||
# Set disaggregation params if needed
|
# Set disaggregation params if needed
|
||||||
if server_args.disaggregation_mode != DisaggregationMode.NULL:
|
if server_args.disaggregation_mode != DisaggregationMode.NULL.value:
|
||||||
warmup_request_kwargs["disaggregated_params"] = (
|
warmup_request_kwargs["disaggregated_params"] = (
|
||||||
sglang_scheduler_pb2.DisaggregatedParams(
|
sglang_scheduler_pb2.DisaggregatedParams(
|
||||||
bootstrap_host=FAKE_BOOTSTRAP_HOST,
|
bootstrap_host=FAKE_BOOTSTRAP_HOST,
|
||||||
|
|||||||
@@ -489,7 +489,7 @@ async def health_generate(request: Request) -> Response:
|
|||||||
)
|
)
|
||||||
if (
|
if (
|
||||||
_global_state.tokenizer_manager.server_args.disaggregation_mode
|
_global_state.tokenizer_manager.server_args.disaggregation_mode
|
||||||
!= DisaggregationMode.NULL
|
!= DisaggregationMode.NULL.value
|
||||||
):
|
):
|
||||||
gri.bootstrap_host = FAKE_BOOTSTRAP_HOST
|
gri.bootstrap_host = FAKE_BOOTSTRAP_HOST
|
||||||
gri.bootstrap_room = 0
|
gri.bootstrap_room = 0
|
||||||
|
|||||||
@@ -104,9 +104,11 @@ class SchedulerMetricsMixin:
|
|||||||
)
|
)
|
||||||
|
|
||||||
if self.enable_metrics:
|
if self.enable_metrics:
|
||||||
if self.server_args.disaggregation_mode == DisaggregationMode.PREFILL:
|
if self.server_args.disaggregation_mode == DisaggregationMode.PREFILL.value:
|
||||||
engine_type = "prefill"
|
engine_type = "prefill"
|
||||||
elif self.server_args.disaggregation_mode == DisaggregationMode.DECODE:
|
elif (
|
||||||
|
self.server_args.disaggregation_mode == DisaggregationMode.DECODE.value
|
||||||
|
):
|
||||||
engine_type = "decode"
|
engine_type = "decode"
|
||||||
else:
|
else:
|
||||||
engine_type = "unified"
|
engine_type = "unified"
|
||||||
|
|||||||
Reference in New Issue
Block a user