[Revert] Revert "[Log] Wire stat loggers into AsyncOmniEngine to match AsyncLL… (#2716)

Signed-off-by: amy-why-3459 <wuhaiyan17@huawei.com>
This commit is contained in:
amy-why-3459 authored and GitHub committed 2026-04-13 14:55:16 +08:00
1 parent 22261430b4
commit 2b70e89535
6 files changed
+8 -144

No files matched your search

+1 -25
View File
@@ -22,8 +22,6 @@ from vllm.outputs import RequestOutput
from vllm.pooling_params import PoolingParams
from vllm.sampling_params import SamplingParams
from vllm.v1.engine import EngineCoreOutputs
from vllm.v1.metrics.loggers import StatLoggerManager
from vllm.v1.metrics.stats import IterationStats
from vllm_omni.distributed.omni_connectors.adapter import compute_talker_prompt_ids_length
from vllm_omni.engine import (
@@ -124,7 +122,6 @@ class Orchestrator:
stage_vllm_configs: list[Any],
*,
async_chunk: bool = False,
logger_manager: StatLoggerManager | None = None,
) -> None:
self.request_async_queue = request_async_queue
self.output_async_queue = output_async_queue
@@ -136,8 +133,6 @@ class Orchestrator:
self.stage_clients: list[Any] = stage_clients
self.output_processors: list[Any] = output_processors
self.stage_vllm_configs: list[Any] = stage_vllm_configs
self.logger_manager: StatLoggerManager | None = logger_manager
self.log_stats = self.logger_manager is not None
# Per-request state
self.request_states: dict[str, OrchestratorRequestState] = {}
@@ -629,13 +624,10 @@ class Orchestrator:
"""
processor = self.output_processors[stage_id]
num_outputs = len(raw_outputs.outputs)
iteration_stats = IterationStats() if (self.log_stats and num_outputs) else None
processed = processor.process_outputs(
raw_outputs.outputs,
raw_outputs.timestamp,
iteration_stats,
None,
)
if processed.reqs_to_abort:
@@ -644,22 +636,6 @@ class Orchestrator:
if raw_outputs.scheduler_stats is not None:
processor.update_scheduler_stats(raw_outputs.scheduler_stats)
# Mirror vLLM AsyncLLM output_handler: feed stats to the logger
# manager so LoggingStatLogger can periodically print KV cache /
# prefix cache hit rate, and PrometheusStatLogger can publish.
if self.logger_manager is not None:
try:
self.logger_manager.record(
engine_idx=stage_id,
scheduler_stats=raw_outputs.scheduler_stats,
iteration_stats=iteration_stats,
)
except Exception:
logger.exception(
"[Orchestrator] stat logger record failed for stage-%s",
stage_id,
)
return processed.request_outputs
async def _handle_add_request(self, msg: dict[str, Any]) -> None: