[Perf][VoxCPM2] Add unified decode graph optimization (#4606)

Signed-off-by: Sy03 <1370724210@qq.com>
Co-authored-by: Yueqian Lin <70319226+linyueqian@users.noreply.github.com>
This commit is contained in:
Sy03andYueqian Lin authored and GitHub committed 2026-07-04 05:30:48 +00:00
1 parent 51c14eeaed
commit 9eaca72382
15 files changed
+1542 -169

No files matched your search

+14 -1
View File
@@ -1264,7 +1264,7 @@ if "daily-omni" not in OPENAI_COMPATIBLE_BACKENDS:
# Prevent import order from causing patch failures
from vllm.benchmarks import serve
from vllm.benchmarks.lib.ready_checker import wait_for_endpoint
from vllm.benchmarks.serve import TaskType, _merge_overrides, calculate_metrics_for_embeddings, get_request
from vllm.benchmarks.serve import TaskType, calculate_metrics_for_embeddings, get_request
from vllm_omni.benchmarks.metrics.metrics import (
MultiModalsBenchmarkMetrics,
@@ -1276,6 +1276,19 @@ from vllm_omni.benchmarks.metrics.metrics import (
benchmark_old = serve.benchmark
def _merge_overrides(base: dict | None, overrides: dict | None) -> dict | None:
"""Merge benchmark extra_body with per-request overrides.
vLLM 0.24 removed the private helper from ``vllm.benchmarks.serve``.
Keep the same shallow-merge behavior here, with request overrides winning.
"""
if not base and not overrides:
return None
merged = dict(base or {})
merged.update(overrides or {})
return merged
async def benchmark(
task_type: TaskType,
endpoint_type: str,