[Metrics] Expose cached prompt tokens by cache tier (#56318)

Signed-off-by: Cam Quilici <cjquilici@gmail.com>
Signed-off-by: Cam Quilici <cameron@semianalysis.com>
Co-authored-by: Cam Quilici <cameron@semianalysis.com>
Co-authored-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: Yifan Qiao <yifanqiao@inferact.ai>
This commit is contained in:
Cameron Quilici
2026-10-06 14:52:05 -04:00
committed by GitHub
co-authored by Cam Quilici Nick Hill Yifan Qiao
parent ad0f67a7ec
commit e37e51dd24
36 changed files with 1005 additions and 44 deletions
@@ -23,6 +23,7 @@ from vllm.config import (
)
from vllm.config.cache import MambaCacheMode
from vllm.config.kv_events import KVEventsConfig
from vllm.distributed.kv_transfer.kv_connector.v1 import KVConnectorRole
from vllm.distributed.kv_transfer.kv_connector.v1.simple_cpu_offload_connector import (
SimpleCPUOffloadConnector,
)
@@ -53,6 +54,7 @@ from vllm.v1.kv_cache_interface import (
MambaSpec,
SlidingWindowSpec,
)
from vllm.v1.metrics.cache_hit_source import CacheHitSource
from vllm.v1.outputs import KVConnectorOutput
from vllm.v1.request import Request
from vllm.v1.simple_kv_offload.manager import SimpleCPUOffloadScheduler
@@ -213,6 +215,37 @@ class SchedulerFixture:
num_groups: int = 1
@pytest.mark.parametrize(
("backend", "source"),
[("cpu", CacheHitSource.HOST), ("disk", CacheHitSource.DISK)],
)
def test_connector_reports_configured_cache_source(
tmp_path, backend: str, source: CacheHitSource
):
kv_cache_config = _make_kv_cache_config(num_blocks=4)
vllm_config = _make_vllm_config()
extra_config = {
"cpu_bytes_to_use_per_rank": _BYTES_PER_BLOCK * 4,
"kv_offload_backend": backend,
}
if backend == "disk":
extra_config.update(
disk_path=str(tmp_path / "kv-offload.bin"),
disk_capacity_bytes=_BYTES_PER_BLOCK * 4,
)
assert vllm_config.kv_transfer_config is not None
vllm_config.kv_transfer_config.kv_connector_extra_config = extra_config
connector = SimpleCPUOffloadConnector(
vllm_config,
KVConnectorRole.SCHEDULER,
kv_cache_config,
)
assert connector.get_external_cache_hit_sources(None, 32) == {source: 32} # type: ignore[arg-type]
assert connector.get_external_cache_hit_sources(None, 0) == {} # type: ignore[arg-type]
def make_scheduler(
num_cpu_blocks: int = 8,
num_gpu_blocks: int = 16,