[Metrics] Add group-aware KV cache capacity to vllm:cache_config_info (#42206)

The startup log already reports the correct group-aware KV cache capacity for
hybrid models, but Prometheus did not expose matching info in 'vllm:cache_config_info`.

This PR adds kv_cache_size_tokens and kv_cache_max_concurrency.

Signed-off-by: Ethan Feng <ethan.fengch@gmail.com>
This commit is contained in:
Ethan Feng
2026-06-12 11:49:44 +00:00
committed by GitHub
parent 8af550b399
commit b7f9b6ab27
7 changed files with 75 additions and 26 deletions
@@ -289,6 +289,17 @@ async def test_metrics_exist(
continue
assert metric in response.text
cache_config_samples = [
sample
for family in text_string_to_metric_families(response.text)
if family.name == "vllm:cache_config_info"
for sample in family.samples
]
assert cache_config_samples
for sample in cache_config_samples:
assert sample.labels.get("kv_cache_size_tokens") not in (None, "None", "")
assert sample.labels.get("kv_cache_max_concurrency") not in (None, "None", "")
@pytest.mark.asyncio
async def test_abort_metrics_reset(
+6
View File
@@ -28,6 +28,7 @@ from vllm.v1.core.kv_cache_utils import (
estimate_max_model_len,
generate_block_hash_extra_keys,
generate_scheduler_kv_cache_config,
get_kv_cache_capacity,
get_kv_cache_configs,
get_max_concurrency_for_kv_cache_config,
get_request_block_hasher,
@@ -1459,6 +1460,11 @@ def test_get_max_concurrency_for_kv_cache_config():
vllm_config, kv_cache_config_hybrid_model
)
assert max_concurrency_hybrid_model == 3
num_tokens, max_concurrency = get_kv_cache_capacity(
vllm_config, kv_cache_config_hybrid_model
)
assert num_tokens == max_concurrency_hybrid_model * max_model_len
assert max_concurrency == max_concurrency_hybrid_model
def test_allocate_with_lookahead():