xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-05-29 23:37:12 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Metrics] Scheduled removal of deprecated metrics (#29330)

Browse Source 
Signed-off-by: Mark McLoughlin <markmc@redhat.com>
This commit is contained in:
Mark McLoughlin 2025-11-25 03:15:13 +00:00 committed by GitHub
parent 7012d8b45e
commit 9cf4edae6e
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 37 additions and 100 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
tests/entrypoints/openai/test_metrics.py
View File

@ -183,9 +183,6 @@ async def test_metrics_counts(
EXPECTED_METRICS_V1 = [ EXPECTED_METRICS_V1 = [
"vllm:num_requests_running", "vllm:num_requests_running",
"vllm:num_requests_waiting", "vllm:num_requests_waiting",
"vllm:gpu_cache_usage_perc",
"vllm:gpu_prefix_cache_queries",
"vllm:gpu_prefix_cache_hits",
"vllm:kv_cache_usage_perc", "vllm:kv_cache_usage_perc",
"vllm:prefix_cache_queries", "vllm:prefix_cache_queries",
"vllm:prefix_cache_hits", "vllm:prefix_cache_hits",

134
vllm/v1/metrics/loggers.py
View File

@ -440,57 +440,6 @@ class PrometheusStatLogger(AggregateStatLoggerBase):
# Setting default values # Setting default values
self.record_sleep_state() self.record_sleep_state()
# GPU cache
#
# Deprecated in 0.9.2 - Renamed as vllm:kv_cache_usage_perc
# With 0.11.x you can enable with --show-hidden-metrics-for-version=0.10
# TODO: remove in 0.12.0
if self.show_hidden_metrics:
gauge_gpu_cache_usage = self._gauge_cls(
name="vllm:gpu_cache_usage_perc",
documentation=(
"GPU KV-cache usage. 1 means 100 percent usage."
"DEPRECATED: Use vllm:kv_cache_usage_perc instead."
),
multiprocess_mode="mostrecent",
labelnames=labelnames,
)
self.gauge_gpu_cache_usage = make_per_engine(
gauge_gpu_cache_usage, engine_indexes, model_name
)
# Deprecated in 0.9.2 - Renamed as vllm:prefix_cache_queries
# With 0.11.x you can enable with --show-hidden-metrics-for-version=0.10
# TODO: remove in 0.12.0
if self.show_hidden_metrics:
counter_gpu_prefix_cache_queries = self._counter_cls(
name="vllm:gpu_prefix_cache_queries",
documentation=(
"GPU prefix cache queries, in terms of number of queried"
"tokens. DEPRECATED: Use vllm:prefix_cache_queries instead."
),
labelnames=labelnames,
)
self.counter_gpu_prefix_cache_queries = make_per_engine(
counter_gpu_prefix_cache_queries, engine_indexes, model_name
)
# Deprecated in 0.9.2 - Renamed as vllm:prefix_cache_hits
# With 0.11.x you can enable with --show-hidden-metrics-for-version=0.10
# TODO: remove in 0.12.0
if self.show_hidden_metrics:
counter_gpu_prefix_cache_hits = self._counter_cls(
name="vllm:gpu_prefix_cache_hits",
documentation=(
"GPU prefix cache hits, in terms of number of cached "
"tokens. DEPRECATED: Use vllm:prefix_cache_hits instead."
),
labelnames=labelnames,
)
self.counter_gpu_prefix_cache_hits = make_per_engine(
counter_gpu_prefix_cache_hits, engine_indexes, model_name
)
gauge_kv_cache_usage = self._gauge_cls( gauge_kv_cache_usage = self._gauge_cls(
name="vllm:kv_cache_usage_perc", name="vllm:kv_cache_usage_perc",
documentation="KV-cache usage. 1 means 100 percent usage.", documentation="KV-cache usage. 1 means 100 percent usage.",
@ -735,39 +684,41 @@ class PrometheusStatLogger(AggregateStatLoggerBase):
) )
# Deprecated in 0.11 - Renamed as vllm:inter_token_latency_seconds # Deprecated in 0.11 - Renamed as vllm:inter_token_latency_seconds
# TODO: in 0.12, only enable if show_hidden_metrics=True # With 0.12.x you can enable with --show-hidden-metrics-for-version=0.11
histogram_time_per_output_token = self._histogram_cls( # TODO: remove in 0.13.0
name="vllm:time_per_output_token_seconds", if self.show_hidden_metrics:
documentation=( histogram_time_per_output_token = self._histogram_cls(
"Histogram of time per output token in seconds." name="vllm:time_per_output_token_seconds",
"DEPRECATED: Use vllm:inter_token_latency_seconds instead." documentation=(
), "Histogram of time per output token in seconds."
buckets=[ "DEPRECATED: Use vllm:inter_token_latency_seconds instead."
0.01, ),
0.025, buckets=[
0.05, 0.01,
0.075, 0.025,
0.1, 0.05,
0.15, 0.075,
0.2, 0.1,
0.3, 0.15,
0.4, 0.2,
0.5, 0.3,
0.75, 0.4,
1.0, 0.5,
2.5, 0.75,
5.0, 1.0,
7.5, 2.5,
10.0, 5.0,
20.0, 7.5,
40.0, 10.0,
80.0, 20.0,
], 40.0,
labelnames=labelnames, 80.0,
) ],
self.histogram_time_per_output_token = make_per_engine( labelnames=labelnames,
histogram_time_per_output_token, engine_indexes, model_name )
) self.histogram_time_per_output_token = make_per_engine(
histogram_time_per_output_token, engine_indexes, model_name
)
histogram_inter_token_latency = self._histogram_cls( histogram_inter_token_latency = self._histogram_cls(
name="vllm:inter_token_latency_seconds", name="vllm:inter_token_latency_seconds",
@ -966,20 +917,8 @@ class PrometheusStatLogger(AggregateStatLoggerBase):
self.gauge_scheduler_waiting[engine_idx].set( self.gauge_scheduler_waiting[engine_idx].set(
scheduler_stats.num_waiting_reqs scheduler_stats.num_waiting_reqs
) )
if self.show_hidden_metrics:
self.gauge_gpu_cache_usage[engine_idx].set(
scheduler_stats.kv_cache_usage
)
self.gauge_kv_cache_usage[engine_idx].set(scheduler_stats.kv_cache_usage) self.gauge_kv_cache_usage[engine_idx].set(scheduler_stats.kv_cache_usage)
if self.show_hidden_metrics:
self.counter_gpu_prefix_cache_queries[engine_idx].inc(
scheduler_stats.prefix_cache_stats.queries
)
self.counter_gpu_prefix_cache_hits[engine_idx].inc(
scheduler_stats.prefix_cache_stats.hits
)
self.counter_prefix_cache_queries[engine_idx].inc( self.counter_prefix_cache_queries[engine_idx].inc(
scheduler_stats.prefix_cache_stats.queries scheduler_stats.prefix_cache_stats.queries
) )
@ -1050,7 +989,8 @@ class PrometheusStatLogger(AggregateStatLoggerBase):
self.histogram_time_to_first_token[engine_idx].observe(ttft) self.histogram_time_to_first_token[engine_idx].observe(ttft)
for itl in iteration_stats.inter_token_latencies_iter: for itl in iteration_stats.inter_token_latencies_iter:
self.histogram_inter_token_latency[engine_idx].observe(itl) self.histogram_inter_token_latency[engine_idx].observe(itl)
self.histogram_time_per_output_token[engine_idx].observe(itl) if self.show_hidden_metrics:
self.histogram_time_per_output_token[engine_idx].observe(itl)
for finished_request in iteration_stats.finished_requests: for finished_request in iteration_stats.finished_requests:
self.counter_request_success[finished_request.finish_reason][ self.counter_request_success[finished_request.finish_reason][