Add vllm:prefix_cache_hits and vllm:prefix_cache_queries counters#358
Open
InfraWhisperer wants to merge 2 commits intollm-d:mainfrom
Open
Add vllm:prefix_cache_hits and vllm:prefix_cache_queries counters#358InfraWhisperer wants to merge 2 commits intollm-d:mainfrom
InfraWhisperer wants to merge 2 commits intollm-d:mainfrom
Conversation
|
Unsigned commits detected! Please sign your commits. For instructions on how to set up GPG/SSH signing and verify your commits, please see GitHub Documentation. |
InfraWhisperer
force-pushed
the
prefix-cache-metrics
branch
2 times, most recently
from
06409c6 to
c2cb138
Compare
Expose token-level prefix cache metrics matching vLLM v1 semantics. Both counters increment per-request: queries by total prompt tokens, hits by tokens found already cached. Enables computing cache hit rate via rate(hits) / rate(queries) for scorer strategy benchmarking. Closes llm-d#356 Signed-off-by: InfraWhisperer <raghav.potluri21@gmail.com>
InfraWhisperer
force-pushed
the
prefix-cache-metrics
branch
from
c2cb138 to
bc803b4
Compare
mayabar
reviewed
Feb 25, 2026
Collaborator
mayabar
left a comment
mayabar
left a comment
There was a problem hiding this comment.
@InfraWhisperer thanks for you PR
Looks good, some general comments:
- please add support for fake metrics: check new values validity on initialization, e.g. check that values are non-negative, initialize prometheus values from the fake values given in the configuration, ...
- tests for the new metrics are missing: need to test calculation in both scenarios, real and fake metrics
Add PrefixCacheHits and PrefixCacheQueries fields to the fake metrics config struct with validation (non-negative, must be specified together, hits <= queries). Initialize Prometheus counters from fake values in setInitialPrometheusMetrics. Add integration tests covering real prefix cache metrics (sequential requests with shared prefixes), fake prefix cache metrics (values appear on /metrics), and fake value immutability (real requests don't mutate fake counters). Add config validation tests for all error paths including the partial specification guard. Signed-off-by: Raghav Potluri <raghav.potluri21@gmail.com>
Author
Hi @mayabar I have added tests as per your comments. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
vllm:prefix_cache_hitsandvllm:prefix_cache_queriesPrometheus counters matching vLLM v1 token-level semanticsqueriesincrements by total prompt tokens per request,hitsby cached tokens — enablesrate(hits) / rate(queries)for cache effectiveness measurementkvCacheUsageChan→kvCacheUsageUpdater)Closes #356
Test plan
go build ./...passes--enable-kvcacheand confirm both counters appear on/metricsprefix_cache_hitsincrements on subsequent requests--enable-kvcacheis not setrate(vllm:prefix_cache_hits[5m]) / rate(vllm:prefix_cache_queries[5m])produces expected hit rate in Prometheus