[V1][Metrics] Add per-request prompt/generation_tokens histograms #12516
Conversation
Observe these values as requests are finished. Requires keeping track of generated tokens per-request to handle streaming delta updates. Signed-off-by: Mark McLoughlin <markmc@redhat.com>
markmc
requested review from
DarkLight1337,
robertgshaw2-redhat,
simon-mo,
WoosukKwon,
njhill,
ywang96,
comaniac and
alexm-redhat
as code owners
|
👋 Hi! Thank you for contributing to the vLLM project. Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging. To run CI, PR reviewers can do one of these:
🚀 |
| @@ -116,9 +131,42 @@ def log(self, scheduler_stats: SchedulerStats, | |||
| self.counter_generation_tokens.inc( | |||
| iteration_stats.num_generation_tokens) | |||
|
|
|||
| for finished_request in iteration_stats.finished_requests: | |||
| self.histogram_num_prompt_tokens_request.observe( | |||
There was a problem hiding this comment.
QQ, in V0, we did:
histogram.labels(**self.labels).observe(datum)Line 541 in 0f657bd
Do you know why this is or is not needed?
There was a problem hiding this comment.
Ill hit automerge to unblock you
There was a problem hiding this comment.
labels() is a pretty expensive factory method and our labels aren't ever changing, so I'm building these labelled metrics in the PrometheusStatLogger constructor
self.histogram_num_generation_tokens_request = \
prometheus_client.Histogram(
name="vllm:request_generation_tokens",
documentation="Number of generation tokens processed.",
buckets=build_1_2_5_buckets(max_model_len),
labelnames=labelnames).labels(*labelvalues)
Not so easy to spot it though!
robertgshaw2-redhat
added
the
ready
…lm-project#12516) Signed-off-by: Mark McLoughlin <markmc@redhat.com>
Follow on from #12478, part of #10582
Observe these values as requests are finished.
Requires keeping track of generated tokens per-request to handle streaming delta updates.