When vLLM runs with Ray Prometheus `vllm:request_success{finished_reason=...}`
only ever increments the repetition bucket regardless of the request's actual finish
reason; stop, length, abort, and error stay at zero. Root cause was `labels()` mutated
the wrapped Ray metric's default tags in place and returned self, so every `.labels(...)`
call on a given wrapper returned the same object.
Co-authored-by: Marwan Sarieddine <sarieddine.marwan@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
Signed-off-by: Marwan Sarieddine <sarieddine.marwan@gmail.com>
Signed-off-by: Seiji Eicher <seiji@anyscale.com>
Enables external `StatLogger` plugins to correlate per-request metrics
with request-level context. Also, this is a pre-requisite for Prometheus
exemplars in #30972.
Signed-off-by: Csrayz <33659823+Csrayz@users.noreply.github.com>
Related to `Counters can only be incremented by non-negative amounts`
error with the `vllm:prompt_tokens_by_source_total` metric.
Signed-off-by: Mark McLoughlin <markmc@redhat.com>
Co-authored-by: Or Ozeri <or@ozery.com>
Add labeled Prometheus metrics to distinguish where prompt tokens come
from in P/D disaggregated deployments.
In P/D disaggregation, decode instances receive KV cache from prefill instances.
Currently, decode reports inflated prompt throughput because it counts all
prompt tokens as "computed", even though most were transferred.
This PR adds labeled metrics so users can understand actual compute work vs
transferred work:
vllm:prompt_tokens_by_source_total{source="local_compute"} # Tokens prefilled locally
vllm:prompt_tokens_by_source_total{source="external_kv_transfer"} # Tokens received via KV transfer
vllm:prompt_tokens_by_source_total{source="local_cache_hit"} # Tokens from local prefix cache
vllm:prompt_tokens_cached_total # Total cached (local + external, -1 when all
Signed-off-by: Zhanqiu Hu <zh338@cornell.edu>
