[Rust Frontend] Cache metric handles for scheduler & request stats (#47444)
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com> Signed-off-by: Bugen Zhao <i@bugenzhao.com>
This commit is contained in:
co-authored by
mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent
8b79971bb9
commit
8f4c69b222
@@ -512,6 +512,7 @@ impl EngineCoreClient {
|
||||
|
||||
Ok(EngineCoreOutputStream::new(
|
||||
request_id,
|
||||
engine_id.engine_index().unwrap_or(0),
|
||||
self.abort_tx.clone(),
|
||||
rx,
|
||||
))
|
||||
|
||||
@@ -15,7 +15,7 @@ use crate::client::state::{OutputReceiver, RequestRegistry, UtilityReceiver, Uti
|
||||
use crate::client::stream::EngineCoreStreamOutput;
|
||||
use crate::client::{AbortCause, AbortRequest};
|
||||
use crate::error::{client_closed, dispatcher_closed, unexpected_dispatcher_output};
|
||||
use crate::metrics::{LoraInfoExporter, record_scheduler_stats};
|
||||
use crate::metrics::{LoraInfoExporter, SchedulerStatsRecorder};
|
||||
use crate::protocol::encode_msgpack;
|
||||
use crate::protocol::output::{EngineCoreOutput, EngineCoreOutputs};
|
||||
use crate::protocol::request::EngineCoreRequestType;
|
||||
@@ -29,6 +29,7 @@ pub(crate) struct ClientInner {
|
||||
/// The runtime handle used for sending messages to the engine.
|
||||
handle: Handle,
|
||||
model_name: String,
|
||||
scheduler_stats_recorder: SchedulerStatsRecorder,
|
||||
request_reg: Mutex<RequestRegistry>,
|
||||
utility_reg: Mutex<UtilityRegistry>,
|
||||
health_error: ArcSwapOption<Error>,
|
||||
@@ -43,10 +44,13 @@ impl ClientInner {
|
||||
model_name: String,
|
||||
engines: &[ConnectedEngine],
|
||||
) -> Self {
|
||||
let scheduler_stats_recorder =
|
||||
SchedulerStatsRecorder::new(&METRICS.scheduler, &model_name, engines);
|
||||
Self {
|
||||
input_send,
|
||||
handle,
|
||||
model_name,
|
||||
scheduler_stats_recorder,
|
||||
request_reg: Mutex::new(RequestRegistry::new(engines)),
|
||||
utility_reg: Mutex::new(UtilityRegistry::default()),
|
||||
health_error: ArcSwapOption::empty(),
|
||||
@@ -389,12 +393,7 @@ pub(crate) async fn run_output_dispatcher_loop(
|
||||
"dropping scheduler stats for unknown engine"
|
||||
);
|
||||
}
|
||||
record_scheduler_stats(
|
||||
&METRICS.scheduler,
|
||||
inner.model_name(),
|
||||
batch.engine_index,
|
||||
scheduler_stats,
|
||||
);
|
||||
inner.scheduler_stats_recorder.record(batch.engine_index, scheduler_stats);
|
||||
}
|
||||
|
||||
// The engine's scheduler stats never carry adapter names;
|
||||
|
||||
@@ -45,6 +45,7 @@ impl Deref for EngineCoreStreamOutput {
|
||||
/// `finish_reason` is non-`None`.
|
||||
pub struct EngineCoreOutputStream {
|
||||
request_id: String,
|
||||
engine_index: u32,
|
||||
abort_tx: mpsc::UnboundedSender<AbortRequest>,
|
||||
state: State,
|
||||
rx: OutputReceiver,
|
||||
@@ -53,11 +54,13 @@ pub struct EngineCoreOutputStream {
|
||||
impl EngineCoreOutputStream {
|
||||
pub(crate) fn new(
|
||||
request_id: String,
|
||||
engine_index: u32,
|
||||
abort_tx: mpsc::UnboundedSender<AbortRequest>,
|
||||
rx: OutputReceiver,
|
||||
) -> Self {
|
||||
Self {
|
||||
request_id,
|
||||
engine_index,
|
||||
abort_tx,
|
||||
state: State::Running,
|
||||
rx,
|
||||
@@ -68,6 +71,11 @@ impl EngineCoreOutputStream {
|
||||
pub fn request_id(&self) -> &str {
|
||||
&self.request_id
|
||||
}
|
||||
|
||||
/// Return the index of the engine that owns this request.
|
||||
pub fn engine_index(&self) -> u32 {
|
||||
self.engine_index
|
||||
}
|
||||
}
|
||||
|
||||
impl Stream for EngineCoreOutputStream {
|
||||
|
||||
@@ -1,90 +1,190 @@
|
||||
use std::collections::BTreeMap;
|
||||
use std::collections::BTreeSet;
|
||||
use std::time::{SystemTime, UNIX_EPOCH};
|
||||
|
||||
use vllm_metrics::{
|
||||
EngineLabels, EnginePositionLabels, LoraAdapterNames, LoraInfoLabels, SchedulerMetrics,
|
||||
EngineLabels, EnginePositionLabels, F64Gauge, Family, HistogramMetric, LoraAdapterNames,
|
||||
LoraInfoLabels, SchedulerLogStatsAccumulator, SchedulerMetrics, U64Counter, U64Gauge,
|
||||
WaitingReasonLabels,
|
||||
};
|
||||
|
||||
use crate::protocol::stats::SchedulerStats;
|
||||
use crate::transport::ConnectedEngine;
|
||||
|
||||
const WAITING_REASON_CAPACITY: &str = "capacity";
|
||||
const WAITING_REASON_DEFERRED: &str = "deferred";
|
||||
|
||||
/// Record the scheduler-stats-backed metrics for one engine at one point in
|
||||
/// time.
|
||||
pub(crate) fn record_scheduler_stats(
|
||||
metrics: &SchedulerMetrics,
|
||||
model_name: impl Into<String>,
|
||||
engine: u32,
|
||||
stats: &SchedulerStats,
|
||||
) {
|
||||
let model_name = model_name.into();
|
||||
let labels = EngineLabels {
|
||||
model_name: model_name.clone(),
|
||||
engine,
|
||||
};
|
||||
/// Cached scheduler-stats metric handles for all engines connected to one
|
||||
/// frontend client.
|
||||
pub(crate) struct SchedulerStatsRecorder {
|
||||
engines: BTreeMap<u32, SchedulerStatsHandles>,
|
||||
}
|
||||
|
||||
/// Per-engine cached metric handles used while recording `SchedulerStats`.
|
||||
struct SchedulerStatsHandles {
|
||||
// Base labels reused for dynamic child labels.
|
||||
labels: EngineLabels,
|
||||
|
||||
// Scheduler state gauges.
|
||||
metrics.scheduler_running.get_or_create(&labels).set(stats.num_running_reqs);
|
||||
metrics
|
||||
.scheduler_waiting
|
||||
.get_or_create(&labels)
|
||||
.set(stats.num_waiting_reqs + stats.num_skipped_waiting_reqs);
|
||||
metrics
|
||||
.scheduler_waiting_by_reason
|
||||
.get_or_create(&WaitingReasonLabels {
|
||||
model_name: model_name.clone(),
|
||||
engine,
|
||||
reason: WAITING_REASON_CAPACITY,
|
||||
})
|
||||
.set(stats.num_waiting_reqs);
|
||||
metrics
|
||||
.scheduler_waiting_by_reason
|
||||
.get_or_create(&WaitingReasonLabels {
|
||||
model_name: model_name.clone(),
|
||||
engine,
|
||||
reason: WAITING_REASON_DEFERRED,
|
||||
})
|
||||
.set(stats.num_skipped_waiting_reqs);
|
||||
metrics.kv_cache_usage.get_or_create(&labels).set(stats.kv_cache_usage);
|
||||
scheduler_running: U64Gauge,
|
||||
scheduler_waiting: U64Gauge,
|
||||
scheduler_waiting_capacity: U64Gauge,
|
||||
scheduler_waiting_deferred: U64Gauge,
|
||||
kv_cache_usage: F64Gauge,
|
||||
|
||||
// Prefix-cache counters, including the connector-backed external cache path.
|
||||
metrics
|
||||
.prefix_cache_queries
|
||||
.get_or_create(&labels)
|
||||
.inc_by(stats.prefix_cache_stats.base.queries);
|
||||
metrics
|
||||
.prefix_cache_hits
|
||||
.get_or_create(&labels)
|
||||
.inc_by(stats.prefix_cache_stats.base.hits);
|
||||
prefix_cache_queries: U64Counter,
|
||||
prefix_cache_hits: U64Counter,
|
||||
external_prefix_cache_queries: U64Counter,
|
||||
external_prefix_cache_hits: U64Counter,
|
||||
|
||||
// Speculative decoding counters.
|
||||
spec_decode_num_drafts: U64Counter,
|
||||
spec_decode_num_draft_tokens: U64Counter,
|
||||
spec_decode_num_accepted_tokens: U64Counter,
|
||||
spec_decode_num_accepted_tokens_per_pos: Family<EnginePositionLabels, U64Counter>,
|
||||
|
||||
// Per-engine performance / MFU counters.
|
||||
estimated_flops_per_gpu: U64Counter,
|
||||
estimated_read_bytes_per_gpu: U64Counter,
|
||||
estimated_write_bytes_per_gpu: U64Counter,
|
||||
|
||||
// Sampled KV-cache residency histograms.
|
||||
kv_block_lifetime_seconds: HistogramMetric,
|
||||
kv_block_idle_before_evict_seconds: HistogramMetric,
|
||||
kv_block_reuse_gap_seconds: HistogramMetric,
|
||||
|
||||
// Non-Prometheus interval accumulator for periodic text-log helpers.
|
||||
log_stats: SchedulerLogStatsAccumulator,
|
||||
}
|
||||
|
||||
impl SchedulerStatsRecorder {
|
||||
/// Resolve the fixed-label metric handles for the connected engines.
|
||||
pub(crate) fn new(
|
||||
metrics: &SchedulerMetrics,
|
||||
model_name: &str,
|
||||
engines: &[ConnectedEngine],
|
||||
) -> Self {
|
||||
let engines = engines
|
||||
.iter()
|
||||
.filter_map(|engine| {
|
||||
let engine = engine.engine_id.engine_index()?;
|
||||
Some((
|
||||
engine,
|
||||
resolve_scheduler_stats_handles(metrics, model_name, engine),
|
||||
))
|
||||
})
|
||||
.collect();
|
||||
|
||||
Self { engines }
|
||||
}
|
||||
|
||||
/// Record one scheduler-stats payload for the given engine index.
|
||||
pub(crate) fn record(&self, engine_index: u32, stats: &SchedulerStats) {
|
||||
if let Some(handles) = self.engines.get(&engine_index) {
|
||||
record_scheduler_stats_with_handles(handles, stats);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
/// Resolve all fixed-label scheduler metrics for one engine.
|
||||
fn resolve_scheduler_stats_handles(
|
||||
metrics: &SchedulerMetrics,
|
||||
model_name: &str,
|
||||
engine: u32,
|
||||
) -> SchedulerStatsHandles {
|
||||
let labels = EngineLabels {
|
||||
model_name: model_name.to_string(),
|
||||
engine,
|
||||
};
|
||||
let capacity = WaitingReasonLabels {
|
||||
model_name: model_name.to_string(),
|
||||
engine,
|
||||
reason: WAITING_REASON_CAPACITY,
|
||||
};
|
||||
let deferred = WaitingReasonLabels {
|
||||
model_name: model_name.to_string(),
|
||||
engine,
|
||||
reason: WAITING_REASON_DEFERRED,
|
||||
};
|
||||
|
||||
SchedulerStatsHandles {
|
||||
scheduler_running: metrics.scheduler_running.get_or_create_owned(&labels),
|
||||
scheduler_waiting: metrics.scheduler_waiting.get_or_create_owned(&labels),
|
||||
scheduler_waiting_capacity: metrics
|
||||
.scheduler_waiting_by_reason
|
||||
.get_or_create_owned(&capacity),
|
||||
scheduler_waiting_deferred: metrics
|
||||
.scheduler_waiting_by_reason
|
||||
.get_or_create_owned(&deferred),
|
||||
kv_cache_usage: metrics.kv_cache_usage.get_or_create_owned(&labels),
|
||||
prefix_cache_queries: metrics.prefix_cache_queries.get_or_create_owned(&labels),
|
||||
prefix_cache_hits: metrics.prefix_cache_hits.get_or_create_owned(&labels),
|
||||
external_prefix_cache_queries: metrics
|
||||
.external_prefix_cache_queries
|
||||
.get_or_create_owned(&labels),
|
||||
external_prefix_cache_hits: metrics.external_prefix_cache_hits.get_or_create_owned(&labels),
|
||||
spec_decode_num_drafts: metrics.spec_decode_num_drafts.get_or_create_owned(&labels),
|
||||
spec_decode_num_draft_tokens: metrics
|
||||
.spec_decode_num_draft_tokens
|
||||
.get_or_create_owned(&labels),
|
||||
spec_decode_num_accepted_tokens: metrics
|
||||
.spec_decode_num_accepted_tokens
|
||||
.get_or_create_owned(&labels),
|
||||
spec_decode_num_accepted_tokens_per_pos: metrics
|
||||
.spec_decode_num_accepted_tokens_per_pos
|
||||
.clone(),
|
||||
log_stats: metrics.log_stats.get_or_create_owned(&labels),
|
||||
estimated_flops_per_gpu: metrics.estimated_flops_per_gpu.get_or_create_owned(&labels),
|
||||
estimated_read_bytes_per_gpu: metrics
|
||||
.estimated_read_bytes_per_gpu
|
||||
.get_or_create_owned(&labels),
|
||||
estimated_write_bytes_per_gpu: metrics
|
||||
.estimated_write_bytes_per_gpu
|
||||
.get_or_create_owned(&labels),
|
||||
kv_block_lifetime_seconds: metrics.kv_block_lifetime_seconds.get_or_create_owned(&labels),
|
||||
kv_block_idle_before_evict_seconds: metrics
|
||||
.kv_block_idle_before_evict_seconds
|
||||
.get_or_create_owned(&labels),
|
||||
kv_block_reuse_gap_seconds: metrics.kv_block_reuse_gap_seconds.get_or_create_owned(&labels),
|
||||
labels,
|
||||
}
|
||||
}
|
||||
|
||||
/// Record scheduler-stats values through pre-resolved metric handles.
|
||||
fn record_scheduler_stats_with_handles(handles: &SchedulerStatsHandles, stats: &SchedulerStats) {
|
||||
// Scheduler state gauges.
|
||||
handles.scheduler_running.set(stats.num_running_reqs);
|
||||
handles
|
||||
.scheduler_waiting
|
||||
.set(stats.num_waiting_reqs + stats.num_skipped_waiting_reqs);
|
||||
handles.scheduler_waiting_capacity.set(stats.num_waiting_reqs);
|
||||
handles.scheduler_waiting_deferred.set(stats.num_skipped_waiting_reqs);
|
||||
handles.kv_cache_usage.set(stats.kv_cache_usage);
|
||||
|
||||
// Prefix-cache counters, including the connector-backed external cache path.
|
||||
handles.prefix_cache_queries.inc_by(stats.prefix_cache_stats.base.queries);
|
||||
handles.prefix_cache_hits.inc_by(stats.prefix_cache_stats.base.hits);
|
||||
|
||||
if let Some(connector_prefix_cache_stats) = &stats.connector_prefix_cache_stats {
|
||||
metrics
|
||||
handles
|
||||
.external_prefix_cache_queries
|
||||
.get_or_create(&labels)
|
||||
.inc_by(connector_prefix_cache_stats.base.queries);
|
||||
metrics
|
||||
handles
|
||||
.external_prefix_cache_hits
|
||||
.get_or_create(&labels)
|
||||
.inc_by(connector_prefix_cache_stats.base.hits);
|
||||
}
|
||||
|
||||
// Speculative decoding counters.
|
||||
if let Some(spec_decoding_stats) = &stats.spec_decoding_stats {
|
||||
metrics
|
||||
.spec_decode_num_drafts
|
||||
.get_or_create(&labels)
|
||||
.inc_by(spec_decoding_stats.num_drafts);
|
||||
metrics
|
||||
handles.spec_decode_num_drafts.inc_by(spec_decoding_stats.num_drafts);
|
||||
handles
|
||||
.spec_decode_num_draft_tokens
|
||||
.get_or_create(&labels)
|
||||
.inc_by(spec_decoding_stats.num_draft_tokens);
|
||||
metrics
|
||||
handles
|
||||
.spec_decode_num_accepted_tokens
|
||||
.get_or_create(&labels)
|
||||
.inc_by(spec_decoding_stats.num_accepted_tokens);
|
||||
metrics.log_stats.get_or_create(&labels).observe_spec_decode(
|
||||
handles.log_stats.observe_spec_decode(
|
||||
spec_decoding_stats.num_drafts,
|
||||
&spec_decoding_stats.num_accepted_tokens_per_pos,
|
||||
);
|
||||
@@ -92,11 +192,11 @@ pub(crate) fn record_scheduler_stats(
|
||||
for (position, accepted_tokens) in
|
||||
spec_decoding_stats.num_accepted_tokens_per_pos.iter().copied().enumerate()
|
||||
{
|
||||
metrics
|
||||
handles
|
||||
.spec_decode_num_accepted_tokens_per_pos
|
||||
.get_or_create(&EnginePositionLabels {
|
||||
model_name: model_name.clone(),
|
||||
engine,
|
||||
model_name: handles.labels.model_name.clone(),
|
||||
engine: handles.labels.engine,
|
||||
position: position as u32,
|
||||
})
|
||||
.inc_by(accepted_tokens);
|
||||
@@ -109,22 +209,13 @@ pub(crate) fn record_scheduler_stats(
|
||||
|| perf_stats.num_read_bytes_per_gpu != 0
|
||||
|| perf_stats.num_write_bytes_per_gpu != 0)
|
||||
{
|
||||
metrics
|
||||
.estimated_flops_per_gpu
|
||||
.get_or_create(&labels)
|
||||
.inc_by(perf_stats.num_flops_per_gpu);
|
||||
metrics
|
||||
.estimated_read_bytes_per_gpu
|
||||
.get_or_create(&labels)
|
||||
.inc_by(perf_stats.num_read_bytes_per_gpu);
|
||||
metrics
|
||||
.estimated_write_bytes_per_gpu
|
||||
.get_or_create(&labels)
|
||||
.inc_by(perf_stats.num_write_bytes_per_gpu);
|
||||
handles.estimated_flops_per_gpu.inc_by(perf_stats.num_flops_per_gpu);
|
||||
handles.estimated_read_bytes_per_gpu.inc_by(perf_stats.num_read_bytes_per_gpu);
|
||||
handles.estimated_write_bytes_per_gpu.inc_by(perf_stats.num_write_bytes_per_gpu);
|
||||
}
|
||||
|
||||
if let Some(cudagraph_stats) = &stats.cudagraph_stats {
|
||||
metrics.log_stats.get_or_create(&labels).observe_cudagraph(
|
||||
handles.log_stats.observe_cudagraph(
|
||||
cudagraph_stats.num_unpadded_tokens,
|
||||
cudagraph_stats.num_padded_tokens,
|
||||
cudagraph_stats.num_paddings,
|
||||
@@ -134,16 +225,11 @@ pub(crate) fn record_scheduler_stats(
|
||||
|
||||
// Sampled KV-cache residency histograms.
|
||||
if !stats.kv_cache_eviction_events.is_empty() {
|
||||
let kv_block_lifetime_seconds = metrics.kv_block_lifetime_seconds.get_or_create(&labels);
|
||||
let kv_block_idle_before_evict_seconds =
|
||||
metrics.kv_block_idle_before_evict_seconds.get_or_create(&labels);
|
||||
let kv_block_reuse_gap_seconds = metrics.kv_block_reuse_gap_seconds.get_or_create(&labels);
|
||||
|
||||
for event in &stats.kv_cache_eviction_events {
|
||||
kv_block_lifetime_seconds.observe(event.lifetime_seconds);
|
||||
kv_block_idle_before_evict_seconds.observe(event.idle_seconds);
|
||||
handles.kv_block_lifetime_seconds.observe(event.lifetime_seconds);
|
||||
handles.kv_block_idle_before_evict_seconds.observe(event.idle_seconds);
|
||||
for reuse_gap_seconds in &event.reuse_gaps_seconds {
|
||||
kv_block_reuse_gap_seconds.observe(*reuse_gap_seconds);
|
||||
handles.kv_block_reuse_gap_seconds.observe(*reuse_gap_seconds);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user