feat(mocker): emit forward pass metrics to event plane for planner consumption (#8032)

Signed-off-by: jthomson04 <jwillthomson19@gmail.com>
Signed-off-by: jthomson04 <jothomson@nvidia.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

lib/llm/src/mocker.rs

@@ -303,6 +303,8 @@ pub struct MockEngine {
     bootstrap_server: Arc<OnceCell<Arc<BootstrapServer>>>,
     /// Keep schedulers alive so their CancelGuards don't fire prematurely.
     _schedulers: OnceCell<Vec<Box<dyn SchedulerHandle>>>,
+    /// Forward pass metrics publisher (kept alive for the engine lifetime).
+    _fpm_publisher: OnceCell<crate::fpm_publisher::FpmDirectPublisher>,
 }
 
 impl MockEngine {
@@ -316,6 +318,7 @@ impl MockEngine {
             unset_dp_rank_counter: AtomicU32::new(0),
             bootstrap_server: Arc::new(OnceCell::new()),
             _schedulers: OnceCell::new(),
+            _fpm_publisher: OnceCell::new(),
         }
     }
 
@@ -361,11 +364,33 @@ impl MockEngine {
             None
         };
 
+        // Create FPM publisher upfront and get per-dp-rank sink handles.
+        let worker_id = component.drt().connection_id().to_string();
+        let fpm_sinks = match crate::fpm_publisher::FpmDirectPublisher::new(
+            component.clone(),
+            worker_id,
+            self.engine_args.dp_size,
+        )
+        .await
+        {
+            Ok((publisher, sinks)) => {
+                let _ = self._fpm_publisher.set(publisher);
+                sinks
+            }
+            Err(e) => {
+                tracing::error!("Failed to start FPM publisher: {e}");
+                (0..self.engine_args.dp_size)
+                    .map(|_| dynamo_mocker::common::protocols::FpmPublisher::default())
+                    .collect()
+            }
+        };
+
         let schedulers = self
-            .start_schedulers(kv_component, cancel_token.clone())
+            .start_schedulers(kv_component, cancel_token.clone(), fpm_sinks)
             .await;
 
-        Self::start_metrics_publishing(&schedulers, component, cancel_token.clone()).await?;
+        Self::start_metrics_publishing(&schedulers, component.clone(), cancel_token.clone())
+            .await?;
 
         let _ = self._schedulers.set(schedulers);
 
@@ -395,17 +420,18 @@ impl MockEngine {
         let _ = senders[dp_rank].send(request);
     }
 
-    /// Create schedulers and spawn their background tasks for distributing token notifications
+    /// Create schedulers and spawn their background tasks for distributing token notifications.
     async fn start_schedulers(
         &self,
         component: Option<&Component>,
         cancel_token: CancellationToken,
+        fpm_sinks: Vec<dynamo_mocker::common::protocols::FpmPublisher>,
     ) -> Vec<Box<dyn SchedulerHandle>> {
         let args = &self.engine_args;
         let mut schedulers = Vec::<Box<dyn SchedulerHandle>>::new();
         let mut senders = Vec::with_capacity(args.dp_size as usize);
 
-        for dp_rank in 0..args.dp_size {
+        for (dp_rank, fpm_publisher) in (0..args.dp_size).zip(fpm_sinks) {
             let (output_tx, mut output_rx) = mpsc::unbounded_channel::<Vec<OutputSignal>>();
 
             let (kv_event_publishers, relay_publisher): (
@@ -493,6 +519,7 @@ impl MockEngine {
                 Some(output_tx),
                 kv_event_publishers,
                 Some(cancel_token.clone()),
+                fpm_publisher,
             );
 
             senders.push(scheduler.request_sender());

lib/mocker/src/common/protocols.rs

@@ -80,6 +80,59 @@ impl KvEventPublishers {
     }
 }
 
+/// Per-iteration forward pass snapshot, mirroring the Python `ForwardPassMetrics`
+/// schema in `components/src/dynamo/common/forward_pass_metrics.py`.
+///
+/// Produced by the scheduler core after each `execute_pass_internal()` call.
+/// The runtime-dependent layer (`lib/llm`) wraps this with identity fields
+/// (worker_id, dp_rank, counter_id) and serializes to msgpack for the event plane.
+#[derive(Debug, Clone, Default)]
+pub struct ForwardPassSnapshot {
+    // -- scheduled requests (executed this iteration) --
+    pub num_prefill_requests: u32,
+    pub sum_prefill_tokens: u64,
+    pub var_prefill_length: f64,
+    pub sum_prefill_kv_tokens: u64,
+    pub num_decode_requests: u32,
+    pub sum_decode_kv_tokens: u64,
+    pub var_decode_kv_tokens: f64,
+    // -- queued requests (waiting, not scheduled) --
+    pub num_queued_prefill: u32,
+    pub sum_queued_prefill_tokens: u64,
+    pub var_queued_prefill_length: f64,
+    pub num_queued_decode: u32,
+    pub sum_queued_decode_kv_tokens: u64,
+    pub var_queued_decode_kv_tokens: f64,
+    // -- timing --
+    pub wall_time_secs: f64,
+}
+
+/// Trait for publishing forward pass metrics snapshots.
+/// This abstracts the FPM publishing pipeline so mocker schedulers remain generic.
+pub trait FpmSink: Send + Sync {
+    fn publish(&self, snapshot: ForwardPassSnapshot) -> anyhow::Result<()>;
+}
+
+/// Optional FPM sink used by schedulers.
+/// Wraps `Option<Arc<dyn FpmSink>>` for ergonomic passing and no-op default behavior.
+#[derive(Clone, Default)]
+pub struct FpmPublisher {
+    sink: Option<Arc<dyn FpmSink>>,
+}
+
+impl FpmPublisher {
+    pub fn new(sink: Option<Arc<dyn FpmSink>>) -> Self {
+        Self { sink }
+    }
+
+    pub fn publish(&self, snapshot: ForwardPassSnapshot) -> anyhow::Result<()> {
+        if let Some(sink) = &self.sink {
+            sink.publish(snapshot)?;
+        }
+        Ok(())
+    }
+}
+
 pub type NumBlocks = usize;
 
 /// Represents different block movement operations in the cache

lib/mocker/src/engine.rs

@@ -6,7 +6,9 @@
 use tokio::sync::mpsc;
 use tokio_util::sync::CancellationToken;
 
-use crate::common::protocols::{EngineType, KvEventPublishers, MockEngineArgs, OutputSignal};
+use crate::common::protocols::{
+    EngineType, FpmPublisher, KvEventPublishers, MockEngineArgs, OutputSignal,
+};
 use crate::scheduler::{Scheduler, SchedulerHandle, SglangScheduler};
 
 /// Create a scheduler for the configured engine type.
@@ -19,6 +21,7 @@ pub fn create_engine(
     output_tx: Option<mpsc::UnboundedSender<Vec<OutputSignal>>>,
     kv_event_publishers: KvEventPublishers,
     cancellation_token: Option<CancellationToken>,
+    fpm_publisher: FpmPublisher,
 ) -> Box<dyn SchedulerHandle> {
     match args.engine_type {
         EngineType::Vllm => Box::new(Scheduler::new(
@@ -27,6 +30,7 @@ pub fn create_engine(
             output_tx,
             kv_event_publishers,
             cancellation_token,
+            fpm_publisher,
         )),
         EngineType::Sglang => Box::new(SglangScheduler::new(
             args,
@@ -34,6 +38,7 @@ pub fn create_engine(
             output_tx,
             kv_event_publishers,
             cancellation_token,
+            fpm_publisher,
         )),
     }
 }

lib/mocker/src/replay/online/live_runtime.rs

@@ -11,7 +11,7 @@ use tokio::task::JoinSet;
 use tokio::time::Instant;
 use tokio_util::sync::CancellationToken;
 
-use crate::common::protocols::{DirectRequest, MockEngineArgs, OutputSignal};
+use crate::common::protocols::{DirectRequest, FpmPublisher, MockEngineArgs, OutputSignal};
 use crate::loadgen::WorkloadDriver;
 use crate::replay::{ReplayPrefillLoadEstimator, ReplayRouterMode, TraceSimulationReport};
 use crate::scheduler::{AdmissionEvent, EngineScheduler, SchedulerHandle};
@@ -68,6 +68,7 @@ impl LiveRuntime {
                 router.sink(worker_idx as _),
                 Some(cancel_token.clone()),
                 Some(admission_tx.clone()),
+                FpmPublisher::default(),
             );
             senders.push(scheduler.request_sender());
             schedulers.push(scheduler);

lib/mocker/src/scheduler/kv_event_sink.rs

@@ -6,7 +6,10 @@ use std::sync::{Arc, Mutex};
 use anyhow::Result;
 use dynamo_kv_router::protocols::{KvCacheEvent, RouterEvent, WorkerId};
 
-use crate::common::protocols::{KvCacheEventSink, KvEventPublishers, RawKvEvent, RawKvEventSink};
+use crate::common::protocols::{
+    ForwardPassSnapshot, FpmPublisher, KvCacheEventSink, KvEventPublishers, RawKvEvent,
+    RawKvEventSink,
+};
 
 /// Captures router-ready events for offline replay and scheduler tests.
 ///
@@ -156,3 +159,30 @@ pub(crate) fn publish_deferred_kv_events(
         }
     }
 }
+
+/// Captures FPM snapshots for the live scheduler so it can flush them at the
+/// correct pass phase, matching the deferred KV event pattern.
+#[derive(Clone, Default)]
+pub(crate) struct DeferredFpmBuffer {
+    snapshots: Arc<Mutex<Vec<ForwardPassSnapshot>>>,
+}
+
+impl DeferredFpmBuffer {
+    pub(crate) fn push(&self, snapshot: ForwardPassSnapshot) {
+        self.snapshots.lock().unwrap().push(snapshot);
+    }
+
+    pub(crate) fn drain(&self) -> Vec<ForwardPassSnapshot> {
+        std::mem::take(&mut *self.snapshots.lock().unwrap())
+    }
+}
+
+/// Forwards buffered FPM snapshots to the real sink once the pass reaches
+/// the configured visibility point.
+pub(crate) fn publish_deferred_fpm(sink: &FpmPublisher, snapshots: Vec<ForwardPassSnapshot>) {
+    for snapshot in snapshots {
+        if let Err(error) = sink.publish(snapshot) {
+            tracing::warn!("Failed to forward buffered FPM snapshot: {error}");
+        }
+    }
+}

lib/mocker/src/scheduler/mod.rs

@@ -8,16 +8,106 @@ mod kv_event_sink;
 pub mod sglang;
 pub mod vllm;
 
-use crate::common::protocols::{DirectRequest, KvEventPublishers, OutputSignal};
+pub use crate::common::protocols::ForwardPassSnapshot;
+use crate::common::protocols::{DirectRequest, FpmPublisher, KvEventPublishers, OutputSignal};
 use dynamo_kv_router::protocols::RouterEvent;
 pub(crate) use kv_event_sink::{
-    CapturedRouterEventBuffer, capture_deferred_kv_publish_sink, capture_router_event_sink,
-    publish_deferred_kv_events,
+    CapturedRouterEventBuffer, DeferredFpmBuffer, capture_deferred_kv_publish_sink,
+    capture_router_event_sink, publish_deferred_fpm, publish_deferred_kv_events,
 };
 use tokio::sync::mpsc;
 use tokio_util::sync::CancellationToken;
 use uuid::Uuid;
 
+/// Welford's online algorithm for count / sum / population-variance.
+///
+/// Mirrors the Python `WelfordAccumulator` in `forward_pass_metrics.py`.
+#[derive(Default)]
+pub(crate) struct WelfordAcc {
+    pub(crate) count: u32,
+    pub(crate) sum: f64,
+    mean: f64,
+    m2: f64,
+}
+
+impl WelfordAcc {
+    pub(crate) fn add(&mut self, v: f64) {
+        self.count += 1;
+        self.sum += v;
+        let delta = v - self.mean;
+        self.mean += delta / self.count as f64;
+        let delta2 = v - self.mean;
+        self.m2 += delta * delta2;
+    }
+
+    pub(crate) fn variance(&self) -> f64 {
+        if self.count == 0 {
+            return 0.0;
+        }
+        self.m2 / self.count as f64
+    }
+}
+
+/// Build a [`ForwardPassSnapshot`] from engine-agnostic iterators.
+///
+/// Each engine (vLLM, SGLang) calls this with its own iterators, avoiding
+/// duplicated variance/accumulation logic.
+///
+/// - `scheduled_prefills`: `(prompt_len, prefix_tokens, tokens_computed)` per request
+/// - `scheduled_decodes`: `sequence_len` per request
+/// - `queued_prefills`: `prompt_len` per waiting prefill request
+/// - `queued_decodes`: `kv_tokens` per preempted decode request
+pub(crate) fn build_fpm_snapshot(
+    scheduled_prefills: impl Iterator<Item = (u64, u64, u64)>,
+    scheduled_decodes: impl Iterator<Item = u64>,
+    queued_prefills: impl Iterator<Item = u64>,
+    queued_decodes: impl Iterator<Item = u64>,
+    wall_time_secs: f64,
+) -> ForwardPassSnapshot {
+    let mut prefill_acc = WelfordAcc::default();
+    let mut decode_acc = WelfordAcc::default();
+    let mut sum_prefill_tokens: u64 = 0;
+    let mut sum_prefill_kv_tokens: u64 = 0;
+
+    for (prompt_len, prefix_tokens, tokens_computed) in scheduled_prefills {
+        sum_prefill_tokens += tokens_computed;
+        sum_prefill_kv_tokens += prefix_tokens;
+        prefill_acc.add(prompt_len as f64);
+    }
+
+    for sequence_len in scheduled_decodes {
+        decode_acc.add(sequence_len as f64);
+    }
+
+    let mut queued_prefill_acc = WelfordAcc::default();
+    let mut queued_decode_acc = WelfordAcc::default();
+
+    for prompt_len in queued_prefills {
+        queued_prefill_acc.add(prompt_len as f64);
+    }
+
+    for kv_tokens in queued_decodes {
+        queued_decode_acc.add(kv_tokens as f64);
+    }
+
+    ForwardPassSnapshot {
+        num_prefill_requests: prefill_acc.count,
+        sum_prefill_tokens,
+        var_prefill_length: prefill_acc.variance(),
+        sum_prefill_kv_tokens,
+        num_decode_requests: decode_acc.count,
+        sum_decode_kv_tokens: decode_acc.sum as u64,
+        var_decode_kv_tokens: decode_acc.variance(),
+        num_queued_prefill: queued_prefill_acc.count,
+        sum_queued_prefill_tokens: queued_prefill_acc.sum as u64,
+        var_queued_prefill_length: queued_prefill_acc.variance(),
+        num_queued_decode: queued_decode_acc.count,
+        sum_queued_decode_kv_tokens: queued_decode_acc.sum as u64,
+        var_queued_decode_kv_tokens: queued_decode_acc.variance(),
+        wall_time_secs,
+    }
+}
+
 pub(crate) use sglang::SglangCore;
 pub use sglang::SglangScheduler;
 pub(crate) use vllm::VllmCore;
@@ -41,6 +131,8 @@ pub(crate) struct EnginePassResult {
     pub(crate) router_event_visibility: RouterEventVisibility,
     /// Router-visible KV events emitted during this pass.
     pub(crate) kv_events: Vec<RouterEvent>,
+    /// Forward pass metrics snapshot for this iteration.
+    pub(crate) fpm: Option<ForwardPassSnapshot>,
 }
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
@@ -112,6 +204,7 @@ impl EngineScheduler {
         kv_event_publishers: KvEventPublishers,
         cancellation_token: Option<CancellationToken>,
         admission_tx: Option<mpsc::UnboundedSender<AdmissionEvent>>,
+        fpm_publisher: FpmPublisher,
     ) -> Self {
         match args.engine_type {
             crate::common::protocols::EngineType::Vllm => {
@@ -122,6 +215,7 @@ impl EngineScheduler {
                     kv_event_publishers,
                     cancellation_token,
                     admission_tx,
+                    fpm_publisher,
                 ))
             }
             crate::common::protocols::EngineType::Sglang => {
@@ -132,6 +226,7 @@ impl EngineScheduler {
                     kv_event_publishers,
                     cancellation_token,
                     admission_tx,
+                    fpm_publisher,
                 ))
             }
         }
@@ -179,3 +274,58 @@ pub trait SchedulerHandle: Send + Sync {
 /// Shared test utilities for scheduler stress tests.
 #[cfg(test)]
 pub(crate) mod test_utils;
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn welford_acc_empty() {
+        let acc = WelfordAcc::default();
+        assert_eq!(acc.count, 0);
+        assert_eq!(acc.sum, 0.0);
+        assert_eq!(acc.variance(), 0.0);
+    }
+
+    #[test]
+    fn welford_acc_single_value() {
+        let mut acc = WelfordAcc::default();
+        acc.add(42.0);
+        assert_eq!(acc.count, 1);
+        assert_eq!(acc.sum, 42.0);
+        assert_eq!(acc.variance(), 0.0);
+    }
+
+    #[test]
+    fn welford_acc_population_variance() {
+        let mut acc = WelfordAcc::default();
+        // Values: 2, 4, 4, 4, 5, 5, 7, 9
+        // Mean = 5, Population variance = 4.0
+        for v in [2.0, 4.0, 4.0, 4.0, 5.0, 5.0, 7.0, 9.0] {
+            acc.add(v);
+        }
+        assert_eq!(acc.count, 8);
+        assert_eq!(acc.sum, 40.0);
+        assert!((acc.variance() - 4.0).abs() < 1e-10);
+    }
+
+    #[test]
+    fn welford_acc_matches_python() {
+        // Reproduce the Python WelfordAccumulator behavior:
+        // values = [100, 200, 300], mean = 200,
+        // population variance = ((100-200)^2 + (200-200)^2 + (300-200)^2) / 3
+        //                     = (10000 + 0 + 10000) / 3 = 6666.666...
+        let mut acc = WelfordAcc::default();
+        acc.add(100.0);
+        acc.add(200.0);
+        acc.add(300.0);
+        assert_eq!(acc.count, 3);
+        assert_eq!(acc.sum, 600.0);
+        let expected = 20000.0 / 3.0;
+        assert!(
+            (acc.variance() - expected).abs() < 1e-10,
+            "expected {expected}, got {}",
+            acc.variance()
+        );
+    }
+}

lib/mocker/src/scheduler/sglang/core.rs

@@ -17,7 +17,8 @@ use super::policy::apply_schedule_policy;
 use super::prefill::get_new_batch_prefill;
 use super::request::{SglangRequest, direct_to_sglang};
 use crate::scheduler::{
-    CapturedRouterEventBuffer, EnginePassResult, RouterEventVisibility, capture_router_event_sink,
+    CapturedRouterEventBuffer, EnginePassResult, RouterEventVisibility, build_fpm_snapshot,
+    capture_router_event_sink,
 };
 
 pub(crate) struct SglangCore {
@@ -130,6 +131,9 @@ impl SglangCore {
             }
         }
 
+        // Capture per-request prefill FPM data before dispersing can_run.
+        let prefill_fpm = admit.prefill_fpm;
+
         let batch_size = admit.can_run.len();
         let mean_isl = if batch_size > 0 {
             admit.total_isl / batch_size
@@ -153,6 +157,13 @@ impl SglangCore {
             }
         }
 
+        // Capture scheduled decode data before the decode step modifies running.
+        let scheduled_decode_lens: Vec<u64> = self
+            .running
+            .iter()
+            .map(|req| req.current_sequence_len() as u64)
+            .collect();
+
         let decode_start_ms = now_ms + prefill_time.as_secs_f64() * 1000.0;
         let mut decode = simulate_decode_step(
             &mut self.running,
@@ -178,6 +189,27 @@ impl SglangCore {
         self.new_token_ratio = (self.new_token_ratio - self.config.new_token_ratio_decay_step)
             .max(self.config.min_new_token_ratio);
 
+        // Build FPM snapshot now that all state has settled.
+        let fpm = build_fpm_snapshot(
+            prefill_fpm.iter().map(|p| {
+                (
+                    p.prompt_len as u64,
+                    p.prefix_tokens as u64,
+                    p.tokens_computed as u64,
+                )
+            }),
+            scheduled_decode_lens.into_iter(),
+            self.waiting
+                .iter()
+                .filter(|req| req.output_len() == 0)
+                .map(|req| req.prompt_len() as u64),
+            self.waiting
+                .iter()
+                .filter(|req| req.output_len() > 0)
+                .map(|req| req.current_sequence_len() as u64),
+            (decode.end_ms - now_ms) / 1000.0,
+        );
+
         debug_assert_sglang_scheduler_state(&self.waiting, &self.running, self.config.block_size);
         EnginePassResult {
             end_ms: decode.end_ms,
@@ -195,6 +227,7 @@ impl SglangCore {
                 .as_ref()
                 .map(CapturedRouterEventBuffer::drain)
                 .unwrap_or_default(),
+            fpm: Some(fpm),
         }
     }
 

lib/mocker/src/scheduler/sglang/live.rs

@@ -7,11 +7,13 @@ use std::time::Instant;
 use tokio::sync::mpsc;
 use tokio_util::sync::CancellationToken;
 
-use crate::common::protocols::{DirectRequest, KvEventPublishers, MockEngineArgs, OutputSignal};
+use crate::common::protocols::{
+    DirectRequest, FpmPublisher, KvEventPublishers, MockEngineArgs, OutputSignal,
+};
 use crate::common::utils::sleep_until_precise;
 use crate::scheduler::{
-    AdmissionEvent, MockerMetrics, RouterEventVisibility, SchedulerHandle,
-    capture_deferred_kv_publish_sink, publish_deferred_kv_events,
+    AdmissionEvent, DeferredFpmBuffer, MockerMetrics, RouterEventVisibility, SchedulerHandle,
+    capture_deferred_kv_publish_sink, publish_deferred_fpm, publish_deferred_kv_events,
 };
 
 use super::core::SglangCore;
@@ -39,6 +41,7 @@ impl SglangScheduler {
         output_tx: Option<mpsc::UnboundedSender<Vec<OutputSignal>>>,
         kv_event_publishers: KvEventPublishers,
         cancellation_token: Option<CancellationToken>,
+        fpm_publisher: FpmPublisher,
     ) -> Self {
         Self::new_internal(
             args,
@@ -47,6 +50,7 @@ impl SglangScheduler {
             kv_event_publishers,
             cancellation_token,
             None,
+            fpm_publisher,
         )
     }
 
@@ -57,6 +61,7 @@ impl SglangScheduler {
         kv_event_publishers: KvEventPublishers,
         cancellation_token: Option<CancellationToken>,
         admission_tx: Option<mpsc::UnboundedSender<AdmissionEvent>>,
+        fpm_publisher: FpmPublisher,
     ) -> Self {
         Self::new_internal(
             args,
@@ -65,6 +70,7 @@ impl SglangScheduler {
             kv_event_publishers,
             cancellation_token,
             admission_tx,
+            fpm_publisher,
         )
     }
 
@@ -75,6 +81,7 @@ impl SglangScheduler {
         kv_event_publishers: KvEventPublishers,
         cancellation_token: Option<CancellationToken>,
         admission_tx: Option<mpsc::UnboundedSender<AdmissionEvent>>,
+        fpm_publisher: FpmPublisher,
     ) -> Self {
         let (request_tx, mut request_rx) = mpsc::unbounded_channel::<DirectRequest>();
         let total_blocks = args.num_gpu_blocks as u64;
@@ -89,6 +96,7 @@ impl SglangScheduler {
         tokio::spawn(async move {
             let (deferred_kv_events, buffering_publishers) =
                 capture_deferred_kv_publish_sink(kv_event_publishers.raw_enabled());
+            let deferred_fpm = DeferredFpmBuffer::default();
             let mut core = SglangCore::new_with_sink(args, dp_rank, buffering_publishers);
 
             loop {
@@ -111,8 +119,12 @@ impl SglangScheduler {
                         let _ = admission_tx.send(admission.clone());
                     }
                 }
+                if let Some(fpm) = pass.fpm {
+                    deferred_fpm.push(fpm);
+                }
                 if pass.router_event_visibility == RouterEventVisibility::PassStart {
                     publish_deferred_kv_events(&kv_event_publishers, deferred_kv_events.drain());
+                    publish_deferred_fpm(&fpm_publisher, deferred_fpm.drain());
                 }
                 let total_time = std::time::Duration::from_secs_f64(pass.end_ms / 1000.0);
                 if total_time > std::time::Duration::ZERO {
@@ -120,10 +132,12 @@ impl SglangScheduler {
                 }
                 if pass.router_event_visibility == RouterEventVisibility::PassEnd {
                     publish_deferred_kv_events(&kv_event_publishers, deferred_kv_events.drain());
+                    publish_deferred_fpm(&fpm_publisher, deferred_fpm.drain());
                 }
                 let active_decode_blocks = pass.active_decode_blocks;
                 flush_output_signals(&output_tx, pass.output_signals);
                 publish_deferred_kv_events(&kv_event_publishers, deferred_kv_events.drain());
+                publish_deferred_fpm(&fpm_publisher, deferred_fpm.drain());
                 let _ = metrics_tx.send(MockerMetrics::new(
                     dp_rank,
                     active_decode_blocks,

lib/mocker/src/scheduler/sglang/prefill.rs

@@ -8,12 +8,21 @@ use super::config::{SglangConfig, ceil_to_block};
 use super::request::SglangRequest;
 use crate::kv_manager::SglangKvManager;
 
+/// Per-request prefill data needed for FPM snapshot construction.
+pub(super) struct PrefillFpmItem {
+    pub(super) prompt_len: usize,
+    pub(super) tokens_computed: usize,
+    pub(super) prefix_tokens: usize,
+}
+
 pub(super) struct AdmitResult {
     pub(super) can_run: Vec<SglangRequest>,
     pub(super) admissions: Vec<AdmissionEvent>,
     pub(super) total_isl: usize,
     pub(super) total_prefix: usize,
     pub(super) oom: bool,
+    /// Per-request prefill info for building FPM snapshots.
+    pub(super) prefill_fpm: Vec<PrefillFpmItem>,
 }
 
 pub(super) fn get_new_batch_prefill(
@@ -50,6 +59,7 @@ pub(super) fn get_new_batch_prefill(
 
     let mut can_run = Vec::new();
     let mut admissions = Vec::new();
+    let mut prefill_fpm = Vec::new();
     let mut rejected = VecDeque::new();
     let mut oom = false;
     let mut total_isl = 0usize;
@@ -139,6 +149,11 @@ pub(super) fn get_new_batch_prefill(
             uuid: req.uuid,
             reused_input_tokens: alloc.prefix_len,
         });
+        prefill_fpm.push(PrefillFpmItem {
+            prompt_len: req.prompt_len(),
+            tokens_computed: chunk_tokens,
+            prefix_tokens: alloc.prefix_len,
+        });
 
         total_isl += chunk_end;
         total_prefix += alloc.prefix_len;
@@ -162,5 +177,6 @@ pub(super) fn get_new_batch_prefill(
         total_isl,
         total_prefix,
         oom,
+        prefill_fpm,
     }
 }

lib/mocker/src/scheduler/sglang/tests.rs

@@ -19,7 +19,8 @@ use super::policy::apply_schedule_policy;
 use super::prefill::get_new_batch_prefill;
 use super::request::SglangRequest;
 use crate::common::protocols::{
-    DirectRequest, EngineType, KvEventPublishers, MockEngineArgs, OutputSignal, SglangArgs,
+    DirectRequest, EngineType, FpmPublisher, KvEventPublishers, MockEngineArgs, OutputSignal,
+    SglangArgs,
 };
 use crate::kv_manager::SglangKvManager;
 use crate::scheduler::test_utils::{
@@ -95,8 +96,14 @@ mod scheduling {
             .unwrap();
 
         let (output_tx, mut output_rx) = mpsc::unbounded_channel::<Vec<OutputSignal>>();
-        let scheduler =
-            SglangScheduler::new(args, 0, Some(output_tx), KvEventPublishers::default(), None);
+        let scheduler = SglangScheduler::new(
+            args,
+            0,
+            Some(output_tx),
+            KvEventPublishers::default(),
+            None,
+            FpmPublisher::default(),
+        );
 
         let num_requests = 5;
         let max_output = 3;
@@ -616,8 +623,14 @@ mod router_events {
             }))
             .build()
             .unwrap();
-        let scheduler =
-            SglangScheduler::new(args, 0, Some(output_tx), KvEventPublishers::default(), None);
+        let scheduler = SglangScheduler::new(
+            args,
+            0,
+            Some(output_tx),
+            KvEventPublishers::default(),
+            None,
+            FpmPublisher::default(),
+        );
 
         assert_sglang_scheduler_completes_all(
             &scheduler,
@@ -836,6 +849,7 @@ mod router_events {
             Some(output_tx),
             KvEventPublishers::new(Some(sink.clone()), None),
             None,
+            FpmPublisher::default(),
         );
 
         for _ in 0..8 {
@@ -908,3 +922,408 @@ mod router_events {
         assert_eq!(signal.handoff_delay_ms, Some(8.0));
     }
 }
+
+mod forward_pass_metrics {
+    use super::*;
+
+    fn fpm_args() -> MockEngineArgs {
+        MockEngineArgs::builder()
+            .engine_type(EngineType::Sglang)
+            .block_size(4)
+            .num_gpu_blocks(16)
+            .max_num_batched_tokens(Some(16))
+            .max_num_seqs(Some(4))
+            .speedup_ratio(0.0)
+            .sglang(Some(SglangArgs {
+                page_size: Some(4),
+                chunked_prefill_size: Some(16),
+                ..Default::default()
+            }))
+            .build()
+            .unwrap()
+    }
+
+    #[test]
+    fn test_fpm_single_prefill_request() {
+        let mut core = SglangCore::new(fpm_args());
+        core.receive(DirectRequest {
+            tokens: (0..8).collect(),
+            max_output_tokens: 1,
+            uuid: Some(Uuid::from_u128(1)),
+            dp_rank: 0,
+            arrival_timestamp_ms: None,
+        });
+
+        let mut collector = crate::replay::TraceCollector::default();
+        let pass = core.execute_pass(&mut collector, 0.0);
+        let fpm = pass.fpm.expect("FPM should be present");
+
+        assert_eq!(fpm.num_prefill_requests, 1);
+        assert!(
+            fpm.sum_prefill_tokens > 0,
+            "prefill tokens should be computed"
+        );
+        // In SGLang, after prefill the request immediately joins running and
+        // participates in the decode step of the same pass.
+        assert_eq!(fpm.num_decode_requests, 1);
+        assert_eq!(fpm.num_queued_prefill, 0);
+        assert_eq!(fpm.num_queued_decode, 0);
+        assert!(fpm.wall_time_secs > 0.0);
+    }
+
+    #[test]
+    fn test_fpm_prefill_and_decode_mixed_batch() {
+        let mut core = SglangCore::new(fpm_args());
+
+        // r1: 4-token prompt, 3 output tokens
+        core.receive(DirectRequest {
+            tokens: (0..4).collect(),
+            max_output_tokens: 3,
+            uuid: Some(Uuid::from_u128(1)),
+            dp_rank: 0,
+            arrival_timestamp_ms: None,
+        });
+
+        let mut collector = crate::replay::TraceCollector::default();
+
+        // Pass 1: prefill r1
+        let pass1 = core.execute_pass(&mut collector, 0.0);
+        let fpm1 = pass1.fpm.expect("FPM should be present");
+        assert_eq!(fpm1.num_prefill_requests, 1);
+
+        // r2: arriving while r1 is decoding
+        core.receive(DirectRequest {
+            tokens: (100..104).collect(),
+            max_output_tokens: 3,
+            uuid: Some(Uuid::from_u128(2)),
+            dp_rank: 0,
+            arrival_timestamp_ms: None,
+        });
+
+        // Pass 2: r2 prefill + decode step runs on all running (r1 + r2)
+        let pass2 = core.execute_pass(&mut collector, 1.0);
+        let fpm2 = pass2.fpm.expect("FPM should be present");
+        assert_eq!(fpm2.num_prefill_requests, 1, "r2 is prefilling");
+        // In SGLang, after r2 prefill completes it joins running alongside r1,
+        // so the decode step sees both.
+        assert_eq!(fpm2.num_decode_requests, 2, "r1 + r2 both in decode step");
+        assert!(
+            fpm2.sum_decode_kv_tokens > 0,
+            "decode requests should have KV context"
+        );
+    }
+
+    #[test]
+    fn test_fpm_empty_pass_is_zeroed() {
+        let mut core = SglangCore::new(fpm_args());
+
+        // Submit and fully drain a request first so the core isn't empty
+        // (empty core blocks in receive_requests in live mode, but
+        // execute_pass_internal works fine on an empty core).
+        let pass = core.execute_hidden_pass(0.0);
+        let fpm = pass.fpm.expect("FPM should be present even for empty pass");
+
+        assert_eq!(fpm.num_prefill_requests, 0);
+        assert_eq!(fpm.num_decode_requests, 0);
+        assert_eq!(fpm.num_queued_prefill, 0);
+        assert_eq!(fpm.num_queued_decode, 0);
+        assert_eq!(fpm.sum_prefill_tokens, 0);
+        assert_eq!(fpm.sum_decode_kv_tokens, 0);
+    }
+
+    #[test]
+    fn test_fpm_queued_requests() {
+        // Very limited KV to force queuing.
+        let args = MockEngineArgs::builder()
+            .engine_type(EngineType::Sglang)
+            .block_size(4)
+            .num_gpu_blocks(4)
+            .max_num_batched_tokens(Some(8))
+            .max_num_seqs(Some(2))
+            .speedup_ratio(0.0)
+            .sglang(Some(SglangArgs {
+                page_size: Some(4),
+                chunked_prefill_size: Some(8),
+                ..Default::default()
+            }))
+            .build()
+            .unwrap();
+        let mut core = SglangCore::new(args);
+
+        // Two 8-token requests but limited KV
+        core.receive(DirectRequest {
+            tokens: (0..8).collect(),
+            max_output_tokens: 1,
+            uuid: Some(Uuid::from_u128(1)),
+            dp_rank: 0,
+            arrival_timestamp_ms: None,
+        });
+        core.receive(DirectRequest {
+            tokens: (100..108).collect(),
+            max_output_tokens: 1,
+            uuid: Some(Uuid::from_u128(2)),
+            dp_rank: 0,
+            arrival_timestamp_ms: None,
+        });
+
+        let mut collector = crate::replay::TraceCollector::default();
+        let pass = core.execute_pass(&mut collector, 0.0);
+        let fpm = pass.fpm.expect("FPM should be present");
+
+        let total_scheduled = fpm.num_prefill_requests + fpm.num_decode_requests;
+        assert!(
+            total_scheduled >= 1,
+            "at least one request should be scheduled"
+        );
+        // With tight KV, the second request should be queued.
+        let total_queued = fpm.num_queued_prefill + fpm.num_queued_decode;
+        assert!(
+            total_queued >= 1,
+            "at least one request should be queued, got {total_queued}"
+        );
+    }
+
+    #[test]
+    fn test_fpm_var_prefill_length_with_multiple_requests() {
+        let args = MockEngineArgs::builder()
+            .engine_type(EngineType::Sglang)
+            .block_size(4)
+            .num_gpu_blocks(32)
+            .max_num_batched_tokens(Some(32))
+            .max_num_seqs(Some(4))
+            .speedup_ratio(0.0)
+            .sglang(Some(SglangArgs {
+                page_size: Some(4),
+                chunked_prefill_size: Some(32),
+                ..Default::default()
+            }))
+            .build()
+            .unwrap();
+        let mut core = SglangCore::new(args);
+
+        // Two prefill requests with different prompt lengths
+        core.receive(DirectRequest {
+            tokens: (0..4).collect(), // prompt_len = 4
+            max_output_tokens: 1,
+            uuid: Some(Uuid::from_u128(1)),
+            dp_rank: 0,
+            arrival_timestamp_ms: None,
+        });
+        core.receive(DirectRequest {
+            tokens: (100..112).collect(), // prompt_len = 12
+            max_output_tokens: 1,
+            uuid: Some(Uuid::from_u128(2)),
+            dp_rank: 0,
+            arrival_timestamp_ms: None,
+        });
+
+        let mut collector = crate::replay::TraceCollector::default();
+        let pass = core.execute_pass(&mut collector, 0.0);
+        let fpm = pass.fpm.expect("FPM should be present");
+
+        assert_eq!(fpm.num_prefill_requests, 2);
+        // Population variance of [4, 12]: mean=8, var=((4-8)^2+(12-8)^2)/2 = 16
+        assert!(
+            (fpm.var_prefill_length - 16.0).abs() < 1e-6,
+            "expected var=16.0, got {}",
+            fpm.var_prefill_length
+        );
+    }
+
+    #[test]
+    fn test_fpm_chunked_prefill_reports_chunk_not_full_prompt() {
+        // With chunked_prefill_size=8 and a 16-token prompt, the request
+        // should be chunked. Each pass should report only the chunk size
+        // in sum_prefill_tokens, not the full prompt length.
+        let args = MockEngineArgs::builder()
+            .engine_type(EngineType::Sglang)
+            .block_size(4)
+            .num_gpu_blocks(32)
+            .max_num_batched_tokens(Some(32))
+            .max_num_seqs(Some(4))
+            .speedup_ratio(0.0)
+            .sglang(Some(SglangArgs {
+                page_size: Some(4),
+                chunked_prefill_size: Some(8),
+                ..Default::default()
+            }))
+            .build()
+            .unwrap();
+        let mut core = SglangCore::new(args);
+
+        core.receive(DirectRequest {
+            tokens: (0..16).collect(),
+            max_output_tokens: 2,
+            uuid: Some(Uuid::from_u128(1)),
+            dp_rank: 0,
+            arrival_timestamp_ms: None,
+        });
+
+        let mut collector = crate::replay::TraceCollector::default();
+
+        // Pass 1: first chunk
+        let pass1 = core.execute_pass(&mut collector, 0.0);
+        let fpm1 = pass1.fpm.expect("FPM should be present");
+        assert_eq!(fpm1.num_prefill_requests, 1);
+        assert!(
+            fpm1.sum_prefill_tokens <= 8,
+            "chunk should be at most 8 tokens, got {}",
+            fpm1.sum_prefill_tokens
+        );
+        assert!(fpm1.sum_prefill_tokens > 0);
+
+        // Pass 2: remaining chunk
+        let pass2 = core.execute_pass(&mut collector, 1.0);
+        let fpm2 = pass2.fpm.expect("FPM should be present");
+        assert_eq!(fpm2.num_prefill_requests, 1, "still prefilling");
+        assert!(
+            fpm2.sum_prefill_tokens <= 8,
+            "second chunk should also be at most 8 tokens, got {}",
+            fpm2.sum_prefill_tokens
+        );
+
+        // Total across both chunks should equal the full prompt length
+        assert_eq!(
+            fpm1.sum_prefill_tokens + fpm2.sum_prefill_tokens,
+            16,
+            "total prefill tokens across chunks should equal full prompt"
+        );
+    }
+
+    #[test]
+    fn test_fpm_retracted_decode_becomes_queued_decode() {
+        // Very tight KV to force decode retraction. Fill the KV with running
+        // requests, then the decode step should retract some, and those should
+        // appear as queued decodes in the next pass's FPM.
+        let args = MockEngineArgs::builder()
+            .engine_type(EngineType::Sglang)
+            .block_size(4)
+            .num_gpu_blocks(6) // 24 tokens — very tight
+            .max_num_batched_tokens(Some(32))
+            .max_num_seqs(Some(4))
+            .speedup_ratio(0.0)
+            .sglang(Some(SglangArgs {
+                page_size: Some(4),
+                chunked_prefill_size: Some(32),
+                ..Default::default()
+            }))
+            .build()
+            .unwrap();
+        let mut core = SglangCore::new(args);
+        let mut collector = crate::replay::TraceCollector::default();
+
+        // Two requests with 4-token prompts and long outputs to fill KV
+        core.receive(DirectRequest {
+            tokens: (0..4).collect(),
+            max_output_tokens: 20,
+            uuid: Some(Uuid::from_u128(1)),
+            dp_rank: 0,
+            arrival_timestamp_ms: None,
+        });
+        core.receive(DirectRequest {
+            tokens: (100..104).collect(),
+            max_output_tokens: 20,
+            uuid: Some(Uuid::from_u128(2)),
+            dp_rank: 0,
+            arrival_timestamp_ms: None,
+        });
+
+        // Run several passes to build up KV pressure
+        for i in 0..4 {
+            core.execute_pass(&mut collector, i as f64);
+        }
+
+        // Add a third request to increase memory pressure
+        core.receive(DirectRequest {
+            tokens: (200..212).collect(), // 12 tokens
+            max_output_tokens: 10,
+            uuid: Some(Uuid::from_u128(3)),
+            dp_rank: 0,
+            arrival_timestamp_ms: None,
+        });
+
+        // Run more passes — at some point retraction should occur
+        let mut saw_queued_decode = false;
+        for i in 4..10 {
+            let pass = core.execute_pass(&mut collector, i as f64);
+            let fpm = pass.fpm.expect("FPM should be present");
+            if fpm.num_queued_decode > 0 {
+                saw_queued_decode = true;
+                assert!(
+                    fpm.sum_queued_decode_kv_tokens > 0,
+                    "retracted decode should have KV context"
+                );
+                break;
+            }
+        }
+
+        // If retraction didn't happen (KV was sufficient), that's also valid —
+        // just verify we always get Some(fpm).
+        if !saw_queued_decode {
+            // Verify the requests completed or are still running with valid FPM
+            let pass = core.execute_hidden_pass(10.0);
+            assert!(pass.fpm.is_some(), "FPM should always be present");
+        }
+    }
+
+    #[tokio::test]
+    async fn test_fpm_sent_through_sink() {
+        use std::sync::Arc;
+
+        use crate::common::protocols::FpmSink;
+        use crate::scheduler::test_utils::CapturingFpmSink;
+
+        let args = MockEngineArgs::builder()
+            .engine_type(EngineType::Sglang)
+            .block_size(4)
+            .num_gpu_blocks(16)
+            .max_num_batched_tokens(Some(16))
+            .max_num_seqs(Some(4))
+            .speedup_ratio(0.0)
+            .sglang(Some(SglangArgs {
+                page_size: Some(4),
+                chunked_prefill_size: Some(16),
+                ..Default::default()
+            }))
+            .build()
+            .unwrap();
+
+        let (output_tx, mut output_rx) = mpsc::unbounded_channel::<Vec<OutputSignal>>();
+        let fpm_sink = Arc::new(CapturingFpmSink::default());
+        let fpm_publisher = FpmPublisher::new(Some(fpm_sink.clone() as Arc<dyn FpmSink>));
+
+        let scheduler = SglangScheduler::new(
+            args,
+            0,
+            Some(output_tx),
+            KvEventPublishers::default(),
+            None,
+            fpm_publisher,
+        );
+
+        scheduler.receive(DirectRequest {
+            tokens: (0..8).collect(),
+            max_output_tokens: 2,
+            uuid: Some(Uuid::from_u128(1)),
+            dp_rank: 0,
+            arrival_timestamp_ms: None,
+        });
+
+        // Wait for at least one output signal — ensures the scheduler has
+        // completed at least one pass and drained the deferred FPM buffer.
+        tokio::time::timeout(Duration::from_secs(5), output_rx.recv())
+            .await
+            .expect("timed out waiting for output")
+            .expect("output channel closed");
+
+        let snapshots = fpm_sink.take();
+        assert!(
+            !snapshots.is_empty(),
+            "should have received at least one FPM snapshot"
+        );
+        let fpm = &snapshots[0];
+        assert_eq!(fpm.num_prefill_requests, 1);
+        assert!(fpm.sum_prefill_tokens > 0);
+        assert!(fpm.wall_time_secs > 0.0);
+    }
+}

lib/mocker/src/scheduler/test_utils.rs

 // SPDX-FileCopyrightText: Copyright (c) 2024-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 // SPDX-License-Identifier: Apache-2.0
 
-use std::sync::Arc;
+use std::sync::{Arc, Mutex};
 
 use anyhow::anyhow;
 use dynamo_kv_router::indexer::{
@@ -17,8 +17,8 @@ use tokio::task::JoinHandle;
 use tokio::time::Duration;
 use tokio_util::sync::CancellationToken;
 
-use super::{DirectRequest, OutputSignal, SchedulerHandle};
-use crate::common::protocols::KvCacheEventSink;
+use super::{DirectRequest, ForwardPassSnapshot, OutputSignal, SchedulerHandle};
+use crate::common::protocols::{FpmSink, KvCacheEventSink};
 
 pub(crate) struct RouterIndexerHarness {
     indexer: Arc<LocalKvIndexer>,
@@ -207,6 +207,25 @@ pub(crate) fn removed_event_count(events: &[RouterEvent]) -> usize {
         .count()
 }
 
+/// Test sink that captures FPM snapshots for assertion.
+#[derive(Default)]
+pub(crate) struct CapturingFpmSink {
+    snapshots: Mutex<Vec<ForwardPassSnapshot>>,
+}
+
+impl FpmSink for CapturingFpmSink {
+    fn publish(&self, snapshot: ForwardPassSnapshot) -> anyhow::Result<()> {
+        self.snapshots.lock().unwrap().push(snapshot);
+        Ok(())
+    }
+}
+
+impl CapturingFpmSink {
+    pub(crate) fn take(&self) -> Vec<ForwardPassSnapshot> {
+        std::mem::take(&mut *self.snapshots.lock().unwrap())
+    }
+}
+
 /// Send `num_requests` to a scheduler, collect all output signals, and assert
 /// that the scheduler produces exactly `num_requests * max_output_tokens` signals
 /// and returns to idle (0 active decode blocks).

lib/mocker/src/scheduler/vllm/core.rs

@@ -19,8 +19,8 @@ use crate::common::utils::compute_prefill_handoff_delay_ms;
 use crate::kv_manager::KvManager;
 use crate::replay::TraceCollector;
 use crate::scheduler::{
-    AdmissionEvent, CapturedRouterEventBuffer, EnginePassResult, RouterEventVisibility,
-    capture_router_event_sink,
+    AdmissionEvent, CapturedRouterEventBuffer, EnginePassResult, ForwardPassSnapshot,
+    RouterEventVisibility, build_fpm_snapshot, capture_router_event_sink,
 };
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
@@ -56,6 +56,12 @@ struct ScheduledWork {
     total_tokens: usize,
     prompt_tokens: usize,
     prefix_tokens: usize,
+    /// Full prompt length, captured at schedule time for FPM variance calculation.
+    prompt_len: usize,
+    /// Total sequence length (prompt + generated) at schedule time, used for
+    /// decode KV context in FPM. Captured here because completed requests are
+    /// removed from state before `compute_fpm` runs.
+    sequence_len: usize,
 }
 
 enum ScheduleOutcome {
@@ -380,6 +386,8 @@ impl VllmCore {
         let (decode_time, output_signals) = self.emit_ready_tokens(collector, decode_start_ms);
         let end_ms = decode_start_ms + decode_time.as_secs_f64() * 1000.0;
 
+        let fpm = self.compute_fpm(&scheduled, (end_ms - now_ms) / 1000.0);
+
         debug_assert_vllm_scheduler_state(&self.state);
         EnginePassResult {
             end_ms,
@@ -393,6 +401,7 @@ impl VllmCore {
                 .as_ref()
                 .map(CapturedRouterEventBuffer::drain)
                 .unwrap_or_default(),
+            fpm: Some(fpm),
         }
     }
 
@@ -406,6 +415,52 @@ impl VllmCore {
         self.state.complete(&uuid);
     }
 
+    /// Compute a forward pass metrics snapshot from the just-completed pass.
+    ///
+    /// `scheduled` contains the work items that were scheduled in this iteration.
+    /// Per-request metadata (prompt_len, sequence_len) is captured in `ScheduledWork`
+    /// at schedule time, so this method does not depend on `self.state.requests` for
+    /// scheduled requests — completed requests may have already been removed.
+    /// Queue metrics are derived from `self.state.waiting` at the moment of the call.
+    fn compute_fpm(
+        &self,
+        scheduled: &FxHashMap<Uuid, ScheduledWork>,
+        wall_time_secs: f64,
+    ) -> ForwardPassSnapshot {
+        let scheduled_prefills = scheduled.values().filter_map(|work| {
+            (work.prompt_tokens > 0).then_some((
+                work.prompt_len as u64,
+                work.prefix_tokens as u64,
+                work.total_tokens as u64,
+            ))
+        });
+
+        let scheduled_decodes = scheduled
+            .values()
+            .filter_map(|work| (work.prompt_tokens == 0).then_some(work.sequence_len as u64));
+
+        let queued_prefills = self.state.waiting.iter().filter_map(|uuid| {
+            let request = self.state.requests.get(uuid)?;
+            matches!(request.status, RequestStatus::Waiting)
+                .then_some(request.sequence.num_input_tokens() as u64)
+        });
+
+        let queued_decodes = self.state.waiting.iter().filter_map(|uuid| {
+            let request = self.state.requests.get(uuid)?;
+            matches!(request.status, RequestStatus::Preempted).then_some(
+                (request.sequence.num_input_tokens() + request.sequence.generated_tokens()) as u64,
+            )
+        });
+
+        build_fpm_snapshot(
+            scheduled_prefills,
+            scheduled_decodes,
+            queued_prefills,
+            queued_decodes,
+            wall_time_secs,
+        )
+    }
+
     #[allow(clippy::too_many_arguments)]
     fn schedule_request(
         &mut self,
@@ -539,12 +594,20 @@ impl VllmCore {
 
         let prompt_after = actual_computed_after.min(prompt_len);
         let prompt_tokens = prompt_after.saturating_sub(prompt_before);
+        let sequence_len = self
+            .state
+            .requests
+            .get(&uuid)
+            .map(|r| r.sequence.len())
+            .unwrap_or(0);
         scheduled.insert(
             uuid,
             ScheduledWork {
                 total_tokens: tokens_used,
                 prompt_tokens,
                 prefix_tokens: prompt_before,
+                prompt_len,
+                sequence_len,
             },
         );
         if prompt_tokens > 0 && self.args.worker_type != WorkerType::Decode {

lib/mocker/src/scheduler/vllm/live.rs

@@ -7,11 +7,13 @@ use std::time::Instant;
 use tokio::sync::mpsc;
 use tokio_util::sync::CancellationToken;
 
-use crate::common::protocols::{DirectRequest, KvEventPublishers, MockEngineArgs, OutputSignal};
+use crate::common::protocols::{
+    DirectRequest, FpmPublisher, KvEventPublishers, MockEngineArgs, OutputSignal,
+};
 use crate::common::utils::sleep_until_precise;
 use crate::scheduler::{
-    AdmissionEvent, RouterEventVisibility, SchedulerHandle, capture_deferred_kv_publish_sink,
-    publish_deferred_kv_events,
+    AdmissionEvent, DeferredFpmBuffer, RouterEventVisibility, SchedulerHandle,
+    capture_deferred_kv_publish_sink, publish_deferred_fpm, publish_deferred_kv_events,
 };
 
 use super::core::VllmCore;
@@ -66,6 +68,7 @@ impl Scheduler {
         output_tx: Option<mpsc::UnboundedSender<Vec<OutputSignal>>>,
         kv_event_publishers: KvEventPublishers,
         cancellation_token: Option<CancellationToken>,
+        fpm_publisher: FpmPublisher,
     ) -> Self {
         Self::new_internal(
             args,
@@ -74,6 +77,7 @@ impl Scheduler {
             kv_event_publishers,
             cancellation_token,
             None,
+            fpm_publisher,
         )
     }
 
@@ -84,6 +88,7 @@ impl Scheduler {
         kv_event_publishers: KvEventPublishers,
         cancellation_token: Option<CancellationToken>,
         admission_tx: Option<mpsc::UnboundedSender<AdmissionEvent>>,
+        fpm_publisher: FpmPublisher,
     ) -> Self {
         Self::new_internal(
             args,
@@ -92,6 +97,7 @@ impl Scheduler {
             kv_event_publishers,
             cancellation_token,
             admission_tx,
+            fpm_publisher,
         )
     }
 
@@ -102,6 +108,7 @@ impl Scheduler {
         kv_event_publishers: KvEventPublishers,
         cancellation_token: Option<CancellationToken>,
         admission_tx: Option<mpsc::UnboundedSender<AdmissionEvent>>,
+        fpm_publisher: FpmPublisher,
     ) -> Self {
         let (request_tx, mut request_rx) = mpsc::unbounded_channel::<DirectRequest>();
         let total_blocks = args.num_gpu_blocks as u64;
@@ -115,6 +122,7 @@ impl Scheduler {
         tokio::spawn(async move {
             let (deferred_kv_events, buffering_publishers) =
                 capture_deferred_kv_publish_sink(kv_event_publishers.raw_enabled());
+            let deferred_fpm = DeferredFpmBuffer::default();
             let mut core = VllmCore::new_with_sink(args, dp_rank, buffering_publishers);
 
             loop {
@@ -128,17 +136,23 @@ impl Scheduler {
                 let iteration_start = Instant::now();
                 let pass = core.execute_pass_internal(None, 0.0, admission_tx.as_ref());
                 let total_time = std::time::Duration::from_secs_f64(pass.end_ms / 1000.0);
+                if let Some(fpm) = pass.fpm {
+                    deferred_fpm.push(fpm);
+                }
                 if pass.router_event_visibility == RouterEventVisibility::PassStart {
                     publish_deferred_kv_events(&kv_event_publishers, deferred_kv_events.drain());
+                    publish_deferred_fpm(&fpm_publisher, deferred_fpm.drain());
                 }
                 if total_time > std::time::Duration::ZERO {
                     sleep_until_precise(iteration_start + total_time).await;
                 }
                 if pass.router_event_visibility == RouterEventVisibility::PassEnd {
                     publish_deferred_kv_events(&kv_event_publishers, deferred_kv_events.drain());
+                    publish_deferred_fpm(&fpm_publisher, deferred_fpm.drain());
                 }
                 flush_output_signals(&mut core, &output_tx, pass.output_signals);
                 publish_deferred_kv_events(&kv_event_publishers, deferred_kv_events.drain());
+                publish_deferred_fpm(&fpm_publisher, deferred_fpm.drain());
                 let _ = metrics_tx.send(MockerMetrics::new(
                     dp_rank,
                     core.kv_manager.num_active_blocks() as u64,

lib/mocker/src/scheduler/vllm/tests.rs

@@ -12,7 +12,7 @@ use tokio::time::interval;
 use uuid::Uuid;
 
 use crate::common::protocols::{
-    DirectRequest, KvCacheEventSink, KvEventPublishers, MockEngineArgs, OutputSignal,
+    DirectRequest, FpmPublisher, KvCacheEventSink, KvEventPublishers, MockEngineArgs, OutputSignal,
     PreemptionMode, RawKvEvent, RawKvEventSink,
 };
 use crate::common::sequence::ActiveSequence;
@@ -577,8 +577,14 @@ mod live_scheduler {
             .build()
             .unwrap();
 
-        let scheduler =
-            Scheduler::new(args, 0, Some(output_tx), KvEventPublishers::default(), None);
+        let scheduler = Scheduler::new(
+            args,
+            0,
+            Some(output_tx),
+            KvEventPublishers::default(),
+            None,
+            FpmPublisher::default(),
+        );
 
         crate::scheduler::test_utils::assert_scheduler_completes_all(
             &scheduler,
@@ -608,8 +614,14 @@ mod live_scheduler {
             .build()
             .unwrap();
 
-        let scheduler =
-            Scheduler::new(args, 0, Some(output_tx), KvEventPublishers::default(), None);
+        let scheduler = Scheduler::new(
+            args,
+            0,
+            Some(output_tx),
+            KvEventPublishers::default(),
+            None,
+            FpmPublisher::default(),
+        );
         let identical_tokens: Vec<u32> = (0..token_length).collect();
 
         for _ in 0..num_requests {
@@ -660,8 +672,14 @@ mod live_scheduler {
             .build()
             .unwrap();
 
-        let scheduler =
-            Scheduler::new(args, 0, Some(output_tx), KvEventPublishers::default(), None);
+        let scheduler = Scheduler::new(
+            args,
+            0,
+            Some(output_tx),
+            KvEventPublishers::default(),
+            None,
+            FpmPublisher::default(),
+        );
         scheduler.receive(DirectRequest {
             tokens: (0..256).collect(),
             max_output_tokens: 200,
@@ -717,6 +735,7 @@ mod live_scheduler {
             Some(output_tx),
             KvEventPublishers::new(None, Some(sink.clone())),
             None,
+            FpmPublisher::default(),
         );
 
         scheduler.receive(DirectRequest {
@@ -771,6 +790,7 @@ mod live_scheduler {
             Some(output_tx),
             KvEventPublishers::new(Some(sink.clone()), None),
             None,
+            FpmPublisher::default(),
         );
 
         for _ in 0..8 {
@@ -813,3 +833,439 @@ mod live_scheduler {
         harness.shutdown();
     }
 }
+
+mod forward_pass_metrics {
+    use super::*;
+
+    /// Helper to build args with specific parameters for FPM tests.
+    fn fpm_args() -> MockEngineArgs {
+        MockEngineArgs::builder()
+            .block_size(4)
+            .num_gpu_blocks(16)
+            .max_num_batched_tokens(Some(16))
+            .max_num_seqs(Some(4))
+            .enable_chunked_prefill(true)
+            .enable_prefix_caching(false)
+            .speedup_ratio(0.0)
+            .build()
+            .unwrap()
+    }
+
+    #[test]
+    fn test_fpm_single_prefill_request() {
+        let mut core = VllmCore::new(fpm_args());
+        core.receive(DirectRequest {
+            tokens: (0..8).collect(),
+            max_output_tokens: 1,
+            uuid: Some(Uuid::from_u128(1)),
+            dp_rank: 0,
+            arrival_timestamp_ms: None,
+        });
+
+        let mut collector = crate::replay::TraceCollector::default();
+        let pass = core.execute_pass(&mut collector, 0.0);
+        let fpm = pass.fpm.expect("FPM should be present");
+
+        assert_eq!(fpm.num_prefill_requests, 1);
+        assert_eq!(fpm.sum_prefill_tokens, 8, "all 8 prompt tokens computed");
+        assert_eq!(fpm.sum_prefill_kv_tokens, 0, "no prefix cache");
+        assert_eq!(fpm.num_decode_requests, 0);
+        assert_eq!(fpm.num_queued_prefill, 0);
+        assert_eq!(fpm.num_queued_decode, 0);
+        assert!(fpm.wall_time_secs > 0.0);
+    }
+
+    #[test]
+    fn test_fpm_prefill_and_decode_mixed_batch() {
+        let mut core = VllmCore::new(fpm_args());
+
+        // r1: 4-token prompt, 3 output tokens
+        let r1 = Uuid::from_u128(1);
+        core.receive(DirectRequest {
+            tokens: (0..4).collect(),
+            max_output_tokens: 3,
+            uuid: Some(r1),
+            dp_rank: 0,
+            arrival_timestamp_ms: None,
+        });
+
+        let mut collector = crate::replay::TraceCollector::default();
+
+        // Pass 1: prefill r1 (4 tokens) + first decode token
+        let pass1 = core.execute_pass(&mut collector, 0.0);
+        let fpm1 = pass1.fpm.expect("FPM should be present");
+        assert_eq!(fpm1.num_prefill_requests, 1);
+        assert_eq!(fpm1.sum_prefill_tokens, 4);
+
+        // r2: 4-token prompt arriving while r1 is decoding
+        let r2 = Uuid::from_u128(2);
+        core.receive(DirectRequest {
+            tokens: (100..104).collect(),
+            max_output_tokens: 3,
+            uuid: Some(r2),
+            dp_rank: 0,
+            arrival_timestamp_ms: None,
+        });
+
+        // Pass 2: r1 decode + r2 prefill (mixed batch)
+        let pass2 = core.execute_pass(&mut collector, 1.0);
+        let fpm2 = pass2.fpm.expect("FPM should be present");
+        assert_eq!(fpm2.num_prefill_requests, 1, "r2 is prefilling");
+        assert_eq!(fpm2.num_decode_requests, 1, "r1 is decoding");
+        assert_eq!(fpm2.sum_prefill_tokens, 4);
+        assert!(
+            fpm2.sum_decode_kv_tokens > 0,
+            "decode request should have KV context"
+        );
+    }
+
+    #[test]
+    fn test_fpm_completed_requests_metrics_correct() {
+        // This tests the fix: completed requests should still contribute
+        // correct metrics even though they're removed from state before
+        // compute_fpm runs.
+        let mut core = VllmCore::new(fpm_args());
+
+        // Request with 4-token prompt and 1 output token — completes in 1 pass
+        let r1 = Uuid::from_u128(1);
+        core.receive(DirectRequest {
+            tokens: (0..4).collect(),
+            max_output_tokens: 1,
+            uuid: Some(r1),
+            dp_rank: 0,
+            arrival_timestamp_ms: None,
+        });
+
+        let mut collector = crate::replay::TraceCollector::default();
+        let pass = core.execute_pass(&mut collector, 0.0);
+        let fpm = pass.fpm.expect("FPM should be present");
+
+        // r1 completes in this pass. The bug was that prompt_len would be 0
+        // because the request was removed from state before compute_fpm ran.
+        assert_eq!(fpm.num_prefill_requests, 1);
+        assert_eq!(fpm.sum_prefill_tokens, 4);
+        // var_prefill_length should reflect the actual prompt length (4), not 0.
+        // With a single request, variance is 0 regardless, so check sum_prefill_tokens
+        // as the main indicator.
+        assert!(pass.completed_requests > 0, "request should have completed");
+    }
+
+    #[test]
+    fn test_fpm_completed_decode_request_has_kv_context() {
+        // Decode request that completes — its KV context should be captured
+        // correctly even though it's removed from state.
+        let args = MockEngineArgs::builder()
+            .block_size(4)
+            .num_gpu_blocks(16)
+            .max_num_batched_tokens(Some(16))
+            .max_num_seqs(Some(4))
+            .enable_chunked_prefill(true)
+            .enable_prefix_caching(false)
+            .speedup_ratio(0.0)
+            .build()
+            .unwrap();
+        let mut core = VllmCore::new(args);
+
+        let r1 = Uuid::from_u128(1);
+        core.receive(DirectRequest {
+            tokens: (0..4).collect(),
+            max_output_tokens: 2,
+            uuid: Some(r1),
+            dp_rank: 0,
+            arrival_timestamp_ms: None,
+        });
+
+        let mut collector = crate::replay::TraceCollector::default();
+
+        // Pass 1: prefill + first decode token
+        core.execute_pass(&mut collector, 0.0);
+
+        // Pass 2: second decode token (completes the request)
+        let pass2 = core.execute_pass(&mut collector, 1.0);
+        let fpm2 = pass2.fpm.expect("FPM should be present");
+
+        assert_eq!(fpm2.num_decode_requests, 1);
+        // The completed decode request should have contributed its KV context
+        // (prompt_len + generated_so_far at schedule time).
+        assert!(
+            fpm2.sum_decode_kv_tokens > 0,
+            "completed decode request should still contribute KV context, got {}",
+            fpm2.sum_decode_kv_tokens
+        );
+    }
+
+    #[test]
+    fn test_fpm_queued_requests() {
+        let args = MockEngineArgs::builder()
+            .block_size(4)
+            .num_gpu_blocks(4) // Very limited KV — only room for one request
+            .max_num_batched_tokens(Some(8))
+            .max_num_seqs(Some(2))
+            .enable_chunked_prefill(true)
+            .enable_prefix_caching(false)
+            .speedup_ratio(0.0)
+            .build()
+            .unwrap();
+        let mut core = VllmCore::new(args);
+
+        // r1 and r2 both have 8-token prompts but only 4 blocks available
+        let r1 = Uuid::from_u128(1);
+        let r2 = Uuid::from_u128(2);
+        core.receive(DirectRequest {
+            tokens: (0..8).collect(),
+            max_output_tokens: 1,
+            uuid: Some(r1),
+            dp_rank: 0,
+            arrival_timestamp_ms: None,
+        });
+        core.receive(DirectRequest {
+            tokens: (100..108).collect(),
+            max_output_tokens: 1,
+            uuid: Some(r2),
+            dp_rank: 0,
+            arrival_timestamp_ms: None,
+        });
+
+        let mut collector = crate::replay::TraceCollector::default();
+        let pass = core.execute_pass(&mut collector, 0.0);
+        let fpm = pass.fpm.expect("FPM should be present");
+
+        // At least one request should be scheduled, the other might be queued
+        // (depending on KV capacity). Some requests may have completed and
+        // been removed from both scheduled and queued.
+        let total_scheduled = fpm.num_prefill_requests + fpm.num_decode_requests;
+        assert!(
+            total_scheduled >= 1,
+            "at least one request should be scheduled"
+        );
+    }
+
+    #[test]
+    fn test_fpm_var_prefill_length_with_multiple_requests() {
+        let args = MockEngineArgs::builder()
+            .block_size(4)
+            .num_gpu_blocks(32)
+            .max_num_batched_tokens(Some(32))
+            .max_num_seqs(Some(4))
+            .enable_chunked_prefill(true)
+            .enable_prefix_caching(false)
+            .speedup_ratio(0.0)
+            .build()
+            .unwrap();
+        let mut core = VllmCore::new(args);
+
+        // Two prefill requests with different prompt lengths
+        core.receive(DirectRequest {
+            tokens: (0..4).collect(), // prompt_len = 4
+            max_output_tokens: 1,
+            uuid: Some(Uuid::from_u128(1)),
+            dp_rank: 0,
+            arrival_timestamp_ms: None,
+        });
+        core.receive(DirectRequest {
+            tokens: (100..112).collect(), // prompt_len = 12
+            max_output_tokens: 1,
+            uuid: Some(Uuid::from_u128(2)),
+            dp_rank: 0,
+            arrival_timestamp_ms: None,
+        });
+
+        let mut collector = crate::replay::TraceCollector::default();
+        let pass = core.execute_pass(&mut collector, 0.0);
+        let fpm = pass.fpm.expect("FPM should be present");
+
+        assert_eq!(fpm.num_prefill_requests, 2);
+        // Population variance of [4, 12]: mean=8, var=((4-8)^2+(12-8)^2)/2 = 16
+        assert!(
+            (fpm.var_prefill_length - 16.0).abs() < 1e-6,
+            "expected var=16.0, got {}",
+            fpm.var_prefill_length
+        );
+    }
+
+    #[test]
+    fn test_fpm_chunked_prefill_reports_chunk_not_full_prompt() {
+        // With max_num_batched_tokens=8 and a 16-token prompt, chunked prefill
+        // should split across two passes. Each pass should report only the
+        // chunk size in sum_prefill_tokens, not the full prompt length.
+        let args = MockEngineArgs::builder()
+            .block_size(4)
+            .num_gpu_blocks(16)
+            .max_num_batched_tokens(Some(8))
+            .max_num_seqs(Some(4))
+            .enable_chunked_prefill(true)
+            .enable_prefix_caching(false)
+            .speedup_ratio(0.0)
+            .build()
+            .unwrap();
+        let mut core = VllmCore::new(args);
+
+        core.receive(DirectRequest {
+            tokens: (0..16).collect(),
+            max_output_tokens: 2,
+            uuid: Some(Uuid::from_u128(1)),
+            dp_rank: 0,
+            arrival_timestamp_ms: None,
+        });
+
+        let mut collector = crate::replay::TraceCollector::default();
+
+        // Pass 1: first chunk
+        let pass1 = core.execute_pass(&mut collector, 0.0);
+        let fpm1 = pass1.fpm.expect("FPM should be present");
+        assert_eq!(fpm1.num_prefill_requests, 1);
+        assert!(
+            fpm1.sum_prefill_tokens <= 8,
+            "chunk should be at most 8 tokens, got {}",
+            fpm1.sum_prefill_tokens
+        );
+        assert!(fpm1.sum_prefill_tokens > 0);
+
+        // Pass 2: remaining chunk
+        let pass2 = core.execute_pass(&mut collector, 1.0);
+        let fpm2 = pass2.fpm.expect("FPM should be present");
+        assert_eq!(fpm2.num_prefill_requests, 1, "still prefilling");
+        assert!(
+            fpm2.sum_prefill_tokens <= 8,
+            "second chunk should also be at most 8 tokens, got {}",
+            fpm2.sum_prefill_tokens
+        );
+
+        // Total across both chunks should equal the full prompt length
+        assert_eq!(
+            fpm1.sum_prefill_tokens + fpm2.sum_prefill_tokens,
+            16,
+            "total prefill tokens across chunks should equal full prompt"
+        );
+
+        // Variance should be over the full prompt length (16) in both passes
+        assert_eq!(
+            fpm1.var_prefill_length, 0.0,
+            "single request → zero variance"
+        );
+        assert_eq!(
+            fpm2.var_prefill_length, 0.0,
+            "single request → zero variance"
+        );
+    }
+
+    #[test]
+    fn test_fpm_preemption_creates_queued_decode() {
+        // Trigger preemption: fill KV with running requests, then submit a new
+        // one that forces eviction. The preempted request should appear as a
+        // queued decode in FPM.
+        let args = MockEngineArgs::builder()
+            .block_size(4)
+            .num_gpu_blocks(6) // 24 tokens of KV — very tight
+            .max_num_batched_tokens(Some(32))
+            .max_num_seqs(Some(3))
+            .enable_chunked_prefill(true)
+            .enable_prefix_caching(false)
+            .preemption_mode(PreemptionMode::Lifo)
+            .speedup_ratio(0.0)
+            .build()
+            .unwrap();
+        let mut core = VllmCore::new(args);
+        let mut collector = crate::replay::TraceCollector::default();
+
+        // r1: 4-token prompt, long output (stays running)
+        core.receive(DirectRequest {
+            tokens: (0..4).collect(),
+            max_output_tokens: 20,
+            uuid: Some(Uuid::from_u128(1)),
+            dp_rank: 0,
+            arrival_timestamp_ms: None,
+        });
+
+        // Prefill r1 and decode a few tokens to build up KV
+        core.execute_pass(&mut collector, 0.0);
+        core.execute_pass(&mut collector, 1.0);
+        core.execute_pass(&mut collector, 2.0);
+
+        // r2: another request that will compete for KV
+        core.receive(DirectRequest {
+            tokens: (100..116).collect(), // 16 tokens — will pressure KV
+            max_output_tokens: 5,
+            uuid: Some(Uuid::from_u128(2)),
+            dp_rank: 0,
+            arrival_timestamp_ms: None,
+        });
+
+        // This pass should trigger preemption
+        let pass = core.execute_pass(&mut collector, 3.0);
+        let fpm = pass.fpm.expect("FPM should be present");
+
+        // We should see at least one queued decode (preempted request) OR one
+        // queued prefill (if the new request couldn't be scheduled). The key
+        // assertion is that queued metrics are non-zero when KV pressure exists.
+        let total_queued = fpm.num_queued_prefill + fpm.num_queued_decode;
+        if total_queued > 0 {
+            // Preemption occurred — verify the preempted decode has KV context
+            if fpm.num_queued_decode > 0 {
+                assert!(
+                    fpm.sum_queued_decode_kv_tokens > 0,
+                    "preempted decode should have KV context"
+                );
+            }
+        }
+        // Regardless, at least one request should be scheduled
+        let total_scheduled = fpm.num_prefill_requests + fpm.num_decode_requests;
+        assert!(total_scheduled >= 1);
+    }
+
+    #[tokio::test]
+    async fn test_fpm_sent_through_sink() {
+        use crate::scheduler::test_utils::CapturingFpmSink;
+
+        let args = MockEngineArgs::builder()
+            .block_size(4)
+            .num_gpu_blocks(16)
+            .max_num_batched_tokens(Some(16))
+            .max_num_seqs(Some(4))
+            .enable_chunked_prefill(true)
+            .enable_prefix_caching(false)
+            .speedup_ratio(0.0)
+            .build()
+            .unwrap();
+
+        let (output_tx, mut output_rx) = mpsc::unbounded_channel::<Vec<OutputSignal>>();
+        let fpm_sink = Arc::new(CapturingFpmSink::default());
+        let fpm_publisher = crate::common::protocols::FpmPublisher::new(Some(
+            fpm_sink.clone() as Arc<dyn crate::common::protocols::FpmSink>
+        ));
+
+        let scheduler = Scheduler::new(
+            args,
+            0,
+            Some(output_tx),
+            KvEventPublishers::default(),
+            None,
+            fpm_publisher,
+        );
+
+        scheduler.receive(DirectRequest {
+            tokens: (0..8).collect(),
+            max_output_tokens: 2,
+            uuid: Some(Uuid::from_u128(1)),
+            dp_rank: 0,
+            arrival_timestamp_ms: None,
+        });
+
+        // Wait for at least one output signal — ensures the scheduler has
+        // completed at least one pass and drained the deferred FPM buffer.
+        tokio::time::timeout(Duration::from_secs(5), output_rx.recv())
+            .await
+            .expect("timed out waiting for output")
+            .expect("output channel closed");
+
+        let snapshots = fpm_sink.take();
+        assert!(
+            !snapshots.is_empty(),
+            "should have received at least one FPM snapshot"
+        );
+        let fpm = &snapshots[0];
+        assert_eq!(fpm.num_prefill_requests, 1);
+        assert!(fpm.sum_prefill_tokens > 0);
+        assert!(fpm.wall_time_secs > 0.0);
+    }
+}