Refactor device timer installation and rename prefill prealloc to bootstrap (#24341)

python/sglang/srt/managers/io_struct.py

@@ -1947,8 +1947,8 @@ class DisaggregationMetrics:
     """PD disaggregation metrics."""
 
     mode: str  # "prefill", "decode", or "null" - not a metric
-    prefill_prealloc_queue_reqs: int = field(
+    prefill_bootstrap_queue_reqs: int = field(
-        default=0, metadata={"metric": ("gauge", "Prefill prealloc queue requests")}
+        default=0, metadata={"metric": ("gauge", "Prefill bootstrap queue requests")}
     )
     prefill_inflight_queue_reqs: int = field(
         default=0, metadata={"metric": ("gauge", "Prefill inflight queue requests")}

python/sglang/srt/managers/scheduler.py

@@ -431,6 +431,7 @@ class Scheduler(
 
         # Launch a model worker and draft model worker if using speculative decoding
         self.init_model_worker()
+        self.install_device_timer_on_runners()
 
         if (t := envs.SGLANG_TEST_STUCK_SCHEDULER_INIT.get()) > 0:
             time.sleep(t)
@@ -709,18 +710,6 @@ class Scheduler(
         else:
             self.model_worker = self.draft_worker
 
-        # Install device timer on model runners for fwd occupancy tracking
-        if hasattr(self, "forward_pass_device_timer"):
-            timer = self.forward_pass_device_timer
-            self.tp_worker.model_runner.device_timer = timer
-            if self.draft_worker is not None:
-                dw = getattr(self.draft_worker, "draft_worker", None)
-                if dw is not None:
-                    if hasattr(dw, "draft_runner"):
-                        dw.draft_runner.device_timer = timer
-                    for r in getattr(dw, "draft_runner_list", []):
-                        r.device_timer = timer
-
         # Get token and memory info from the model worker
         (
             self.max_total_num_tokens,

python/sglang/srt/managers/scheduler_runtime_checker_mixin.py

@@ -512,7 +512,7 @@ class SchedulerRuntimeCheckerMixin:
         )
         self.stats.num_grammar_queue_reqs = len(self.grammar_manager)
         if self.disaggregation_mode == DisaggregationMode.PREFILL:
-            self.stats.num_prefill_prealloc_queue_reqs = QueueCount.from_reqs(
+            self.stats.num_prefill_bootstrap_queue_reqs = QueueCount.from_reqs(
                 self.disagg_prefill_bootstrap_queue.queue, priority_enabled
             )
             self.stats.num_prefill_inflight_queue_reqs = QueueCount.from_reqs(

python/sglang/srt/observability/metrics_collector.py

@@ -107,7 +107,7 @@ class SchedulerStats:
     num_paused_reqs: int = 0
 
     # PD disaggregation
-    num_prefill_prealloc_queue_reqs: QueueCount = field(default_factory=QueueCount)
+    num_prefill_bootstrap_queue_reqs: QueueCount = field(default_factory=QueueCount)
     num_prefill_inflight_queue_reqs: QueueCount = field(default_factory=QueueCount)
     num_decode_prealloc_queue_reqs: QueueCount = field(default_factory=QueueCount)
     num_decode_transfer_queue_reqs: QueueCount = field(default_factory=QueueCount)
@@ -340,9 +340,9 @@ class SchedulerMetricsCollector:
         # =================================================================
         # PD disaggregation
         # =================================================================
-        self.num_prefill_prealloc_queue_reqs = Gauge(
+        self.num_prefill_bootstrap_queue_reqs = Gauge(
-            name="sglang:num_prefill_prealloc_queue_reqs",
+            name="sglang:num_prefill_bootstrap_queue_reqs",
-            documentation="The number of requests in the prefill prealloc queue.",
+            documentation="The number of requests in the prefill bootstrap queue.",
             labelnames=labels.keys(),
             multiprocess_mode="mostrecent",
         )
@@ -1080,7 +1080,8 @@ class SchedulerMetricsCollector:
 
         # PD disaggregation
         self._log_gauge_queue_count(
-            self.num_prefill_prealloc_queue_reqs, stats.num_prefill_prealloc_queue_reqs
+            self.num_prefill_bootstrap_queue_reqs,
+            stats.num_prefill_bootstrap_queue_reqs,
         )
         self._log_gauge_queue_count(
             self.num_prefill_inflight_queue_reqs, stats.num_prefill_inflight_queue_reqs

python/sglang/srt/observability/scheduler_metrics_mixin.py

@@ -178,6 +178,19 @@ class SchedulerMetricsMixin:
             enable_metrics=self.enable_metrics
         )
 
+    def install_device_timer_on_runners(self: Scheduler):
+        if not hasattr(self, "forward_pass_device_timer"):
+            return
+        timer = self.forward_pass_device_timer
+        self.tp_worker.model_runner.device_timer = timer
+        if self.draft_worker is not None:
+            dw = getattr(self.draft_worker, "draft_worker", None)
+            if dw is not None:
+                if hasattr(dw, "draft_runner"):
+                    dw.draft_runner.device_timer = timer
+                for r in getattr(dw, "draft_runner_list", []):
+                    r.device_timer = timer
+
     def init_kv_events(self: Scheduler, kv_events_config: Optional[str]):
         self.enable_kv_cache_events = bool(
             kv_events_config and self.attn_tp_rank == 0 and self.attn_cp_rank == 0
@@ -373,7 +386,7 @@ class SchedulerMetricsMixin:
         )
 
         if self.disaggregation_mode == DisaggregationMode.PREFILL:
-            msg += f"#prealloc-req: {len(self.disagg_prefill_bootstrap_queue.queue)}, "
+            msg += f"#bootstrap-req: {len(self.disagg_prefill_bootstrap_queue.queue)}, "
             msg += f"#inflight-req: {len(self.disagg_prefill_inflight_queue)}, "
 
         if (
@@ -438,7 +451,7 @@ class SchedulerMetricsMixin:
 
             # PD disaggregation
             if self.disaggregation_mode == DisaggregationMode.PREFILL:
-                self.stats.num_prefill_prealloc_queue_reqs = QueueCount.from_reqs(
+                self.stats.num_prefill_bootstrap_queue_reqs = QueueCount.from_reqs(
                     self.disagg_prefill_bootstrap_queue.queue, priority_enabled
                 )
                 self.stats.num_prefill_inflight_queue_reqs = QueueCount.from_reqs(
@@ -615,7 +628,7 @@ class SchedulerMetricsMixin:
 
             # PD disaggregation
             if self.disaggregation_mode == DisaggregationMode.PREFILL:
-                self.stats.num_prefill_prealloc_queue_reqs = QueueCount.from_reqs(
+                self.stats.num_prefill_bootstrap_queue_reqs = QueueCount.from_reqs(
                     self.disagg_prefill_bootstrap_queue.queue, priority_enabled
                 )
                 self.stats.num_prefill_inflight_queue_reqs = QueueCount.from_reqs(
@@ -861,7 +874,7 @@ class SchedulerMetricsMixin:
         disaggregation = None
         if include_all or "disagg" in include:
             mode_str = "null"
-            prefill_prealloc = 0
+            prefill_bootstrap = 0
             prefill_inflight = 0
             decode_prealloc = 0
             decode_transfer = 0
@@ -869,7 +882,7 @@ class SchedulerMetricsMixin:
 
             if self.disaggregation_mode == DisaggregationMode.PREFILL:
                 mode_str = "prefill"
-                prefill_prealloc = len(self.disagg_prefill_bootstrap_queue.queue)
+                prefill_bootstrap = len(self.disagg_prefill_bootstrap_queue.queue)
                 prefill_inflight = len(self.disagg_prefill_inflight_queue)
             elif self.disaggregation_mode == DisaggregationMode.DECODE:
                 mode_str = "decode"
@@ -881,7 +894,7 @@ class SchedulerMetricsMixin:
 
             disaggregation = DisaggregationMetrics(
                 mode=mode_str,
-                prefill_prealloc_queue_reqs=prefill_prealloc,
+                prefill_bootstrap_queue_reqs=prefill_bootstrap,
                 prefill_inflight_queue_reqs=prefill_inflight,
                 decode_prealloc_queue_reqs=decode_prealloc,
                 decode_transfer_queue_reqs=decode_transfer,