[PD+DP] Allow PrefillDelayer in disaggregated-prefill mode (#23588)

python/sglang/srt/managers/prefill_delayer.py

@@ -70,9 +70,6 @@ class PrefillDelayer:
         self._curr_state: Optional[_State] = None
         self.skip_first_delayer = True
 
-        assert (
-            server_args.disaggregation_mode == "null"
-        ), "To use PrefillDelayer, disaggregation_mode must be null."
         assert (
             not server_args.disable_overlap_schedule
         ), "To use PrefillDelayer, disable_overlap_schedule must be False."

python/sglang/srt/managers/scheduler.py

@@ -982,22 +982,28 @@ class Scheduler(
         self.prefill_delayer: Optional[PrefillDelayer] = None
         self.max_prefill_bs: int = 0
         if self.server_args.enable_prefill_delayer:
-            self.prefill_delayer = PrefillDelayer(
-                dp_size=self.dp_size,
-                attn_tp_size=self.attn_tp_size,
-                cpu_group=self.tp_cpu_group,
-                server_args=self.server_args,
-                metrics_collector=(
-                    self.metrics_collector if self.enable_metrics else None
-                ),
-                max_delay_passes=self.server_args.prefill_delayer_max_delay_passes,
-                token_usage_low_watermark=self.server_args.prefill_delayer_token_usage_low_watermark,
-                device=(
-                    self.tp_group.device
-                    if self.server_args.disable_overlap_schedule
-                    else "cpu"
-                ),
-            )
+            if self.server_args.disaggregation_mode == "decode":
+                logger.info(
+                    "Ignoring --enable-prefill-delayer on decode engine "
+                    "(no prefill scheduling path; delayer would be a no-op)."
+                )
+            else:
+                self.prefill_delayer = PrefillDelayer(
+                    dp_size=self.dp_size,
+                    attn_tp_size=self.attn_tp_size,
+                    cpu_group=self.tp_cpu_group,
+                    server_args=self.server_args,
+                    metrics_collector=(
+                        self.metrics_collector if self.enable_metrics else None
+                    ),
+                    max_delay_passes=self.server_args.prefill_delayer_max_delay_passes,
+                    token_usage_low_watermark=self.server_args.prefill_delayer_token_usage_low_watermark,
+                    device=(
+                        self.tp_group.device
+                        if self.server_args.disable_overlap_schedule
+                        else "cpu"
+                    ),
+                )
 
         # NOTE: preemption is enabled by default for priority scheduling.
         self.enable_priority_preemption = (