[PrefillDelayer] support NCCL all-gather for cross-DP info sync (#24768)

Co-authored-by: Byron Hsu <byron@periodiclabs.ai>
Co-authored-by: Cursor <cursoragent@cursor.com>

python/sglang/srt/managers/prefill_delayer.py

@@ -6,6 +6,7 @@ from typing import TYPE_CHECKING, NamedTuple, Optional
 
 import torch
 
+from sglang.srt.environ import envs
 from sglang.srt.utils import get_bool_env_var
 
 if TYPE_CHECKING:
@@ -45,6 +46,7 @@ class PrefillDelayer:
         token_usage_low_watermark: Optional[float],
         metrics_collector: Optional["SchedulerMetricsCollector"] = None,
         device: Optional["torch.device"] = "cpu",
+        device_group=None,
     ):
         self._max_delay_passes = max_delay_passes
         self._token_usage_low_watermark = token_usage_low_watermark
@@ -68,15 +70,32 @@ class PrefillDelayer:
         self.dp_size = dp_size
         self.enable_dp_attention = server_args.enable_dp_attention
         dp_size_dim = dp_size if self.enable_dp_attention else 1
+
+        # Mirror scheduler_dp_attn_mixin's NCCL all-gather path: when the
+        # env flag is on (or overlap scheduling is disabled), ride the NCCL
+        # device group on `device` instead of gloo on CPU.
+        use_nccl = (
+            server_args.disable_overlap_schedule
+            or envs.SGLANG_NCCL_ALL_GATHER_IN_OVERLAP_SCHEDULER_SYNC_BATCH.get()
+        )
+        if use_nccl:
+            assert (
+                device_group is not None
+            ), "device_group is required when using NCCL for PrefillDelayer all-gather"
+            self._gather_group = device_group
+            self._gather_device = device
+        else:
+            self._gather_group = cpu_group
+            self._gather_device = "cpu"
+
         # Fields packed per rank into the all-gather tensor: prefillable,
         # token_watermark_force_allow, running_batch, max_prefill_bs,
         # waiting_queue_len.
         self._global_info_buffer = torch.empty(
             (dp_size_dim, attn_tp_size, 5),
             dtype=torch.int64,
-            device=device,
+            device=self._gather_device,
         )
-        self._cpu_group = cpu_group
 
         self._metrics_collector = metrics_collector
 
@@ -277,13 +296,13 @@ class PrefillDelayer:
                 max_prefill_bs,
                 waiting_queue_len,
             ],
-            device="cpu",
+            device=self._gather_device,
             dtype=torch.int64,
         )
         torch.distributed.all_gather_into_tensor(
             self._global_info_buffer.flatten(),
             local_info,
-            group=self._cpu_group,
+            group=self._gather_group,
         )
         tp0_info = self._global_info_buffer[:, 0, :]
         return tp0_info

python/sglang/srt/managers/scheduler.py

@@ -1119,17 +1119,14 @@ class Scheduler(
                     dp_size=self.dp_size,
                     attn_tp_size=self.attn_tp_size,
                     cpu_group=self.tp_cpu_group,
+                    device_group=self.tp_group.device_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=(
+                    device=self.tp_group.device,
-                        self.tp_group.device
-                        if self.server_args.disable_overlap_schedule
-                        else "cpu"
-                    ),
                 )
 
         # NOTE: preemption is enabled by default for priority scheduling.