fix: expose trigger_completion_at_end through unified API

flashinfer/comm/allreduce.py

@@ -453,6 +453,7 @@ def allreduce_fusion(
     workspace: AllReduceFusionWorkspace,
     pattern: int,
     launch_with_pdl: bool = False,
+    trigger_completion_at_end: Optional[bool] = None,
     # ===== OUTPUT tensors (pre-allocated, will be filled) =====
     output: Optional[torch.Tensor] = None,
     residual_out: Optional[torch.Tensor] = None,
@@ -494,7 +495,13 @@ def allreduce_fusion(
                  - kARResidualRMSNormOutFP8Quant = 4
                  - kARResidualRMSNormOutFP4Quant = 5
                  Note: MNNVL only supports patterns 0 and 1
-        launch_with_pdl: Use Persistent Dependency Launch
+        launch_with_pdl: Use Programmatic Dependent Launch
+        trigger_completion_at_end: Controls when PDL completion is signaled.
+                     None (default): follows launch_with_pdl for backward compatibility.
+                     True: signal completion after the kernel finishes (safe, no overlap).
+                     False: signal completion early, allowing the next PDL-aware kernel
+                     to overlap with this one. Only safe when the subsequent kernel also
+                     uses cudaGridDependencySynchronize().
 
         # ===== OUTPUT tensors (pre-allocated, filled by function) =====
         output: AllReduce output [token_num, hidden_dim]
@@ -611,7 +618,9 @@ def _flatten_checked(t, name):
             hidden_dim=hidden_dim,
             workspace_ptrs=workspace.workspace_tensor,
             launch_with_pdl=launch_with_pdl,
-            trigger_completion_at_end=launch_with_pdl,  # Same meaning
+            trigger_completion_at_end=trigger_completion_at_end
+            if trigger_completion_at_end is not None
+            else launch_with_pdl,
             fp32_acc=fp32_acc,
             pattern_code=pattern,  # type: ignore[arg-type]
             use_oneshot=use_oneshot,