Fix silently wrong EPLB output with --moe-a2a-backend none (rank-invariant dispatch) (#32962)

python/sglang/srt/eplb/expert_location.py

@@ -560,8 +560,11 @@ def _compute_logical_to_all_physical_map(
                 physical_expert_id
             )
 
-    # Replace by the physical expert on local GPU or node if possible
-    if moe_ep_rank is not None:
+    # Replace by the physical expert on local GPU or node if possible. Skipped
+    # without an a2a backend, where all EP ranks must agree on the pick: this
+    # collapse is per-rank, and the full candidate list is what lets the dispatch
+    # spread a hot expert over its replicas. See ExpertLocationDispatchInfo.
+    if moe_ep_rank is not None and server_args.moe_a2a_backend != "none":
         num_local_gpu_physical_experts = num_physical_experts // ep_size
         prefer_same_node = _prefer_same_node_experts(server_args)
         num_gpus_per_node = (

python/sglang/srt/eplb/expert_location_dispatch.py

@@ -23,7 +23,7 @@ from sglang.srt.runtime_context import get_server_args
 
 @dataclass
 class ExpertLocationDispatchInfo:
-    ep_dispatch_algorithm: Literal["static", "random"]
+    ep_dispatch_algorithm: Literal["static", "dynamic", "fake", "lp"]
     # (num_logical_experts,)
     partial_logical_to_rank_dispatch_physical_map: Optional[torch.Tensor]
     # (num_logical_experts, X)
@@ -31,10 +31,17 @@ class ExpertLocationDispatchInfo:
     # (num_logical_experts,)
     partial_logical_to_all_physical_map_num_valid: torch.Tensor
     num_physical_experts: int
+    # Whether every rank must pick the same physical expert for a token. True
+    # without an a2a backend: all EP ranks then run the MoE over the same tokens
+    # and sum their partial outputs, so a rank-dependent pick counts a replicated
+    # logical expert several times. With one, each rank dispatches only its own
+    # tokens and is free to disagree.
+    rank_invariant: bool = False
 
     @classmethod
     def init_new(cls, layer_id: int):
-        ep_dispatch_algorithm = get_server_args().ep_dispatch_algorithm
+        server_args = get_server_args()
+        ep_dispatch_algorithm = server_args.ep_dispatch_algorithm
         expert_location_metadata = get_global_expert_location_metadata()
         assert expert_location_metadata is not None
 
@@ -43,6 +50,7 @@ class ExpertLocationDispatchInfo:
 
         return cls(
             ep_dispatch_algorithm=ep_dispatch_algorithm,
+            rank_invariant=server_args.moe_a2a_backend == "none",
             partial_logical_to_rank_dispatch_physical_map=(
                 expert_location_metadata.logical_to_rank_dispatch_physical_map[
                     layer_id, :
@@ -107,15 +115,37 @@ def _topk_ids_logical_to_physical_static(
 def _topk_ids_logical_to_physical_dynamic(
     topk_ids: torch.Tensor, info: Optional[ExpertLocationDispatchInfo]
 ) -> torch.Tensor:
+    """Spread each (token, logical expert) over that logical expert's replicas.
+
+    Under ``rank_invariant`` the replica is picked by token row rather than at
+    random: ``torch.randint`` reads the default CUDA generator, so agreement
+    across ranks would rest on their philox offsets staying aligned, which
+    nothing asserts, and it would make greedy requests non-reproducible. Both
+    pick evenly, which is the load split EPLB's placement solver assumes.
+
+    Row indexing leaves replicas unused for a single-row batch, where there is no
+    imbalance to fix anyway.
+    """
     topk_ids_original_shape = topk_ids.shape
     original_dtype = topk_ids.dtype
     device = topk_ids.device
     topk_ids = topk_ids.flatten()
 
-    chosen_dispatch_index = (
-        torch.randint(0, 65536, topk_ids.shape, dtype=torch.int32, device=device)
-        % info.partial_logical_to_all_physical_map_num_valid[topk_ids]
-    )
+    num_valid = info.partial_logical_to_all_physical_map_num_valid[topk_ids]
+    if info.rank_invariant:
+        slots_per_token = (
+            topk_ids_original_shape[-1] if len(topk_ids_original_shape) > 1 else 1
+        )
+        row_index = (
+            torch.arange(topk_ids.shape[0], dtype=num_valid.dtype, device=device)
+            // slots_per_token
+        )
+        chosen_dispatch_index = row_index % num_valid
+    else:
+        chosen_dispatch_index = (
+            torch.randint(0, 65536, topk_ids.shape, dtype=torch.int32, device=device)
+            % num_valid
+        )
     topk_ids = info.partial_logical_to_all_physical_map[topk_ids, chosen_dispatch_index]
     if topk_ids.dtype != original_dtype:
         topk_ids = topk_ids.to(original_dtype)

python/sglang/srt/server_args.py

@@ -6710,10 +6710,29 @@ class ServerArgs:
                 "EPLB is enabled. The expert_distribution_recorder_mode is automatically set."
             )
 
+        # Without an a2a backend all EP ranks run the MoE over the same tokens and
+        # sum their partial outputs, so the pick has to agree across ranks.
+        needs_rank_invariant_dispatch = self._resolved().moe_a2a_backend == "none"
+
         if (self.enable_eplb or (self.init_expert_location != "trivial")) and (
             self.ep_dispatch_algorithm is None
         ):
-            self.ep_dispatch_algorithm = "static"
+            self.ep_dispatch_algorithm = (
+                "dynamic" if needs_rank_invariant_dispatch else "static"
+            )
+
+        # `dynamic` / `fake` switch to the row-index pick; `static` reads a
+        # per-rank table and `lp` samples inside its kernel.
+        if needs_rank_invariant_dispatch and self.ep_dispatch_algorithm in (
+            "static",
+            "lp",
+        ):
+            raise ValueError(
+                f"--ep-dispatch-algorithm {self.ep_dispatch_algorithm} picks a "
+                "different physical replica per rank, which only holds up when an "
+                "a2a backend routes each token to a single rank. Use "
+                "--ep-dispatch-algorithm dynamic with --moe-a2a-backend none."
+            )
 
         if self.enable_eplb and self.ep_join_mode != "scale":
             assert self._resolved().ep_size > 1