[NVIDIA][comm] Merge EP+MoE-TP post-experts all-reduces into one _TP reduction (#32963)

python/sglang/srt/layers/communicator.py

@@ -24,7 +24,6 @@ from sglang.srt.distributed import (
     attention_tensor_model_parallel_all_reduce,
     attention_tensor_model_parallel_quant_all_reduce,
     get_tp_group,
-    moe_tensor_model_parallel_all_reduce,
     tensor_model_parallel_all_reduce,
 )
 from sglang.srt.distributed.device_communicators.pynccl_allocator import (
@@ -60,6 +59,8 @@ from sglang.srt.layers.dp_attention import (
 )
 from sglang.srt.layers.flashinfer_comm_fusion import is_flashinfer_allreduce_unavailable
 from sglang.srt.layers.moe import (
+    can_merge_post_experts_all_reduce,
+    deferred_post_experts_all_reduce,
     get_moe_a2a_backend,
     should_use_dp_reduce_scatterv,
     should_use_flashinfer_cutlass_moe_fp4_allgather,
@@ -724,7 +725,9 @@ class LayerCommunicator:
                             )
                         )
                 else:
-                    hidden_states = moe_tensor_model_parallel_all_reduce(hidden_states)
+                    # Fusion was published but this shape can't use the kernel,
+                    # so run the deferred reduction inline.
+                    hidden_states = deferred_post_experts_all_reduce(hidden_states)
                     hidden_states, residual = self.input_layernorm(
                         hidden_states, residual
                     )
@@ -938,16 +941,16 @@ class LayerCommunicator:
         ):
             return False
 
-        # Fusing makes the next layer's residual+LN absorb the post-experts
+        # The fused residual+LN reduces over a single group. Hybrid EP+TP spans
-        # all-reduce, and that fused kernel reduces over a single group. Under
+        # two disjoint groups; post_experts_all_reduce() merges them into one
-        # hybrid EP+TP the post-experts reduction spans two disjoint groups
+        # _TP reduction when moe_dp_size == 1, which the fused kernel can absorb.
-        # (moe_expert_parallel_all_reduce over _MOE_EP, then
+        # When merging is blocked, no single group covers both, so fusion stays off.
-        # moe_tensor_model_parallel_all_reduce over _MOE_TP), and
-        # should_skip_post_experts_all_reduce() skips *both* once fusion is
-        # published -- so the fused reduce would cover only half the peers and
-        # silently return under-reduced activations.
         parallel = get_parallel()
-        if parallel.moe_ep_size > 1 and parallel.moe_tp_size > 1:
+        if (
+            parallel.moe_ep_size > 1
+            and parallel.moe_tp_size > 1
+            and not can_merge_post_experts_all_reduce()
+        ):
             return False
 
         if (

python/sglang/srt/layers/flashinfer_comm_fusion.py

@@ -647,6 +647,37 @@ def _get_workspace_manager(use_attn_tp_group: bool) -> FlashInferWorkspaceManage
     return manager
 
 
+def resolve_fusion_world_size(*, use_attn_tp_group: bool) -> int:
+    """Peer count of the fusion group. Reads sizes only -- deliberately does not
+    reach for a group coordinator, which some callers hit before one exists."""
+    from sglang.srt.layers.moe.utils import can_merge_post_experts_all_reduce
+
+    parallel = get_parallel()
+    if use_attn_tp_group:
+        return parallel.attn_tp_size
+    if can_merge_post_experts_all_reduce():
+        return parallel.tp_size
+    return parallel.moe_ep_size if parallel.moe_ep_size > 1 else parallel.moe_tp_size
+
+
+def resolve_fusion_group(*, use_attn_tp_group: bool):
+    """Return (world_size, rank, coordinator) for the fusion workspace.
+
+    Must match the group the fused residual+LN kernel reduces over; a mismatch
+    silently reduces across the wrong peers.
+    """
+    from sglang.srt.layers.moe.utils import can_merge_post_experts_all_reduce
+
+    parallel = get_parallel()
+    if use_attn_tp_group:
+        return parallel.attn_tp_size, parallel.attn_tp_rank, get_attn_tp_group()
+    if can_merge_post_experts_all_reduce():
+        return parallel.tp_size, parallel.tp_rank, get_tp_group()
+    if parallel.moe_ep_size > 1:
+        return parallel.moe_ep_size, parallel.moe_ep_rank, get_moe_ep_group()
+    return parallel.moe_tp_size, parallel.moe_tp_rank, get_moe_tp_group()
+
+
 def _sync_allreduce_unavailable_across_tp():
     """Synchronize _flashinfer_allreduce_unavailable across all TP ranks.
 
@@ -694,19 +725,9 @@ def ensure_workspace_initialized(
     if not is_flashinfer_available() or _flashinfer_comm is None:
         return False
 
-    if use_attn_tp_group:
+    world_size, rank, coordinator = resolve_fusion_group(
-        world_size = get_parallel().attn_tp_size
+        use_attn_tp_group=use_attn_tp_group
-        rank = get_parallel().attn_tp_rank
+    )
-        coordinator = get_attn_tp_group()
-    else:
-        if get_parallel().moe_ep_size > 1:
-            world_size = get_parallel().moe_ep_size
-            rank = get_parallel().moe_ep_rank
-            coordinator = get_moe_ep_group()
-        else:
-            world_size = get_parallel().moe_tp_size
-            rank = get_parallel().moe_tp_rank
-            coordinator = get_moe_tp_group()
 
     # Always pass the coordinator's groups: flashinfer >=0.6.10 reads the
     # rendezvous group from `group=...` (falling back to WORLD when None),
@@ -814,13 +835,7 @@ def flashinfer_allreduce_residual_rmsnorm(
         )
         return None, None
 
-    if use_attn_tp_group:
+    world_size = resolve_fusion_world_size(use_attn_tp_group=use_attn_tp_group)
-        world_size = get_parallel().attn_tp_size
-    else:
-        if get_parallel().moe_ep_size > 1:
-            world_size = get_parallel().moe_ep_size
-        else:
-            world_size = get_parallel().moe_tp_size
 
     if world_size <= 1:
         logger.debug("Single GPU, no need for allreduce fusion")

python/sglang/srt/layers/moe/__init__.py

@@ -3,6 +3,8 @@ from sglang.srt.layers.moe.utils import (
     DeepEPMode,
     MoeA2ABackend,
     MoeRunnerBackend,
+    can_merge_post_experts_all_reduce,
+    deferred_post_experts_all_reduce,
     get_deepep_config,
     get_deepep_mode,
     get_moe_a2a_backend,
@@ -11,6 +13,7 @@ from sglang.srt.layers.moe.utils import (
     initialize_moe_config,
     is_moe_input_scattered_across_dp_ranks,
     is_tbo_enabled,
+    post_experts_all_reduce,
     should_skip_mlp_all_reduce,
     should_skip_post_experts_all_reduce,
     should_use_dp_reduce_scatterv,
@@ -28,6 +31,9 @@ __all__ = [
     "get_moe_runner_backend",
     "get_deepep_mode",
     "should_skip_mlp_all_reduce",
+    "can_merge_post_experts_all_reduce",
+    "deferred_post_experts_all_reduce",
+    "post_experts_all_reduce",
     "should_skip_post_experts_all_reduce",
     "should_use_dp_reduce_scatterv",
     "should_use_flashinfer_cutlass_moe_fp4_allgather",

python/sglang/srt/layers/moe/utils.py

@@ -731,30 +731,9 @@ def should_skip_mlp_all_reduce() -> bool:
 
 
 def should_skip_post_experts_all_reduce(*, is_tp_path: bool) -> bool:
-    """Whether to skip the post-experts all-reduce (EP or TP) because a
+    """Whether a downstream component will fuse, replace, or absorb the post-experts all-reduce.
-    downstream component will fuse, replace, or absorb it.
 
-    Skip reasons, in order:
+    Pass ``is_tp_path=True`` for the TP all-reduce, ``False`` for the EP one.
-      - ``get_forward().fuse_mlp_allreduce``: LayerCommunicator will fuse the
-        all-reduce with the next layer's residual all-reduce.
-      - ``get_forward().mlp_reduce_scatter``: LayerCommunicator's post-attention
-        scatter will do reduce-scatter, which would double-reduce on top of
-        an all-reduce.
-      - ``should_use_dp_reduce_scatterv()``: the standard dispatcher's combine
-        path replaces the all-reduce with a reduce-scatterv.
-      - ``should_use_flashinfer_cutlass_moe_fp4_allgather()`` (TP path only):
-        the flashinfer cutlass FP4 kernel performs an all-gather that absorbs
-        the post-experts TP all-reduce. Not relevant to the EP all-reduce.
-      - ``get_moe_a2a_backend().is_flashinfer()``: the flashinfer A2A
-        dispatcher's ``MoeAlltoAll.combine`` already alltoall-reduces partial
-        MoE outputs back to the source rank, so any further EP/TP all-reduce
-        would double-count and overflow BF16. Mirrors TRTLLM's
-        ``not enable_alltoall`` gate
-        (``tensorrt_llm/_torch/modules/fused_moe/interface.py:879``).
-
-    The first two reasons come from per-layer ``ForwardFlags`` published by
-    the decoder via ``get_forward().scoped(...)``. Pass ``is_tp_path=True``
-    for the post-experts TP all-reduce, ``False`` for the EP all-reduce.
     """
     if should_skip_mlp_all_reduce():
         return True
@@ -778,6 +757,70 @@ def should_skip_post_experts_all_reduce(*, is_tp_path: bool) -> bool:
     return False
 
 
+def can_merge_post_experts_all_reduce() -> bool:
+    """Whether the EP and MoE-TP reductions can collapse into one _TP all-reduce.
+
+    True when moe_dp_size == 1: the two groups are an orthogonal decomposition
+    of _TP, so reducing over each in turn equals one _TP reduction.
+    """
+    parallel = get_parallel()
+    return (
+        parallel.moe_ep_size > 1
+        and parallel.moe_tp_size > 1
+        and parallel.moe_dp_size == 1
+    )
+
+
+def post_experts_all_reduce(hidden_states: torch.Tensor) -> torch.Tensor:
+    """Reduce the post-experts MoE output across the EP and MoE-TP groups.
+
+    When both are live and mergeable, issues one _TP all-reduce instead of two
+    sequential ones, which also restores the invariant the fused residual+LN path
+    depends on.
+    """
+    from sglang.srt.distributed.communication_op import (
+        moe_expert_parallel_all_reduce,
+        moe_tensor_model_parallel_all_reduce,
+        tensor_model_parallel_all_reduce,
+    )
+
+    parallel = get_parallel()
+    reduce_ep = parallel.moe_ep_size > 1 and not should_skip_post_experts_all_reduce(
+        is_tp_path=False
+    )
+    reduce_tp = parallel.moe_tp_size > 1 and not should_skip_post_experts_all_reduce(
+        is_tp_path=True
+    )
+
+    if reduce_ep and reduce_tp and can_merge_post_experts_all_reduce():
+        return tensor_model_parallel_all_reduce(hidden_states)
+
+    if reduce_ep:
+        hidden_states = moe_expert_parallel_all_reduce(hidden_states)
+    if reduce_tp:
+        hidden_states = moe_tensor_model_parallel_all_reduce(hidden_states)
+    return hidden_states
+
+
+def deferred_post_experts_all_reduce(hidden_states: torch.Tensor) -> torch.Tensor:
+    """Run the post-experts reduction that was deferred to allreduce fusion.
+
+    Called when the fused residual+LN kernel cannot service the shape. Reduces
+    over the same group ``resolve_fusion_group`` builds the workspace on.
+    """
+    from sglang.srt.distributed.communication_op import (
+        moe_expert_parallel_all_reduce,
+        moe_tensor_model_parallel_all_reduce,
+        tensor_model_parallel_all_reduce,
+    )
+
+    if can_merge_post_experts_all_reduce():
+        return tensor_model_parallel_all_reduce(hidden_states)
+    if get_parallel().moe_ep_size > 1:
+        return moe_expert_parallel_all_reduce(hidden_states)
+    return moe_tensor_model_parallel_all_reduce(hidden_states)
+
+
 @contextmanager
 def speculative_moe_backend_context():
     """

python/sglang/srt/models/deepseek_v2.py

@@ -51,11 +51,7 @@ from sglang.srt.configs.model_config import (
     is_deepseek_dsa,
     is_glm_moe_dsa,
 )
-from sglang.srt.distributed import (
+from sglang.srt.distributed import divide, get_pp_group
-    divide,
-    get_pp_group,
-    tensor_model_parallel_all_reduce,
-)
 from sglang.srt.environ import envs
 from sglang.srt.eplb.expert_distribution import get_global_expert_distribution_recorder
 from sglang.srt.eplb.expert_location import ModelConfigForExpertLocation
@@ -95,7 +91,7 @@ from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.moe import (
     get_moe_a2a_backend,
     get_moe_runner_backend,
-    should_skip_post_experts_all_reduce,
+    post_experts_all_reduce,
     should_use_flashinfer_cutlass_moe_fp4_allgather,
 )
 from sglang.srt.layers.moe.ep_moe.layer import get_moe_impl_class
@@ -1037,10 +1033,7 @@ class DeepseekV2MoE(nn.Module):
                 self.routed_scaling_factor,
             )
 
-        if self.tp_size > 1 and not should_skip_post_experts_all_reduce(
+        final_hidden_states = post_experts_all_reduce(final_hidden_states)
-            is_tp_path=True,
-        ):
-            final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
         # TP1 shared experts are replicated, so add them after all-reduce to
         # avoid summing the same shared output once per TP rank.
         if self._shared_expert_tp1:
@@ -1179,10 +1172,7 @@ class DeepseekV2MoE(nn.Module):
             self.routed_scaling_factor,
         )
 
-        if self.tp_size > 1 and not should_skip_post_experts_all_reduce(
+        final_hidden_states = post_experts_all_reduce(final_hidden_states)
-            is_tp_path=True,
-        ):
-            final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
         # TP1 shared experts are replicated, so add them after all-reduce to
         # avoid summing the same shared output once per TP rank.
         if shared_output is not None and self._shared_expert_tp1:
@@ -1240,10 +1230,7 @@ class DeepseekV2MoE(nn.Module):
             ),  # block_size
             True,  # is_vnni
         )
-        if self.tp_size > 1 and not should_skip_post_experts_all_reduce(
+        final_hidden_states = post_experts_all_reduce(final_hidden_states)
-            is_tp_path=True,
-        ):
-            final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
         return final_hidden_states
 
     def forward_deepep(

python/sglang/srt/models/hunyuan_v3.py

@@ -18,10 +18,6 @@ import torch
 from torch import nn
 from transformers import PretrainedConfig
 
-from sglang.srt.distributed import (
-    moe_expert_parallel_all_reduce,
-    moe_tensor_model_parallel_all_reduce,
-)
 from sglang.srt.layers.activation import SiluAndMul
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.linear import (
@@ -31,7 +27,7 @@ from sglang.srt.layers.linear import (
     RowParallelLinear,
 )
 from sglang.srt.layers.logits_processor import LogitsProcessor
-from sglang.srt.layers.moe import should_skip_post_experts_all_reduce
+from sglang.srt.layers.moe import post_experts_all_reduce
 from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoE
 from sglang.srt.layers.moe.topk import TopK
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
@@ -191,17 +187,7 @@ class HYV3MoEFused(nn.Module):
                 hidden_states=hidden_states, topk_output=topk_output
             )
 
-        if self.ep_size > 1 and not should_skip_post_experts_all_reduce(
+        final_hidden_states = post_experts_all_reduce(final_hidden_states)
-            is_tp_path=False,
-        ):
-            final_hidden_states = moe_expert_parallel_all_reduce(final_hidden_states)
-
-        if self.tp_size > 1 and not should_skip_post_experts_all_reduce(
-            is_tp_path=True,
-        ):
-            final_hidden_states = moe_tensor_model_parallel_all_reduce(
-                final_hidden_states
-            )
 
         return final_hidden_states.view(orig_shape)
 
@@ -226,17 +212,7 @@ class HYV3MoEFused(nn.Module):
         current_stream.wait_stream(self.alt_stream)
         final_hidden_states = final_hidden_states + shared_output
 
-        if self.ep_size > 1 and not should_skip_post_experts_all_reduce(
+        final_hidden_states = post_experts_all_reduce(final_hidden_states)
-            is_tp_path=False,
-        ):
-            final_hidden_states = moe_expert_parallel_all_reduce(final_hidden_states)
-
-        if self.tp_size > 1 and not should_skip_post_experts_all_reduce(
-            is_tp_path=True,
-        ):
-            final_hidden_states = moe_tensor_model_parallel_all_reduce(
-                final_hidden_states
-            )
 
         return final_hidden_states.view(orig_shape)
 

python/sglang/srt/models/qwen2_moe.py

@@ -60,6 +60,7 @@ from sglang.srt.layers.linear import (
 )
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.moe import (
+    can_merge_post_experts_all_reduce,
     get_moe_a2a_backend,
     should_skip_post_experts_all_reduce,
 )
@@ -1161,10 +1162,20 @@ class Qwen2MoeModel(nn.Module):
                 and hasattr(hidden_states, "_sglang_needs_allreduce_fusion")
                 and hidden_states._sglang_needs_allreduce_fusion
             ):
+                # The deferred reduction the next layer would have fused; no
+                # layer follows on this rank, so run it here. Unconditional --
+                # the skip flags that deferred it are what got us into this
+                # branch -- so it bypasses post_experts_all_reduce()'s guards
+                # while reusing its merge rule.
+                if can_merge_post_experts_all_reduce():
+                    hidden_states = tensor_model_parallel_all_reduce(hidden_states)
+                else:
                     if get_parallel().moe_ep_size > 1:
                         hidden_states = moe_expert_parallel_all_reduce(hidden_states)
                     if get_parallel().moe_tp_size > 1:
-                    hidden_states = moe_tensor_model_parallel_all_reduce(hidden_states)
+                        hidden_states = moe_tensor_model_parallel_all_reduce(
+                            hidden_states
+                        )
                 hidden_states._sglang_needs_allreduce_fusion = False
             return PPProxyTensors(
                 {

python/sglang/srt/models/qwen3_moe.py

@@ -28,8 +28,6 @@ from transformers import PretrainedConfig
 
 from sglang.srt.distributed import (
     get_pp_group,
-    moe_expert_parallel_all_reduce,
-    moe_tensor_model_parallel_all_reduce,
 )
 from sglang.srt.eplb.expert_distribution import get_global_expert_distribution_recorder
 from sglang.srt.eplb.expert_location import ModelConfigForExpertLocation
@@ -44,7 +42,7 @@ from sglang.srt.layers.linear import (
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.moe import (
     get_moe_a2a_backend,
-    should_skip_post_experts_all_reduce,
+    post_experts_all_reduce,
 )
 from sglang.srt.layers.moe.ep_moe.layer import get_moe_impl_class
 from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoE
@@ -335,17 +333,7 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
             topk_output = self.topk.empty_topk_output(hidden_states.device)
         final_hidden_states = self.experts(hidden_states, topk_output)
 
-        if self.ep_size > 1 and not should_skip_post_experts_all_reduce(
+        final_hidden_states = post_experts_all_reduce(final_hidden_states)
-            is_tp_path=False
-        ):
-            final_hidden_states = moe_expert_parallel_all_reduce(final_hidden_states)
-
-        if self.tp_size > 1 and not should_skip_post_experts_all_reduce(
-            is_tp_path=True
-        ):
-            final_hidden_states = moe_tensor_model_parallel_all_reduce(
-                final_hidden_states
-            )
 
         return final_hidden_states.view(num_tokens, hidden_dim)
 

test/registered/unit/layers/test_layer_communicator_fusion_gate.py

@@ -1,9 +1,18 @@
+import contextlib
 import types
 import unittest
-from unittest.mock import patch
+from unittest.mock import MagicMock, patch
+
+import torch
 
 from sglang.srt.layers import communicator as comm
 from sglang.srt.layers.communicator import LayerCommunicator, ScatterMode
+from sglang.srt.layers.moe import (
+    can_merge_post_experts_all_reduce,
+    deferred_post_experts_all_reduce,
+    post_experts_all_reduce,
+)
+from sglang.srt.layers.moe import utils as moe_utils
 from sglang.srt.runtime_context import get_parallel
 from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
@@ -20,19 +29,218 @@ def _fake_communicator(mlp_mode=ScatterMode.TP_ATTN_FULL):
     )
 
 
-class TestFuseMlpAllReduceGate(CustomTestCase):
+@contextlib.contextmanager
-    """Hybrid EP+TP must not fuse the post-experts all-reduce away.
+def _recorded_all_reduces(called, *, moe_ep_size, moe_tp_size, moe_dp_size):
+    """Log which group each all-reduce helper reduces over, under a fixed topo."""
 
-    The fused residual+LN reduces over a single group, but with moe_ep_size > 1
+    def record(name):
-    and moe_tp_size > 1 the post-experts reduction spans two disjoint groups
+        return lambda x: called.append(name) or x
-    (_MOE_EP then _MOE_TP) and should_skip_post_experts_all_reduce() drops both
+
-    once fusion is published. The result is activations reduced over only half
+    with (
-    the peers -- wrong output, no crash. Observed as garbage completions on
+        patch(
-    Qwen3-30B-A3B with --tp-size 4 --ep-size 2.
+            "sglang.srt.distributed.communication_op.tensor_model_parallel_all_reduce",
+            side_effect=record("tp"),
+        ),
+        patch(
+            "sglang.srt.distributed.communication_op.moe_expert_parallel_all_reduce",
+            side_effect=record("ep"),
+        ),
+        patch(
+            "sglang.srt.distributed.communication_op.moe_tensor_model_parallel_all_reduce",
+            side_effect=record("moe_tp"),
+        ),
+        get_parallel().override(
+            moe_ep_size=moe_ep_size,
+            moe_tp_size=moe_tp_size,
+            moe_dp_size=moe_dp_size,
+            tp_size=moe_ep_size * moe_tp_size * moe_dp_size,
+        ),
+    ):
+        yield
+
+
+class TestPostExpertsAllReduceMerge(CustomTestCase):
+    """The two post-experts reductions collapse into one _TP reduction.
+
+    _MOE_EP and _MOE_TP are orthogonal subgroups of _TP, so with
+    moe_dp_size == 1 reducing over each in turn equals one _TP reduction --
+    one collective instead of two. With moe_dp_size > 1 they cover only part of
+    _TP and merging would sum across DP replicas, which hold different tokens.
+    """
+
+    def _calls(self, *, moe_ep_size, moe_tp_size, moe_dp_size=1, skip=False):
+        called = []
+        with (
+            patch.object(
+                moe_utils, "should_skip_post_experts_all_reduce", return_value=skip
+            ),
+            _recorded_all_reduces(
+                called,
+                moe_ep_size=moe_ep_size,
+                moe_tp_size=moe_tp_size,
+                moe_dp_size=moe_dp_size,
+            ),
+        ):
+            post_experts_all_reduce(torch.zeros(2, 2))
+        return called
+
+    def test_hybrid_issues_one_tp_reduction(self):
+        self.assertEqual(self._calls(moe_ep_size=2, moe_tp_size=2), ["tp"])
+
+    def test_moe_dp_keeps_the_two_step_form(self):
+        # Server args reject moe_ep_size > 1 together with moe_tp_size > 1 and
+        # moe_dp_size > 1 (they force ep_size * moe_dp_size == tp_size), so this
+        # pins the guard rather than a topology that can be launched today.
+        self.assertEqual(
+            self._calls(moe_ep_size=2, moe_tp_size=2, moe_dp_size=2), ["ep", "moe_tp"]
+        )
+
+    def test_single_dimension_issues_one_reduction(self):
+        self.assertEqual(self._calls(moe_ep_size=1, moe_tp_size=4), ["moe_tp"])
+        self.assertEqual(self._calls(moe_ep_size=4, moe_tp_size=1), ["ep"])
+
+    def test_skipped_when_deferred_to_fusion(self):
+        self.assertEqual(self._calls(moe_ep_size=2, moe_tp_size=2, skip=True), [])
+
+
+class TestDeferredPostExpertsAllReduce(CustomTestCase):
+    """The inline fallback must reduce over the same peers the fused kernel would.
+
+    _MOE_TP holds a single rank under pure EP, so reducing over it there is a
+    no-op that drops the deferred reduction instead of performing it.
+    """
+
+    def _calls(self, *, moe_ep_size, moe_tp_size, moe_dp_size=1):
+        called = []
+        with _recorded_all_reduces(
+            called,
+            moe_ep_size=moe_ep_size,
+            moe_tp_size=moe_tp_size,
+            moe_dp_size=moe_dp_size,
+        ):
+            deferred_post_experts_all_reduce(torch.zeros(2, 2))
+        return called
+
+    def test_hybrid_reduces_over_tp(self):
+        self.assertEqual(self._calls(moe_ep_size=2, moe_tp_size=2), ["tp"])
+
+    def test_pure_ep_reduces_over_ep(self):
+        self.assertEqual(self._calls(moe_ep_size=4, moe_tp_size=1), ["ep"])
+
+    def test_pure_tp_reduces_over_moe_tp(self):
+        self.assertEqual(self._calls(moe_ep_size=1, moe_tp_size=4), ["moe_tp"])
+
+    def test_moe_dp_reduces_over_moe_tp(self):
+        self.assertEqual(
+            self._calls(moe_ep_size=1, moe_tp_size=2, moe_dp_size=2), ["moe_tp"]
+        )
+
+
+class TestCanMergePostExpertsAllReduce(CustomTestCase):
+    def _can_merge(self, *, moe_ep_size, moe_tp_size, moe_dp_size=1):
+        with get_parallel().override(
+            moe_ep_size=moe_ep_size,
+            moe_tp_size=moe_tp_size,
+            moe_dp_size=moe_dp_size,
+            tp_size=moe_ep_size * moe_tp_size * moe_dp_size,
+        ):
+            return can_merge_post_experts_all_reduce()
+
+    def test_hybrid_ep_tp_merges(self):
+        self.assertTrue(self._can_merge(moe_ep_size=2, moe_tp_size=2))
+
+    def test_single_dimension_does_not_merge(self):
+        self.assertFalse(self._can_merge(moe_ep_size=1, moe_tp_size=4))
+        self.assertFalse(self._can_merge(moe_ep_size=4, moe_tp_size=1))
+
+
+class TestResolveFusionGroup(CustomTestCase):
+    """EP2/MoE-TP2/DP1 (e.g. DeepSeek-V4-Flash with --tp-size 4 --ep-size 2) must
+    resolve to the _TP group with world_size=4 and the TP rank."""
+
+    def _resolve(self, *, moe_ep_size, moe_tp_size, moe_dp_size=1, tp_rank=0):
+        from sglang.srt.layers.flashinfer_comm_fusion import (
+            resolve_fusion_group,
+            resolve_fusion_world_size,
+        )
+
+        fake_tp_group = MagicMock(name="tp_group")
+        fake_ep_group = MagicMock(name="ep_group")
+        fake_moe_tp_group = MagicMock(name="moe_tp_group")
+        tp_size = moe_ep_size * moe_tp_size * moe_dp_size
+
+        with (
+            get_parallel().override(
+                moe_ep_size=moe_ep_size,
+                moe_tp_size=moe_tp_size,
+                moe_dp_size=moe_dp_size,
+                tp_size=tp_size,
+                tp_rank=tp_rank,
+                moe_ep_rank=tp_rank % moe_ep_size,
+                moe_tp_rank=tp_rank % moe_tp_size,
+            ),
+            patch(
+                "sglang.srt.layers.flashinfer_comm_fusion.get_tp_group",
+                return_value=fake_tp_group,
+            ),
+            patch(
+                "sglang.srt.layers.flashinfer_comm_fusion.get_moe_ep_group",
+                return_value=fake_ep_group,
+            ),
+            patch(
+                "sglang.srt.layers.flashinfer_comm_fusion.get_moe_tp_group",
+                return_value=fake_moe_tp_group,
+            ),
+        ):
+            ws = resolve_fusion_world_size(use_attn_tp_group=False)
+            group_tuple = resolve_fusion_group(use_attn_tp_group=False)
+        return ws, group_tuple, (fake_tp_group, fake_ep_group, fake_moe_tp_group)
+
+    def test_hybrid_ep2_tp2_dp1_resolves_to_tp_ws4(self):
+        # EP2/MoE-TP2/DP1 (DeepSeek-V4-Flash on 4 GPUs): workspace must sit on
+        # _TP (ws=4) so the fused kernel reduces over all 4 peers.
+        ws, (size, rank, group), (tp_grp, ep_grp, moe_tp_grp) = self._resolve(
+            moe_ep_size=2, moe_tp_size=2, moe_dp_size=1, tp_rank=3
+        )
+        self.assertEqual(ws, 4)
+        self.assertEqual(size, 4)
+        self.assertEqual(rank, 3)
+        self.assertIs(group, tp_grp)
+
+    def test_pure_ep_resolves_to_ep_group(self):
+        ws, (size, rank, group), (tp_grp, ep_grp, moe_tp_grp) = self._resolve(
+            moe_ep_size=4, moe_tp_size=1, moe_dp_size=1, tp_rank=2
+        )
+        self.assertEqual(ws, 4)
+        self.assertEqual(size, 4)
+        self.assertIs(group, ep_grp)
+
+    def test_pure_tp_resolves_to_moe_tp_group(self):
+        ws, (size, rank, group), (tp_grp, ep_grp, moe_tp_grp) = self._resolve(
+            moe_ep_size=1, moe_tp_size=4, moe_dp_size=1, tp_rank=1
+        )
+        self.assertEqual(ws, 4)
+        self.assertEqual(size, 4)
+        self.assertIs(group, moe_tp_grp)
+
+
+class TestFuseMlpAllReduceGate(CustomTestCase):
+    """Fusion is allowed only when one group covers the whole reduction.
+
+    The fused residual+LN reduces over a single group. Hybrid EP+TP produces two
+    reductions over disjoint groups; merging collapses them to one _TP reduction
+    that the fused kernel can absorb. When merging does not apply
+    (moe_dp_size > 1) there is no such group and fusion must stay off --
+    otherwise the fused reduce covers half the peers and silently under-reduces.
     """
 
     def _should_fuse(
-        self, *, moe_ep_size, moe_tp_size, mlp_mode=ScatterMode.TP_ATTN_FULL
+        self,
+        *,
+        moe_ep_size,
+        moe_tp_size,
+        moe_dp_size=1,
+        mlp_mode=ScatterMode.TP_ATTN_FULL,
     ):
         forward_batch = types.SimpleNamespace(
             input_ids=types.SimpleNamespace(shape=(8,))
@@ -46,15 +254,24 @@ class TestFuseMlpAllReduceGate(CustomTestCase):
                 return_value=types.SimpleNamespace(input_scattered=False),
             ),
             get_parallel().override(
-                moe_ep_size=moe_ep_size, moe_tp_size=moe_tp_size, tp_size=4
+                moe_ep_size=moe_ep_size,
+                moe_tp_size=moe_tp_size,
+                moe_dp_size=moe_dp_size,
+                tp_size=moe_ep_size * moe_tp_size * moe_dp_size,
             ),
         ):
             return LayerCommunicator.should_fuse_mlp_allreduce_with_next_layer(
                 _fake_communicator(mlp_mode), forward_batch
             )
 
-    def test_hybrid_ep_tp_does_not_fuse(self):
+    def test_hybrid_ep_tp_fuses_when_mergeable(self):
-        self.assertFalse(self._should_fuse(moe_ep_size=2, moe_tp_size=2))
+        self.assertTrue(self._should_fuse(moe_ep_size=2, moe_tp_size=2))
+
+    def test_hybrid_ep_tp_does_not_fuse_when_moe_dp_blocks_the_merge(self):
+        # Same caveat as test_moe_dp_keeps_the_two_step_form: unreachable today,
+        # kept so a future relaxation cannot silently re-enable fusion over a
+        # reduction that no single group covers.
+        self.assertFalse(self._should_fuse(moe_ep_size=2, moe_tp_size=2, moe_dp_size=2))
 
     def test_pure_tp_still_fuses(self):
         self.assertTrue(self._should_fuse(moe_ep_size=1, moe_tp_size=4))