[MoE] Enable Shared/Routed Overlap For Latent MoE (Nemotron-H) (#32790)

Signed-off-by: dafrimi <dafrimi@nvidia.com>

tests/kernels/moe/test_shared_fused_moe_routed_transform.py

@@ -0,0 +1,162 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Tests for SharedFusedMoE with routed_input_transform.
+
+Verifies that applying routed_input_transform inside SharedFusedMoE
+produces the same results as applying the transform manually outside.
+"""
+
+import pytest
+import torch
+import torch.nn as nn
+
+from vllm.config import VllmConfig, set_current_vllm_config
+from vllm.forward_context import set_forward_context
+from vllm.model_executor.layers.fused_moe.shared_fused_moe import SharedFusedMoE
+
+
+class SimpleLinear(nn.Module):
+    """A simple linear transform mimicking latent projection in latent MoE."""
+
+    def __init__(self, in_features: int, out_features: int, dtype: torch.dtype):
+        super().__init__()
+        self.weight = nn.Parameter(
+            torch.randn(out_features, in_features, device="cuda", dtype=dtype) / 10
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return nn.functional.linear(x, self.weight)
+
+
+class SimpleSharedExperts(nn.Module):
+    """A simple 2-layer MLP mimicking shared experts."""
+
+    def __init__(self, hidden_size: int, intermediate_size: int, dtype: torch.dtype):
+        super().__init__()
+        self.up = nn.Linear(
+            hidden_size, intermediate_size * 2, bias=False, device="cuda", dtype=dtype
+        )
+        self.down = nn.Linear(
+            intermediate_size, hidden_size, bias=False, device="cuda", dtype=dtype
+        )
+        with torch.no_grad():
+            self.up.weight.div_(10)
+            self.down.weight.div_(10)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        gate_up = self.up(x)
+        gate, up = gate_up.chunk(2, dim=-1)
+        return self.down(nn.functional.silu(gate) * up)
+
+
+@pytest.fixture(autouse=True)
+def setup_cuda():
+    if not torch.cuda.is_available():
+        pytest.skip("CUDA not available")
+    torch.set_default_device("cuda")
+
+
+@pytest.mark.parametrize("num_tokens", [1, 32])
+@pytest.mark.parametrize("hidden_size,latent_size", [(256, 128), (128, 64)])
+@pytest.mark.parametrize("dtype", [torch.bfloat16])
+def test_routed_input_transform_inside_vs_outside(
+    num_tokens: int,
+    hidden_size: int,
+    latent_size: int,
+    dtype: torch.dtype,
+    dist_init,
+    workspace_init,
+):
+    """Compare SharedFusedMoE with transform inside vs manually applying outside.
+    Method A (inside): SharedFusedMoE with routed_input_transform
+    Method B (outside): Manually transform, then SharedFusedMoE without transform
+    """
+    torch.manual_seed(42)
+
+    num_experts = 8
+    top_k = 2
+    intermediate_size = hidden_size * 2
+
+    vllm_config = VllmConfig()
+    vllm_config.compilation_config.static_forward_context = dict()
+
+    shared_experts = SimpleSharedExperts(hidden_size, intermediate_size, dtype)
+    routed_transform = SimpleLinear(hidden_size, latent_size, dtype)
+
+    with set_current_vllm_config(vllm_config):
+        # Method A: SharedFusedMoE WITH routed_input_transform
+        moe_with_transform = SharedFusedMoE(
+            shared_experts=shared_experts,
+            routed_input_transform=routed_transform,
+            num_experts=num_experts,
+            top_k=top_k,
+            hidden_size=latent_size,
+            intermediate_size=intermediate_size,
+            reduce_results=False,
+            renormalize=True,
+            params_dtype=dtype,
+            tp_size=1,
+            dp_size=1,
+            pcp_size=1,
+            prefix="moe_with_transform",
+        )
+
+        # Method B: SharedFusedMoE WITHOUT routed_input_transform
+        # Note: shared_experts=None because when transform is done outside,
+        moe_without_transform = SharedFusedMoE(
+            shared_experts=None,
+            routed_input_transform=None,
+            num_experts=num_experts,
+            top_k=top_k,
+            hidden_size=latent_size,
+            intermediate_size=intermediate_size,
+            reduce_results=False,
+            renormalize=True,
+            params_dtype=dtype,
+            tp_size=1,
+            dp_size=1,
+            pcp_size=1,
+            prefix="moe_without_transform",
+        )
+
+        with torch.no_grad():
+            moe_without_transform.w13_weight.copy_(moe_with_transform.w13_weight)
+            moe_without_transform.w2_weight.copy_(moe_with_transform.w2_weight)
+
+        moe_with_transform.quant_method.process_weights_after_loading(
+            moe_with_transform
+        )
+        moe_without_transform.quant_method.process_weights_after_loading(
+            moe_without_transform
+        )
+
+        hidden_states = torch.randn(num_tokens, hidden_size, device="cuda", dtype=dtype)
+        router_logits = torch.randn(num_tokens, num_experts, device="cuda", dtype=dtype)
+
+        with set_forward_context(None, vllm_config, num_tokens=num_tokens):
+            shared_out_A, routed_out_A = moe_with_transform(
+                hidden_states, router_logits
+            )
+
+            transformed_hidden = routed_transform(hidden_states)
+            shared_out_B, routed_out_B = moe_without_transform(
+                transformed_hidden, router_logits
+            )
+
+        torch.testing.assert_close(
+            routed_out_A,
+            routed_out_B,
+            atol=1e-3,
+            rtol=1e-3,
+            msg="Routed output should match: transform inside vs outside",
+        )
+
+        expected_shared_out = shared_experts(hidden_states)
+
+        torch.testing.assert_close(
+            shared_out_A,
+            expected_shared_out,
+            atol=1e-3,
+            rtol=1e-3,
+        )

vllm/model_executor/layers/fused_moe/layer.py

@@ -1,8 +1,8 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
-from collections.abc import Callable, Iterable
+from collections.abc import Callable, Generator, Iterable
-from contextlib import nullcontext
+from contextlib import contextmanager, nullcontext
 from enum import Enum
 from typing import Literal, cast, get_args, overload
 
@@ -351,6 +351,10 @@ class FusedMoE(CustomOp):
                     "Enabled separate cuda stream for MoE shared_experts", scope="local"
                 )
 
+        # For latent MoE: stores original hidden_states before routed_input_transform
+        # so shared_experts can use it for cloning (they need original dimension)
+        self._shared_experts_input: torch.Tensor | None = None
+
         if params_dtype is None:
             params_dtype = torch.get_default_dtype()
         self.params_dtype = params_dtype
@@ -664,6 +668,39 @@ class FusedMoE(CustomOp):
     def gate(self) -> torch.nn.Module | None:
         return None
 
+    def apply_routed_input_transform(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        """Hook to transform hidden_states before passing to routed experts.
+        For latent MoE: transforms [S, hidden_size] → [S, moe_latent_size].
+        The original hidden_states is saved in _shared_experts_input so
+        shared_experts still receive the original [S, hidden_size].
+
+        Override in subclasses (e.g., SharedFusedMoE) for latent MoE.
+        """
+        return hidden_states
+
+    @contextmanager
+    def _set_shared_experts_input(
+        self, value: torch.Tensor | None
+    ) -> Generator[None, None, None]:
+        """Context manager to safely set/clear _shared_experts_input."""
+        self._shared_experts_input = value
+        try:
+            yield
+        finally:
+            self._shared_experts_input = None
+
+    def _get_shared_experts_input(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        """Get input for shared experts.
+
+        For latent MoE: shared_experts need original [S, hidden_size],
+        not the transformed [S, latent_size] used by routed experts.
+        """
+        return (
+            self._shared_experts_input
+            if self._shared_experts_input is not None
+            else hidden_states
+        )
+
     @property
     def tp_size(self):
         return self.moe_parallel_config.tp_size
@@ -855,9 +892,11 @@ class FusedMoE(CustomOp):
         if use_shared_experts_stream:
             assert self.shared_experts_stream is not None
 
+            shared_experts_input = self._get_shared_experts_input(hidden_states)
+
             # Clone BEFORE switching streams to avoid race condition
             # where routed_expert kernel may mutate hidden_states.
-            hidden_states_clone = hidden_states.clone()
+            hidden_states_clone = shared_experts_input.clone()
 
             # Record that the clone will be used by shared_experts_stream
             # to avoid gc issue from deallocation of hidden_states_clone
@@ -1537,11 +1576,20 @@ class FusedMoE(CustomOp):
         hidden_states: torch.Tensor,
         router_logits: torch.Tensor,
     ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
-        og_hidden_states = hidden_states.shape[-1]
+        # For latent MoE: save ORIGINAL hidden_states before transform
-        if self.hidden_size != og_hidden_states:
+        # (shared_experts need original dimension, routed experts use transformed)
+        original_hidden_states = hidden_states
+        original_hidden_dim = hidden_states.shape[-1]
+
+        # Apply transform for routed experts (e.g., latent projection for latent MoE)
+        hidden_states = self.apply_routed_input_transform(hidden_states)
+
+        # This is the dimension after transform (for routed expert output slicing)
+        transformed_hidden_dim = hidden_states.shape[-1]
+        if self.hidden_size != transformed_hidden_dim:
             hidden_states = F.pad(
                 hidden_states,
-                (0, self.hidden_size - og_hidden_states),
+                (0, self.hidden_size - transformed_hidden_dim),
                 mode="constant",
                 value=0.0,
             )
@@ -1576,22 +1624,31 @@ class FusedMoE(CustomOp):
                 fused_output = torch.ops.vllm.moe_forward(
                     hidden_states, router_logits, encode_layer_name()
                 )
-            return reduce_output(fused_output)[..., :og_hidden_states]
+            return reduce_output(fused_output)[..., :transformed_hidden_dim]
         else:
             if current_platform.is_tpu() or current_platform.is_cpu():
                 # TODO: Once the OOM issue for the TPU backend is resolved, we
                 # will switch to using the moe_forward custom op.
                 # Note: CPU doesn't require wrapped forward_impl.
-                shared_output, fused_output = self.forward_impl(
+                with self._set_shared_experts_input(original_hidden_states):
-                    hidden_states, router_logits
+                    shared_output, fused_output = self.forward_impl(
-                )
+                        hidden_states, router_logits
+                    )
             else:
+                # Custom op handles setting/clearing _shared_experts_input internally
+                # We pass original tensor for shared experts (not transformed)
                 shared_output, fused_output = torch.ops.vllm.moe_forward_shared(
-                    hidden_states, router_logits, encode_layer_name()
+                    hidden_states,
+                    router_logits,
+                    encode_layer_name(),
+                    original_hidden_states,
                 )
+
+            # shared_output uses original dimension (before transform)
+            # fused_output uses transformed dimension (after transform)
             return (
-                reduce_output(shared_output)[..., :og_hidden_states],
+                reduce_output(shared_output)[..., :original_hidden_dim],
-                reduce_output(fused_output)[..., :og_hidden_states],
+                reduce_output(fused_output)[..., :transformed_hidden_dim],
             )
 
     @property
@@ -1831,7 +1888,8 @@ class FusedMoE(CustomOp):
             # because matrix multiply maybe modify the hidden_states.
             if has_separate_shared_experts and not use_shared_experts_stream:
                 assert self.shared_experts is not None
-                shared_output = self.shared_experts(hidden_states)
+                shared_input = self._get_shared_experts_input(hidden_states)
+                shared_output = self.shared_experts(shared_input)
 
             # NOTE: Similar with DP, PCP also needs dispatch and combine. For
             # simplicity, AgRsAll2All was added separately for PCP here. Maybe
@@ -2021,19 +2079,34 @@ def moe_forward_shared(
     hidden_states: torch.Tensor,
     router_logits: torch.Tensor,
     layer_name: str,
+    shared_experts_input: torch.Tensor | None = None,
 ) -> tuple[torch.Tensor, torch.Tensor]:
     self = get_layer_from_name(layer_name)
     assert self.shared_experts is not None
-    return self.forward_impl(hidden_states, router_logits)
+
+    # Set here because torch.compile skips forward_native() setup code
+    # and calls this op directly. forward_impl() reads from this var.
+    with self._set_shared_experts_input(shared_experts_input):
+        return self.forward_impl(hidden_states, router_logits)
 
 
 def moe_forward_shared_fake(
     hidden_states: torch.Tensor,
     router_logits: torch.Tensor,
     layer_name: str,
+    shared_experts_input: torch.Tensor | None = None,
 ) -> tuple[torch.Tensor, torch.Tensor]:
-    shared_out = torch.empty_like(hidden_states)
+    # Output shapes:
+    # - fused_out: same as hidden_states (routed experts use transformed size)
+    # - shared_out: same as shared_experts_input if provided, else same as hidden_states
+    # (For latent MoE: shared experts use original hidden_size, not latent size)
     fused_out = torch.empty_like(hidden_states)
+
+    if shared_experts_input is not None:
+        shared_out = torch.empty_like(shared_experts_input)
+    else:
+        shared_out = torch.empty_like(hidden_states)
+
     return shared_out, fused_out
 
 

vllm/model_executor/layers/fused_moe/shared_fused_moe.py

@@ -23,10 +23,12 @@ class SharedFusedMoE(FusedMoE):
         shared_experts: torch.nn.Module | None,
         gate: torch.nn.Module | None = None,
         use_overlapped: bool = True,
+        routed_input_transform: torch.nn.Module | None = None,
         **kwargs,
     ):
         super().__init__(**kwargs)
         self._shared_experts = shared_experts
+        self._routed_input_transform = routed_input_transform
 
         # Disable shared expert overlap if:
         #   - we are using eplb with non-default backend, because of correctness issues
@@ -56,6 +58,26 @@ class SharedFusedMoE(FusedMoE):
     def is_internal_router(self) -> bool:
         return self.gate is not None
 
+    def apply_routed_input_transform(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        """Apply transform for routed experts (e.g., latent projection).
+
+        This is called by FusedMoE.forward_native. The original hidden_states
+        is saved separately so shared experts get [S, hidden_size] while
+        routed experts get the transformed [S, moe_latent_size].
+
+        TODO: For latent MoE bandwidth optimization, fc2_latent_proj could be
+        moved inside SharedFusedMoE to all-reduce on the smaller latent
+        dimension.
+        """
+        if self._routed_input_transform is not None:
+            result = self._routed_input_transform(hidden_states)
+            # ReplicatedLinear returns (output, extra_bias) tuple.
+            # We only need the output tensor; extra_bias is not used here.
+            if isinstance(result, tuple):
+                return result[0]
+            return result
+        return hidden_states
+
     def forward(
         self,
         hidden_states: torch.Tensor,

vllm/model_executor/models/nemotron_h.py

@@ -188,10 +188,29 @@ class NemotronHMoE(nn.Module):
                 prefix=f"{prefix}.shared_experts",
             )
 
+        if self.use_latent_moe:
+            self.fc1_latent_proj = ReplicatedLinear(
+                input_size=config.hidden_size,
+                output_size=self.moe_hidden_size,
+                bias=config.mlp_bias,
+                quant_config=quant_config,
+                disable_tp=self.is_sequence_parallel,
+                prefix=f"{prefix}.fc1_latent_proj",
+            )
+            self.fc2_latent_proj = ReplicatedLinear(
+                input_size=self.moe_hidden_size,
+                output_size=config.hidden_size,
+                bias=config.mlp_bias,
+                quant_config=quant_config,
+                disable_tp=self.is_sequence_parallel,
+                prefix=f"{prefix}.fc2_latent_proj",
+            )
+        else:
+            self.fc1_latent_proj = None
+            self.fc2_latent_proj = None
+
         self.experts = SharedFusedMoE(
-            # TODO: make it possible for shared experts to have
+            shared_experts=self.shared_experts,
-            # different input in SharedFusedMoE
-            shared_experts=self.shared_experts if not self.use_latent_moe else None,
             num_experts=config.n_routed_experts,
             top_k=config.num_experts_per_tok,
             hidden_size=self.moe_hidden_size,
@@ -211,30 +230,9 @@ class NemotronHMoE(nn.Module):
             num_redundant_experts=self.n_redundant_experts,
             is_sequence_parallel=self.is_sequence_parallel,
             router_logits_dtype=router_logits_dtype,
+            routed_input_transform=self.fc1_latent_proj,
         )
 
-        if self.use_latent_moe:
-            self.fc1_latent_proj = ReplicatedLinear(
-                input_size=config.hidden_size,
-                output_size=self.moe_hidden_size,
-                bias=config.mlp_bias,
-                quant_config=quant_config,
-                disable_tp=self.is_sequence_parallel,
-                prefix=f"{prefix}.fc1_latent_proj",
-            )
-            self.fc2_latent_proj = ReplicatedLinear(
-                input_size=self.moe_hidden_size,
-                output_size=config.hidden_size,
-                bias=config.mlp_bias,
-                quant_config=quant_config,
-                disable_tp=self.is_sequence_parallel,
-                prefix=f"{prefix}.fc2_latent_proj",
-            )
-
-        else:
-            self.fc1_latent_proj = None
-            self.fc2_latent_proj = None
-
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
         num_tokens, hidden_dim = hidden_states.shape
         hidden_states = hidden_states.view(-1, hidden_dim)
@@ -244,38 +242,28 @@ class NemotronHMoE(nn.Module):
 
         # router_logits: (num_tokens, n_experts)
         router_logits, _ = self.gate(hidden_states.to(dtype=torch.float32))
-        shared_output = None
-        if self.use_latent_moe:
-            if self.shared_experts is not None:
-                shared_output = self.shared_experts(hidden_states)
-            hidden_states, _ = self.fc1_latent_proj(hidden_states)
 
-        fused_moe_out = self.experts(
+        # SharedFusedMoE handles:
+        #   - shared experts (with original hidden_states)
+        #   - routed_input_transform (fc1_latent_proj) for latent MoE
+        #   - multistream parallelism between shared and routed experts
+        shared_output, final_hidden_states = self.experts(
             hidden_states=hidden_states, router_logits=router_logits
         )
 
-        if self.use_latent_moe:
-            _, final_hidden_states = fused_moe_out
-        else:
-            shared_output, final_hidden_states = fused_moe_out
-
         # Fix FP16 overflow
         # See DeepseekV2DecoderLayer for more details.
         if hidden_states.dtype != torch.float16:
             final_hidden_states *= self.routed_scaling_factor
         elif self.shared_experts is not None:
-            assert shared_output is not None
             shared_output *= 1.0 / self.routed_scaling_factor
 
-        # TODO: currently latent up_proj is done before all-reduce for simplicity.
+        # TODO: See SharedFusedMoE.apply_routed_input_transform
-        #  if and when shared experts will be part of SharedFusedMoE,
+        # for bandwidth optimization
-        #  we should do the up_proj after all-reduce,
-        #  to have the all-reduce in the smaller latent dimension.
         if self.use_latent_moe:
             final_hidden_states, _ = self.fc2_latent_proj(final_hidden_states)
 
         if self.shared_experts is not None:
-            assert shared_output is not None
             final_hidden_states += shared_output
 
         if self.is_sequence_parallel: