From a776a48b1c753645c547b735ab647867c98a9b0c Mon Sep 17 00:00:00 2001
From: Jackmin801 <56836461+Jackmin801@users.noreply.github.com>
Date: Wed, 8 Apr 2026 12:23:08 -0700
Subject: [PATCH] [MoE] Move DEEP_GEMM into experts/ subdirectory (#39005)

Signed-off-by: Jackmin801 <ongjackm@gmail.com>
Signed-off-by: Robert Shaw <robshaw@redhat.com>
Co-authored-by: Robert Shaw <robshaw@redhat.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
---
 benchmarks/kernels/benchmark_silu_mul_fp8_quant.py     |  2 +-
 docs/design/moe_kernel_features.md                     |  2 +-
 tests/kernels/moe/modular_kernel_tools/mk_objects.py   |  8 ++++----
 tests/kernels/moe/test_batched_deepgemm.py             |  4 ++--
 tests/kernels/moe/test_block_fp8.py                    |  2 +-
 tests/kernels/moe/test_deepep_deepgemm_moe.py          |  6 ++++--
 tests/kernels/moe/test_deepgemm.py                     |  2 +-
 tests/kernels/moe/test_silu_mul_fp8_quant_deep_gemm.py |  2 +-
 vllm/model_executor/layers/fused_moe/__init__.py       | 10 ++++++----
 .../fused_moe/{ => experts}/batched_deep_gemm_moe.py   |  0
 .../layers/fused_moe/{ => experts}/deep_gemm_moe.py    |  0
 vllm/model_executor/layers/fused_moe/oracle/fp8.py     |  2 +-
 .../layers/fused_moe/triton_deep_gemm_moe.py           |  2 +-
 vllm/model_executor/warmup/deep_gemm_warmup.py         |  2 +-
 14 files changed, 24 insertions(+), 20 deletions(-)
 rename vllm/model_executor/layers/fused_moe/{ => experts}/batched_deep_gemm_moe.py (100%)
 rename vllm/model_executor/layers/fused_moe/{ => experts}/deep_gemm_moe.py (100%)

diff --git a/benchmarks/kernels/benchmark_silu_mul_fp8_quant.py b/benchmarks/kernels/benchmark_silu_mul_fp8_quant.py
index 13b97b769..9fcf278f2 100644
--- a/benchmarks/kernels/benchmark_silu_mul_fp8_quant.py
+++ b/benchmarks/kernels/benchmark_silu_mul_fp8_quant.py
@@ -20,7 +20,7 @@ import matplotlib.pyplot as plt
 import numpy as np
 import torch
 
-from vllm.model_executor.layers.fused_moe.batched_deep_gemm_moe import (
+from vllm.model_executor.layers.fused_moe.experts.batched_deep_gemm_moe import (
     persistent_masked_m_silu_mul_quant,
 )
 from vllm.triton_utils import tl, triton
diff --git a/docs/design/moe_kernel_features.md b/docs/design/moe_kernel_features.md
index 7cf7b76d6..c9dc1292d 100644
--- a/docs/design/moe_kernel_features.md
+++ b/docs/design/moe_kernel_features.md
@@ -82,7 +82,7 @@ To be used with a particular `FusedMoEPrepareAndFinalizeModular` subclass, MoE k
 | ------ | ----------------- | ------------ | ------------- | ------------------- | --------------------- | ------- | ------ |
 | triton | standard | all<sup>1</sup> | G,A,T | silu, gelu,</br>swigluoai,</br>silu_no_mul,</br>gelu_no_mul | Y | Y | [`fused_experts`][vllm.model_executor.layers.fused_moe.fused_moe.fused_experts],</br>[`TritonExperts`][vllm.model_executor.layers.fused_moe.fused_moe.TritonExperts] |
 | triton (batched) | batched | all<sup>1</sup> | G,A,T | silu, gelu | <sup>6</sup> | Y | [`BatchedTritonExperts`][vllm.model_executor.layers.fused_moe.fused_batched_moe.BatchedTritonExperts] |
-| deep gemm | standard,</br>batched | fp8 | G(128),A,T | silu, gelu | <sup>6</sup> | Y | </br>[`DeepGemmExperts`][vllm.model_executor.layers.fused_moe.deep_gemm_moe.DeepGemmExperts],</br>[`BatchedDeepGemmExperts`][vllm.model_executor.layers.fused_moe.batched_deep_gemm_moe.BatchedDeepGemmExperts] |
+| deep gemm | standard,</br>batched | fp8 | G(128),A,T | silu, gelu | <sup>6</sup> | Y | </br>[`DeepGemmExperts`][vllm.model_executor.layers.fused_moe.experts.deep_gemm_moe.DeepGemmExperts],</br>[`BatchedDeepGemmExperts`][vllm.model_executor.layers.fused_moe.experts.batched_deep_gemm_moe.BatchedDeepGemmExperts] |
 | cutlass_fp4 | standard,</br>batched | nvfp4 | A,T | silu | Y | Y | [`CutlassExpertsFp4`][vllm.model_executor.layers.fused_moe.cutlass_moe.CutlassExpertsFp4] |
 | cutlass_fp8 | standard,</br>batched | fp8 | A,T | silu, gelu | Y | Y | [`CutlassExpertsFp8`][vllm.model_executor.layers.fused_moe.cutlass_moe.CutlassExpertsFp8],</br>[`CutlasBatchedExpertsFp8`][vllm.model_executor.layers.fused_moe.cutlass_moe.CutlassBatchedExpertsFp8] |
 | flashinfer | standard | nvfp4,</br>fp8 | T | <sup>5</sup> | N | Y | [`FlashInferExperts`][vllm.model_executor.layers.fused_moe.flashinfer_cutlass_moe.FlashInferExperts] |
diff --git a/tests/kernels/moe/modular_kernel_tools/mk_objects.py b/tests/kernels/moe/modular_kernel_tools/mk_objects.py
index a56435379..a39e03abe 100644
--- a/tests/kernels/moe/modular_kernel_tools/mk_objects.py
+++ b/tests/kernels/moe/modular_kernel_tools/mk_objects.py
@@ -7,14 +7,14 @@ import torch
 # Fused experts and PrepareFinalize imports
 import vllm.model_executor.layers.fused_moe.modular_kernel as mk
 from vllm.model_executor.layers.fused_moe import TritonExperts
-from vllm.model_executor.layers.fused_moe.batched_deep_gemm_moe import (
-    BatchedDeepGemmExperts,
-)
 from vllm.model_executor.layers.fused_moe.config import (
     FusedMoEConfig,
     FusedMoEQuantConfig,
 )
-from vllm.model_executor.layers.fused_moe.deep_gemm_moe import DeepGemmExperts
+from vllm.model_executor.layers.fused_moe.experts.batched_deep_gemm_moe import (
+    BatchedDeepGemmExperts,
+)
+from vllm.model_executor.layers.fused_moe.experts.deep_gemm_moe import DeepGemmExperts
 from vllm.model_executor.layers.fused_moe.fused_batched_moe import (
     BatchedTritonExperts,
     NaiveBatchedExperts,
diff --git a/tests/kernels/moe/test_batched_deepgemm.py b/tests/kernels/moe/test_batched_deepgemm.py
index 20763b91d..b11098c82 100644
--- a/tests/kernels/moe/test_batched_deepgemm.py
+++ b/tests/kernels/moe/test_batched_deepgemm.py
@@ -5,10 +5,10 @@ import pytest
 import torch
 
 from vllm.model_executor.layers.fused_moe.activation import MoEActivation
-from vllm.model_executor.layers.fused_moe.batched_deep_gemm_moe import (
+from vllm.model_executor.layers.fused_moe.config import fp8_w8a8_moe_quant_config
+from vllm.model_executor.layers.fused_moe.experts.batched_deep_gemm_moe import (
     BatchedDeepGemmExperts,
 )
-from vllm.model_executor.layers.fused_moe.config import fp8_w8a8_moe_quant_config
 from vllm.model_executor.layers.fused_moe.fused_batched_moe import (
     BatchedPrepareAndFinalize,
     BatchedTritonExperts,
diff --git a/tests/kernels/moe/test_block_fp8.py b/tests/kernels/moe/test_block_fp8.py
index f27fd6f34..0181fc252 100644
--- a/tests/kernels/moe/test_block_fp8.py
+++ b/tests/kernels/moe/test_block_fp8.py
@@ -28,7 +28,7 @@ from vllm.model_executor.layers.fused_moe.all2all_utils import (
 from vllm.model_executor.layers.fused_moe.config import (
     fp8_w8a8_moe_quant_config,
 )
-from vllm.model_executor.layers.fused_moe.deep_gemm_moe import (
+from vllm.model_executor.layers.fused_moe.experts.deep_gemm_moe import (
     _valid_deep_gemm_shape,
 )
 from vllm.model_executor.layers.fused_moe.triton_deep_gemm_moe import (
diff --git a/tests/kernels/moe/test_deepep_deepgemm_moe.py b/tests/kernels/moe/test_deepep_deepgemm_moe.py
index 9dd8b13d6..6bde13e0e 100644
--- a/tests/kernels/moe/test_deepep_deepgemm_moe.py
+++ b/tests/kernels/moe/test_deepep_deepgemm_moe.py
@@ -47,10 +47,12 @@ if has_deep_ep():
     from .parallel_utils import DeepEPHTArgs, DeepEPLLArgs, make_deepep_a2a
 
 if has_deep_gemm():
-    from vllm.model_executor.layers.fused_moe.batched_deep_gemm_moe import (
+    from vllm.model_executor.layers.fused_moe.experts.batched_deep_gemm_moe import (
         BatchedDeepGemmExperts,
     )
-    from vllm.model_executor.layers.fused_moe.deep_gemm_moe import DeepGemmExperts
+    from vllm.model_executor.layers.fused_moe.experts.deep_gemm_moe import (
+        DeepGemmExperts,
+    )
 
 requires_deep_ep = pytest.mark.skipif(
     not has_deep_ep(),
diff --git a/tests/kernels/moe/test_deepgemm.py b/tests/kernels/moe/test_deepgemm.py
index c2949391c..47700f82a 100644
--- a/tests/kernels/moe/test_deepgemm.py
+++ b/tests/kernels/moe/test_deepgemm.py
@@ -175,7 +175,7 @@ def test_deepgemm_vs_triton(m, n, k, topk, num_experts, monkeypatch, workspace_i
         mp.setenv("VLLM_USE_DEEP_GEMM", "1")
 
         _DeepGemmExperts = importlib.import_module(
-            "vllm.model_executor.layers.fused_moe.deep_gemm_moe"
+            "vllm.model_executor.layers.fused_moe.experts.deep_gemm_moe"
         ).DeepGemmExperts
 
         call_counter = {"cnt": 0}
diff --git a/tests/kernels/moe/test_silu_mul_fp8_quant_deep_gemm.py b/tests/kernels/moe/test_silu_mul_fp8_quant_deep_gemm.py
index 4a447ba7c..ed58db62d 100644
--- a/tests/kernels/moe/test_silu_mul_fp8_quant_deep_gemm.py
+++ b/tests/kernels/moe/test_silu_mul_fp8_quant_deep_gemm.py
@@ -7,7 +7,7 @@ import random
 import pytest
 import torch
 
-from vllm.model_executor.layers.fused_moe.batched_deep_gemm_moe import (
+from vllm.model_executor.layers.fused_moe.experts.batched_deep_gemm_moe import (
     persistent_masked_m_silu_mul_quant,
 )
 from vllm.model_executor.layers.quantization.utils.quant_utils import (
diff --git a/vllm/model_executor/layers/fused_moe/__init__.py b/vllm/model_executor/layers/fused_moe/__init__.py
index f56a2e63b..b342e0c6e 100644
--- a/vllm/model_executor/layers/fused_moe/__init__.py
+++ b/vllm/model_executor/layers/fused_moe/__init__.py
@@ -77,16 +77,18 @@ __all__ = [
 
 if HAS_TRITON:
     # import to register the custom ops
-    from vllm.model_executor.layers.fused_moe.batched_deep_gemm_moe import (
-        BatchedDeepGemmExperts,
-    )
     from vllm.model_executor.layers.fused_moe.cutlass_moe import (
         CutlassBatchedExpertsFp8,
         CutlassExpertsFp8,
         CutlassExpertsW4A8Fp8,
         cutlass_moe_w4a8_fp8,
     )
-    from vllm.model_executor.layers.fused_moe.deep_gemm_moe import DeepGemmExperts
+    from vllm.model_executor.layers.fused_moe.experts.batched_deep_gemm_moe import (
+        BatchedDeepGemmExperts,
+    )
+    from vllm.model_executor.layers.fused_moe.experts.deep_gemm_moe import (
+        DeepGemmExperts,
+    )
     from vllm.model_executor.layers.fused_moe.fused_batched_moe import (
         BatchedTritonExperts,
     )
diff --git a/vllm/model_executor/layers/fused_moe/batched_deep_gemm_moe.py b/vllm/model_executor/layers/fused_moe/experts/batched_deep_gemm_moe.py
similarity index 100%
rename from vllm/model_executor/layers/fused_moe/batched_deep_gemm_moe.py
rename to vllm/model_executor/layers/fused_moe/experts/batched_deep_gemm_moe.py
diff --git a/vllm/model_executor/layers/fused_moe/deep_gemm_moe.py b/vllm/model_executor/layers/fused_moe/experts/deep_gemm_moe.py
similarity index 100%
rename from vllm/model_executor/layers/fused_moe/deep_gemm_moe.py
rename to vllm/model_executor/layers/fused_moe/experts/deep_gemm_moe.py
diff --git a/vllm/model_executor/layers/fused_moe/oracle/fp8.py b/vllm/model_executor/layers/fused_moe/oracle/fp8.py
index 36f35ed5e..d3c70d2d0 100644
--- a/vllm/model_executor/layers/fused_moe/oracle/fp8.py
+++ b/vllm/model_executor/layers/fused_moe/oracle/fp8.py
@@ -131,7 +131,7 @@ def backend_to_kernel_cls(
         return [TritonOrDeepGemmExperts]
 
     elif backend == Fp8MoeBackend.BATCHED_DEEPGEMM:
-        from vllm.model_executor.layers.fused_moe.batched_deep_gemm_moe import (
+        from vllm.model_executor.layers.fused_moe.experts.batched_deep_gemm_moe import (
             BatchedDeepGemmExperts,
         )
 
diff --git a/vllm/model_executor/layers/fused_moe/triton_deep_gemm_moe.py b/vllm/model_executor/layers/fused_moe/triton_deep_gemm_moe.py
index b601806b0..d58dbc20e 100644
--- a/vllm/model_executor/layers/fused_moe/triton_deep_gemm_moe.py
+++ b/vllm/model_executor/layers/fused_moe/triton_deep_gemm_moe.py
@@ -9,7 +9,7 @@ from vllm.model_executor.layers.fused_moe.config import (
     FusedMoEConfig,
     FusedMoEQuantConfig,
 )
-from vllm.model_executor.layers.fused_moe.deep_gemm_moe import (
+from vllm.model_executor.layers.fused_moe.experts.deep_gemm_moe import (
     DeepGemmExperts,
     _valid_deep_gemm,
     _valid_deep_gemm_shape,
diff --git a/vllm/model_executor/warmup/deep_gemm_warmup.py b/vllm/model_executor/warmup/deep_gemm_warmup.py
index 07476906e..e41b2c5e1 100644
--- a/vllm/model_executor/warmup/deep_gemm_warmup.py
+++ b/vllm/model_executor/warmup/deep_gemm_warmup.py
@@ -11,8 +11,8 @@ from tqdm import tqdm
 
 import vllm.envs as envs
 from vllm.distributed.parallel_state import get_dp_group, is_global_first_rank
-from vllm.model_executor.layers.fused_moe.deep_gemm_moe import DeepGemmExperts
 from vllm.model_executor.layers.fused_moe.deep_gemm_utils import compute_aligned_M
+from vllm.model_executor.layers.fused_moe.experts.deep_gemm_moe import DeepGemmExperts
 from vllm.model_executor.layers.fused_moe.layer import FusedMoE
 from vllm.model_executor.layers.fused_moe.triton_deep_gemm_moe import (
     TritonOrDeepGemmExperts,