From 46dced64ea5a3cf345464b1cf5392f9b6682cfae Mon Sep 17 00:00:00 2001
From: Qiaolin Yu <liin1211@outlook.com>
Date: Thu, 5 Mar 2026 14:50:40 -0800
Subject: [PATCH] Adjust padding size to improve triton_kernels moe performance
 (#19174)

---
 .../fused_moe_triton/triton_kernels_moe.py    | 22 +++++++++++++++----
 .../sglang/srt/layers/quantization/mxfp4.py   |  8 +++----
 2 files changed, 21 insertions(+), 9 deletions(-)

diff --git a/python/sglang/srt/layers/moe/fused_moe_triton/triton_kernels_moe.py b/python/sglang/srt/layers/moe/fused_moe_triton/triton_kernels_moe.py
index e14ffe950..b7e6cd7b5 100644
--- a/python/sglang/srt/layers/moe/fused_moe_triton/triton_kernels_moe.py
+++ b/python/sglang/srt/layers/moe/fused_moe_triton/triton_kernels_moe.py
@@ -271,7 +271,9 @@ def triton_kernel_fused_experts_with_bias(
     # feature check
     assert inplace is False, "Inplace is not supported in new triton MoE kernel"
 
-    E, _, _ = w1.shape
+    M, K = hidden_states.shape
+    E, _, N = w1.shape
+    n_expts_act = routing_data.n_expts_act
 
     if global_num_experts == -1:
         global_num_experts = E
@@ -292,7 +294,16 @@ def triton_kernel_fused_experts_with_bias(
         2,
     )
 
-    intermediate_cache = matmul_ogs(
+    intermediate_cache = torch.empty(
+        (1, M * n_expts_act, N // 2),
+        device=hidden_states.device,
+        dtype=hidden_states.dtype,
+    )
+    output = torch.empty(
+        (1, M, K), device=hidden_states.device, dtype=hidden_states.dtype
+    )
+
+    matmul_ogs(
         hidden_states,
         w1,
         b1,
@@ -301,14 +312,17 @@ def triton_kernel_fused_experts_with_bias(
         precision_config=w1_pcg,
         gammas=routing_data.gate_scal if apply_router_weight_on_input else None,
         fused_activation=act,
+        y=intermediate_cache,
     )
 
-    return matmul_ogs(
-        intermediate_cache,
+    matmul_ogs(
+        intermediate_cache.view(M * n_expts_act, N // 2),
         w2,
         b2,
         routing_data,
         scatter_indx=scatter_indx,
         precision_config=w2_pcg,
         gammas=None if apply_router_weight_on_input else routing_data.gate_scal,
+        y=output,
     )
+    return output.view(M, K)
diff --git a/python/sglang/srt/layers/quantization/mxfp4.py b/python/sglang/srt/layers/quantization/mxfp4.py
index 81c326966..8a0ec4729 100644
--- a/python/sglang/srt/layers/quantization/mxfp4.py
+++ b/python/sglang/srt/layers/quantization/mxfp4.py
@@ -335,6 +335,7 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
         scale_dtype = torch.uint8
         self.with_bias = with_bias
         mxfp4_block = 32
+        triton_kernels_padding_alignment = 64
 
         # pad the intermediate size to be a multiple of 2 * mxfp4_block
         # for to hold non-uniform sharded tensor as well as swizzling
@@ -347,7 +348,7 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
                 hidden_size = round_up(hidden_size, 256)
             else:
                 intermediate_size_per_partition_after_pad = round_up(
-                    intermediate_size_per_partition, 64
+                    intermediate_size_per_partition, triton_kernels_padding_alignment
                 )
         elif _use_aiter:
 
@@ -362,11 +363,8 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
                 - layer.intermediate_size_per_partition
             )
         elif has_triton_kernels:
-            # TODO: this is a hack to make
-            # intermediate_size_per_partition_after_pad the same as the
-            # per_rank_intermediate_size during weight loading
             intermediate_size_per_partition_after_pad = round_up(
-                intermediate_size_per_partition, mxfp4_block
+                intermediate_size_per_partition, triton_kernels_padding_alignment
             )
 
         self.intermediate_size_per_partition = intermediate_size_per_partition_after_pad