From f8089976a2535683d430e11d78f7cd80c063ccae Mon Sep 17 00:00:00 2001
From: BBuf <1182563586@qq.com>
Date: Wed, 19 Nov 2025 21:34:40 +0800
Subject: [PATCH] delete useless pad kernel in sgl_moe_align_block_size

---
 .../layers/moe/fused_moe_triton/moe_align_block_size.py    | 7 +------
 1 file changed, 1 insertion(+), 6 deletions(-)

diff --git a/python/sglang/srt/layers/moe/fused_moe_triton/moe_align_block_size.py b/python/sglang/srt/layers/moe/fused_moe_triton/moe_align_block_size.py
index 64d0126d6271..ce1cae66e9e8 100644
--- a/python/sglang/srt/layers/moe/fused_moe_triton/moe_align_block_size.py
+++ b/python/sglang/srt/layers/moe/fused_moe_triton/moe_align_block_size.py
@@ -69,11 +69,6 @@ def moe_align_block_size(
         (num_experts + 2,), dtype=torch.int32, device=topk_ids.device
     )
 
-    # Threshold based on benchmark results
-    fuse_sorted_ids_padding = sorted_ids.shape[0] <= 4096
-    if not fuse_sorted_ids_padding:
-        sorted_ids.fill_(topk_ids.numel())
-
     sgl_moe_align_block_size(
         topk_ids,
         num_experts + 1,
@@ -82,6 +77,6 @@ def moe_align_block_size(
         expert_ids,
         num_tokens_post_pad,
         cumsum_buffer,
-        fuse_sorted_ids_padding,
+        True,
     )
     return sorted_ids, expert_ids, num_tokens_post_pad