From bfefdc52d7046bd5c3eae4f4b784a9e0ead2cd05 Mon Sep 17 00:00:00 2001
From: Zaili Wang <109502517+ZailiWang@users.noreply.github.com>
Date: Tue, 21 Jul 2026 09:03:24 +0800
Subject: [PATCH] [CPU] Fix mxfp4 padding size (#31334)

---
 python/sglang/srt/configs/update_config.py    |  5 ++++
 .../sglang/srt/layers/quantization/mxfp4.py   | 23 -------------------
 2 files changed, 5 insertions(+), 23 deletions(-)

diff --git a/python/sglang/srt/configs/update_config.py b/python/sglang/srt/configs/update_config.py
index b0fa5a3d9..4a1968692 100644
--- a/python/sglang/srt/configs/update_config.py
+++ b/python/sglang/srt/configs/update_config.py
@@ -252,6 +252,11 @@ def adjust_config_with_unaligned_cpu_tp(
         )
 
     intermediate_padding_size = tp_size * get_moe_padding_size(weight_block_size)
+    if model_config.quantization == "mxfp4":
+        # For mxfp4 quantization, 2 mxfp4 values are packed to 1 uint8,
+        # so we need to double the intermediate padding size to ensure
+        # the padded intermediate size is divisible by 2 for proper packing.
+        intermediate_padding_size *= 2
     for moe_intermediate_attr in [
         "moe_intermediate_size",
         "intermediate_size",
diff --git a/python/sglang/srt/layers/quantization/mxfp4.py b/python/sglang/srt/layers/quantization/mxfp4.py
index bca97962b..822998fc2 100644
--- a/python/sglang/srt/layers/quantization/mxfp4.py
+++ b/python/sglang/srt/layers/quantization/mxfp4.py
@@ -857,29 +857,6 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
                     layer.w2_weight_bias = Parameter(
                         layer.w2_weight_bias.float(), requires_grad=False
                     )
-                return
-            # Fallback if the TP-sharded layer cannot be AMX-packed
-            from sglang.srt.layers.quantization.mxfp4_tensor import MXFP4QuantizeUtil
-
-            w13_weight = MXFP4QuantizeUtil.dequantize(
-                quantized_data=layer.w13_weight,
-                dtype=torch.bfloat16,
-                scale=layer.w13_weight_scale,
-                block_sizes=[32],
-            )
-            w2_weight = MXFP4QuantizeUtil.dequantize(
-                quantized_data=layer.w2_weight,
-                dtype=torch.bfloat16,
-                scale=layer.w2_weight_scale,
-                block_sizes=[32],
-            )
-            del layer.w13_weight
-            del layer.w2_weight
-            del layer.w13_weight_scale
-            del layer.w2_weight_scale
-            layer.w13_weight = Parameter(w13_weight, requires_grad=False)
-            layer.w2_weight = Parameter(w2_weight, requires_grad=False)
-
             return
         else:
             from triton_kernels.numerics_details.mxfp import upcast_from_mxfp