From cdc9dc1fb17360854fbfbc6a5f6a1f768b9de5b7 Mon Sep 17 00:00:00 2001
From: Lei Zhenyuan <zhenyuan.lei@intel.com>
Date: Wed, 11 Feb 2026 09:50:33 +0800
Subject: [PATCH] fix for tma (#4023)

---
 unsloth/kernels/moe/grouped_gemm/interface.py | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/unsloth/kernels/moe/grouped_gemm/interface.py b/unsloth/kernels/moe/grouped_gemm/interface.py
index 572b02702b..554e5fcc03 100644
--- a/unsloth/kernels/moe/grouped_gemm/interface.py
+++ b/unsloth/kernels/moe/grouped_gemm/interface.py
@@ -4,6 +4,7 @@
 import logging
 import warnings
 from dataclasses import asdict
+from unsloth import DEVICE_TYPE
 
 import torch
 import triton
@@ -41,6 +42,8 @@ logger.addHandler(ch)
 # 1. GPU capability >= 9 (Hopper+)
 # 2. Triton version with TMA API (make_tensor_descriptor or _experimental_make_tensor_descriptor)
 def _check_tma_support():
+    if DEVICE_TYPE == "xpu":
+        return False
     import triton.language as tl
 
     gpu_supports_tma = torch.cuda.get_device_capability()[0] >= 9