Fix CI xeon test with triton 3.3.1 (#8086)

python/sglang/srt/layers/quantization/fp8_kernel.py

@@ -29,6 +29,7 @@ from sglang.srt.utils import (
     direct_register_custom_op,
     get_device_core_count,
     get_device_name,
+    is_cpu,
     is_cuda,
     is_hip,
     log_info_on_rank0,
@@ -37,6 +38,7 @@ from sglang.srt.utils import (
 
 _is_hip = is_hip()
 _is_cuda = is_cuda()
+_is_cpu = is_cpu()
 
 if _is_cuda:
     from sgl_kernel import (
@@ -1168,7 +1170,7 @@ def scaled_fp8_quant(
     return output, scale
 
 
-@triton.autotune(
+fp8_autotune = triton.autotune(
     configs=[
         triton.Config({"BLOCK_M": block_m}, num_warps=num_warps)
         for block_m in [16, 32, 64, 128]
@@ -1176,6 +1178,8 @@ def scaled_fp8_quant(
     ],
     key=["K", "BLOCK_K", "M_ALIGNMENT"],
 )
+
+
 @triton.jit
 def _per_token_group_quant_fp8_hopper_moe_mn_major(
     a,  # (M, K):(K, 1)
@@ -1221,6 +1225,12 @@ def _per_token_group_quant_fp8_hopper_moe_mn_major(
         tl.store(sfa_ptrs, inp_amax / 448.0, mask=coord_m < m)
 
 
+if not _is_cpu:
+    _per_token_group_quant_fp8_hopper_moe_mn_major = fp8_autotune(
+        _per_token_group_quant_fp8_hopper_moe_mn_major
+    )
+
+
 def per_token_group_quant_fp8_hopper_moe_mn_major(
     A: torch.Tensor,
     expert_offsets: torch.Tensor,