[Torch 2.11] Guard torch._C._cpu attribute checks for forward compatibility (#35673)

Signed-off-by: atalman <atalman@fb.com>

vllm/model_executor/layers/fused_moe/cpu_fused_moe.py

@@ -280,7 +280,7 @@ class CPUFusedMOE:
         if not (w13_output_size % 32 == 0 and w2_output_size % 32 == 0):
             return False, "none"
 
-        supports_amx = torch._C._cpu._is_amx_tile_supported()
+        supports_amx = torch.cpu._is_amx_tile_supported()
 
         if (
             supports_amx

vllm/model_executor/layers/quantization/cpu_wna16.py

@@ -292,7 +292,7 @@ class CPUAWQLinearMethod(LinearMethodBase):
 
 
 def _get_isa_hint(dtype: torch.dtype) -> str:
-    supports_amx = torch._C._cpu._is_amx_tile_supported()
+    supports_amx = torch.cpu._is_amx_tile_supported()
     if supports_amx and dtype in (torch.bfloat16,):
         return "amx"
     else:

vllm/model_executor/layers/utils.py

@@ -212,7 +212,7 @@ direct_register_custom_op(
 
 def check_cpu_sgl_kernel(n: int, k: int, dtype: torch.dtype) -> bool:
     return (
-        torch._C._cpu._is_amx_tile_supported()
+        torch.cpu._is_amx_tile_supported()
         and (dtype in (torch.bfloat16, torch.int8))
         and k % 32 == 0
         and n % 16 == 0