[Build] Avoid building too many extensions (#1624)

tests/kernels/test_attention.py

@@ -6,7 +6,7 @@ import torch
 from xformers import ops as xops
 from xformers.ops.fmha.attn_bias import BlockDiagonalCausalMask
 
-from vllm import attention_ops
+from vllm._C import ops
 from vllm.utils import get_max_shared_memory_bytes
 
 FLOAT32_BYTES = torch.finfo(torch.float).bits // 8
@@ -165,7 +165,7 @@ def test_paged_attention(
     # Call the paged attention kernel.
     output = torch.empty_like(query)
     if version == "v1":
-        attention_ops.paged_attention_v1(
+        ops.paged_attention_v1(
             output,
             query,
             key_cache,
@@ -194,7 +194,7 @@ def test_paged_attention(
             device=output.device,
         )
         max_logits = torch.empty_like(exp_sums)
-        attention_ops.paged_attention_v2(
+        ops.paged_attention_v2(
             output,
             exp_sums,
             max_logits,