[easy] Print number of needed GPUs in skip message (vllm-project#17594)

Signed-off-by: rzou <[email protected]>
Signed-off-by: Mu Huai <[email protected]>

Author: zou3519

tests/compile/test_basic_correctness.py

@@ -103,7 +103,8 @@ def test_compile_correctness(
     method = test_setting.method
     fullgraph = test_setting.fullgraph
     if cuda_device_count_stateless() != pp_size * tp_size:
-        pytest.skip("Not correct CUDA devices for the test.")
+        pytest.skip(f"Need exactly {pp_size}*{tp_size} CUDA gpus but got "
+                    f"{cuda_device_count_stateless()}")
 
     with monkeypatch.context() as m:
         m.setenv("VLLM_ATTENTION_BACKEND", attn_backend)