[Core] Make V1 work and enable V1 engine test (#389)

1. Make sure the version is string before parse in collect_env
2. Add basic V1 engine test

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>

tests/ops/test_fused_moe.py

@@ -100,3 +100,4 @@ def test_fused_experts(
                                  e_map)
         # TODO: The native params are: atol=2e-2, rtol=0, maybe related to the nan problem
         torch.testing.assert_close(output, torch_output, atol=4e-2, rtol=1)
+    torch.npu.empty_cache()

tests/test_offline_inference.py

@@ -45,8 +45,6 @@ def test_models(
     dtype: str,
     max_tokens: int,
 ) -> None:
-    os.environ["VLLM_ATTENTION_BACKEND"] = "ASCEND"
-
     # 5042 tokens for gemma2
     # gemma2 has alternating sliding window size of 4096
     # we need a prompt with more than 4096 tokens to test the sliding window
@@ -60,3 +58,8 @@ def test_models(
                     enforce_eager=False,
                     gpu_memory_utilization=0.7) as vllm_model:
         vllm_model.generate_greedy(example_prompts, max_tokens)
+
+
+if __name__ == "__main__":
+    import pytest
+    pytest.main([__file__])