Refactor e2e CI (#2276)

Refactor E2E CI to make it clear and faster
1. remove some uesless e2e test
2. remove some uesless function
3. Make sure all test runs with VLLMRunner to avoid oom error
4. Make sure all ops test end with torch.empty_cache to avoid oom error
5. run the test one by one to avoid resource limit error


- vLLM version: v0.10.1.1
- vLLM main:
a344a5aa0a

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

tests/e2e/singlecard/test_camem.py

@@ -17,10 +17,13 @@
 # limitations under the License.
 #
 
+import gc
+
 import torch
-from vllm import LLM, SamplingParams
+from vllm import SamplingParams
 from vllm.utils import GiB_bytes
 
+from tests.e2e.conftest import VllmRunner
 from tests.e2e.utils import fork_new_process_for_each_test
 from vllm_ascend.device_allocator.camem import CaMemAllocator
 
@@ -57,29 +60,37 @@ def test_basic_camem():
     output = x + y + z
     assert torch.allclose(output, torch.ones_like(output) * 3)
 
+    gc.collect()
+    torch.npu.empty_cache()
+    torch.npu.reset_peak_memory_stats()
+
 
 @fork_new_process_for_each_test
 def test_end_to_end():
     free, total = torch.npu.mem_get_info()
     used_bytes_baseline = total - free  # in case other process is running
-    llm = LLM("Qwen/Qwen2.5-0.5B-Instruct", enable_sleep_mode=True)
+
     prompt = "How are you?"
     sampling_params = SamplingParams(temperature=0, max_tokens=10)
-    output = llm.generate(prompt, sampling_params)
 
-    # the benefit of `llm.sleep(level=2)` is mainly CPU memory usage,
-    # which is difficult to measure in the test. therefore, we only
-    # test sleep level 1 here.
-    llm.sleep(level=1)
+    with VllmRunner("Qwen/Qwen3-0.6B",
+                    enforce_eager=True,
+                    enable_sleep_mode=True) as runner:
 
-    free_gpu_bytes_after_sleep, total = torch.npu.mem_get_info()
-    used_bytes = total - free_gpu_bytes_after_sleep - used_bytes_baseline
-    # now the memory usage should be less than the model weights
-    # (0.5B model, 1GiB weights)
-    assert used_bytes < 1 * GiB_bytes
+        output = runner.model.generate(prompt, sampling_params)
+        # the benefit of `llm.sleep(level=2)` is mainly CPU memory usage,
+        # which is difficult to measure in the test. therefore, we only
+        # test sleep level 1 here.
+        runner.model.sleep(level=1)
 
-    llm.wake_up()
-    output2 = llm.generate(prompt, sampling_params)
+        free_gpu_bytes_after_sleep, total = torch.npu.mem_get_info()
+        used_bytes = total - free_gpu_bytes_after_sleep - used_bytes_baseline
+        # now the memory usage should be less than the model weights
+        # (0.5B model, 1GiB weights)
+        assert used_bytes < 1 * GiB_bytes
+
+        runner.model.wake_up()
+        output2 = runner.model.generate(prompt, sampling_params)
 
     # cmp output
     assert output[0].outputs[0].text == output2[0].outputs[0].text