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_embedding.py

@@ -16,24 +16,29 @@
 # This file is a part of the vllm-ascend project.
 # Adapted from vllm/tests/basic_correctness/test_basic_correctness.py
 #
-from collections.abc import Sequence
-from typing import Optional
-
 from modelscope import snapshot_download  # type: ignore[import-untyped]
 
-from tests.e2e.conftest import HfRunner
-from tests.e2e.utils import check_embeddings_close, matryoshka_fy
+from tests.e2e.conftest import HfRunner, VllmRunner
+from tests.e2e.utils import check_embeddings_close
 
 
-def run_embedding_correctness_test(
-    hf_model: "HfRunner",
-    inputs: list[str],
-    vllm_outputs: Sequence[list[float]],
-    dimensions: Optional[int] = None,
-):
-    hf_outputs = hf_model.encode(inputs)
-    if dimensions:
-        hf_outputs = matryoshka_fy(hf_outputs, dimensions)
+def test_embed_models_correctness():
+    queries = ['What is the capital of China?', 'Explain gravity']
+
+    model_name = snapshot_download("Qwen/Qwen3-Embedding-0.6B")
+    with VllmRunner(
+            model_name,
+            task="embed",
+            enforce_eager=True,
+    ) as vllm_runner:
+        vllm_outputs = vllm_runner.encode(queries)
+
+    with HfRunner(
+            model_name,
+            dtype="float32",
+            is_sentence_transformer=True,
+    ) as hf_runner:
+        hf_outputs = hf_runner.encode(queries)
 
     check_embeddings_close(
         embeddings_0_lst=hf_outputs,
@@ -42,27 +47,3 @@ def run_embedding_correctness_test(
         name_1="vllm",
         tol=1e-2,
     )
-
-
-# dummy to avoid pytest collect nothing and exit code 5
-def test_dummy():
-    assert True
-
-
-def test_embed_models_correctness(hf_runner, vllm_runner):
-    queries = ['What is the capital of China?', 'Explain gravity']
-
-    model_name = snapshot_download("Qwen/Qwen3-Embedding-0.6B")
-    with vllm_runner(
-            model_name,
-            task="embed",
-            enforce_eager=True,
-    ) as vllm_model:
-        vllm_outputs = vllm_model.encode(queries)
-
-    with hf_runner(
-            model_name,
-            dtype="float32",
-            is_sentence_transformer=True,
-    ) as hf_model:
-        run_embedding_correctness_test(hf_model, queries, vllm_outputs)