[CI] Enable linux-aarch64-a2 (64GB) and tp2 * 2 max-parallel to speed up CI (#2065)

### What this PR does / why we need it?
Currently our workflow run time takes about 3 hours in total, which
seriously affects the developer experience, so it is urgent to have a
optimization, after this pr, It is expected that the running time of the
full CI can be shortened to 1h40min.

- Enable linux-aarch64-a2 (64GB) to replace linux-arm64-npu (32GB)
- Change TP4 ---> TP2 * 2 max-parallel
- Move DeepSeek-V2-Lite-W8A8 to single card test

### Does this PR introduce _any_ user-facing change?
No


- vLLM version: v0.10.0
- vLLM main:
a2480251ec

---------

Signed-off-by: wangli <wangli858794774@gmail.com>

tests/e2e/multicard/test_fused_moe_allgather_ep.py

@@ -46,7 +46,7 @@ def test_generate_with_allgather():
     sampling_params = SamplingParams(max_tokens=100, temperature=0.0)
 
     with VllmRunner(snapshot_download("vllm-ascend/DeepSeek-V3-Pruning"),
-                    tensor_parallel_size=4,
+                    tensor_parallel_size=2,
                     enforce_eager=True,
                     max_model_len=1024,
                     dtype="auto",
@@ -74,7 +74,7 @@ def test_generate_with_alltoall():
     sampling_params = SamplingParams(max_tokens=100, temperature=0.0)
 
     with VllmRunner(snapshot_download("vllm-ascend/DeepSeek-V3-Pruning"),
-                    tensor_parallel_size=4,
+                    tensor_parallel_size=2,
                     enforce_eager=True,
                     max_model_len=1024,
                     dtype="auto",

tests/e2e/multicard/test_offline_inference_distributed.py

@@ -42,7 +42,7 @@ def test_models_distributed_QwQ():
     with VllmRunner(
             "Qwen/QwQ-32B",
             dtype=dtype,
-            tensor_parallel_size=4,
+            tensor_parallel_size=2,
             distributed_executor_backend="mp",
     ) as vllm_model:
         vllm_model.generate_greedy(example_prompts, max_tokens)
@@ -57,7 +57,7 @@ def test_models_distributed_DeepSeek_multistream_moe():
     with VllmRunner(
             "vllm-ascend/DeepSeek-V3-Pruning",
             dtype=dtype,
-            tensor_parallel_size=4,
+            tensor_parallel_size=2,
             distributed_executor_backend="mp",
             additional_config={
                 "torchair_graph_config": {
@@ -82,7 +82,7 @@ def test_models_distributed_DeepSeek_dbo():
     with VllmRunner(
             "deepseek-ai/DeepSeek-V2-Lite",
             dtype=dtype,
-            tensor_parallel_size=4,
+            tensor_parallel_size=2,
             distributed_executor_backend="mp",
     ) as vllm_model:
         model_arch = 'DeepseekV2ForCausalLM'
@@ -106,7 +106,7 @@ def test_models_distributed_DeepSeekV3_dbo():
     with VllmRunner(
             "vllm-ascend/DeepSeek-V3-Pruning",
             dtype=dtype,
-            tensor_parallel_size=4,
+            tensor_parallel_size=2,
             distributed_executor_backend="mp",
     ) as vllm_model:
         model_arch = 'DeepseekV3ForCausalLM'
@@ -118,24 +118,6 @@ def test_models_distributed_DeepSeekV3_dbo():
         vllm_model.generate(example_prompts, sampling_params)
 
 
-@pytest.mark.skip(reason="Due to OOM,waiting for 1311pr to merge in")
-def test_models_distributed_DeepSeek_W8A8():
-    example_prompts = [
-        "Hello, my name is",
-    ]
-    max_tokens = 5
-
-    with VllmRunner(
-            snapshot_download("vllm-ascend/DeepSeek-V2-Lite-W8A8"),
-            max_model_len=8192,
-            enforce_eager=True,
-            dtype="auto",
-            tensor_parallel_size=4,
-            quantization="ascend",
-    ) as vllm_model:
-        vllm_model.generate_greedy(example_prompts, max_tokens)
-
-
 def test_models_distributed_pangu():
     example_prompts = [
         "Hello, my name is",
@@ -147,7 +129,7 @@ def test_models_distributed_pangu():
             max_model_len=8192,
             enforce_eager=True,
             dtype="auto",
-            tensor_parallel_size=4,
+            tensor_parallel_size=2,
             distributed_executor_backend="mp",
     ) as vllm_model:
         vllm_model.generate_greedy(example_prompts, max_tokens)
@@ -169,7 +151,7 @@ def test_models_distributed_topk() -> None:
     with VllmRunner(
             "deepseek-ai/DeepSeek-V2-Lite",
             dtype=dtype,
-            tensor_parallel_size=4,
+            tensor_parallel_size=2,
             distributed_executor_backend="mp",
     ) as vllm_model:
         vllm_model.generate(example_prompts, sampling_params)
@@ -186,7 +168,7 @@ def test_models_distributed_Qwen3_W8A8():
             max_model_len=8192,
             enforce_eager=True,
             dtype="auto",
-            tensor_parallel_size=4,
+            tensor_parallel_size=2,
             quantization="ascend",
     ) as vllm_model:
         vllm_model.generate_greedy(example_prompts, max_tokens)

tests/e2e/multicard/test_pipeline_parallel.py

@@ -22,7 +22,7 @@ MODELS = [
     "Qwen/Qwen3-0.6B",
 ]
 
-TENSOR_PARALLELS = [2]
+TENSOR_PARALLELS = [1]
 PIPELINE_PARALLELS = [2]
 DIST_EXECUTOR_BACKEND = ["mp", "ray"]
 

tests/e2e/multicard/test_torchair_graph_mode.py

@@ -30,7 +30,7 @@ os.environ["PYTORCH_NPU_ALLOC_CONF"] = "max_split_size_mb:256"
 def _deepseek_torchair_test_fixture(
     additional_config: Dict,
     *,
-    tensor_parallel_size=4,
+    tensor_parallel_size=2,
 ):
     example_prompts = [
         "Hello, my name is",
@@ -98,7 +98,7 @@ def test_e2e_deepseekv3_with_torchair_ms_mla():
 def _pangu_torchair_test_fixture(
     additional_config: Dict,
     *,
-    tensor_parallel_size=4,
+    tensor_parallel_size=2,
 ):
     example_prompts = [
         "Hello, my name is",