[CI] Add DeepSeek-V3.2-W8A8 nightly ci test (#4633)

### What this PR does / why we need it?
Add DeepSeek-V3.2-W8A8 nightly ci test：

DeepSeek-V3.2-W8A8 1node DP2+TP8
:tests/e2e/nightly/models/test_deepseek_v3_2_w8a8.py

### Does this PR introduce _any_ user-facing change

- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

Signed-off-by: hfadzxy <starmoon_zhang@163.com>

.github/workflows/_e2e_nightly_single_node.yaml

@@ -110,19 +110,6 @@ jobs:
           fi
           cd ..
 
-      - name: Install custom-ops (for DeepSeek-V3.2-Exp)
-        if: ${{ inputs.name == 'deepseek3_2-exp-w8a8' }}
-        shell: bash -l {0}
-        run: |
-          wget https://vllm-ascend.obs.cn-north-4.myhuaweicloud.com/vllm-ascend/a3/CANN-custom_ops-sfa-linux.aarch64.run
-          chmod +x ./CANN-custom_ops-sfa-linux.aarch64.run
-          ./CANN-custom_ops-sfa-linux.aarch64.run --quiet
-          export ASCEND_CUSTOM_OPP_PATH=/usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize:${ASCEND_CUSTOM_OPP_PATH}
-          export LD_LIBRARY_PATH=/usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/op_api/lib/:${LD_LIBRARY_PATH}
-          wget https://vllm-ascend.obs.cn-north-4.myhuaweicloud.com/vllm-ascend/a3/custom_ops-1.0-cp311-cp311-linux_aarch64.whl
-          pip install custom_ops-1.0-cp311-cp311-linux_aarch64.whl
-          . /usr/local/Ascend/ascend-toolkit/set_env.sh
-
       - name: Install Ascend toolkit & triton_ascend
         shell: bash -l {0}
         run: |

.github/workflows/nightly_test_a3.yaml

@@ -150,13 +150,12 @@ jobs:
           - name: kimi-k2-thinking
             os: linux-aarch64-a3-16
             tests: tests/e2e/nightly/single_node/models/test_kimi_k2_thinking.py
-          # TODO: Replace deepseek3.2-exp with deepseek3.2 after nightly tests pass
-          # - name: deepseek3_2-exp-w8a8
-          #   os: linux-aarch64-a3-16
-          #   tests: tests/e2e/nightly/single_node/models/test_deepseek_v3_2_exp_w8a8.py
-          - name: deepseek-r1-w8a8-hmb
+          - name: deepseek-r1-w8a8-hbm
             os: linux-aarch64-a3-16
-            tests: tests/e2e/nightly/single_node/models/test_deepseek_r1_w8a8_hmb.py
+            tests: tests/e2e/nightly/single_node/models/test_deepseek_r1_w8a8_hbm.py
+          - name: deepseek3_2-w8a8
+            os: linux-aarch64-a3-16
+            tests: tests/e2e/nightly/single_node/models/test_deepseek_v3_2_w8a8.py
     uses: ./.github/workflows/_e2e_nightly_single_node.yaml
     with:
       vllm: v0.13.0

tests/e2e/nightly/multi_node/config/DeepSeek-V3_2-Exp-bf16.yaml

@@ -1,51 +0,0 @@
-test_name: "test DeepSeek-V3.2-Exp-bf16 multi-dp"
-model: "Yanguan/DeepSeek-V3.2-Exp-bf16"
-num_nodes: 2
-npu_per_node: 16
-env_common:
-  VLLM_USE_MODELSCOPE: true
-  OMP_PROC_BIND: false
-  OMP_NUM_THREADS: 100
-  HCCL_BUFFSIZE: 1024
-  SERVER_PORT: 8080
-  VLLM_ASCEND_ENABLE_MLAPO: 0
-
-deployment:
-  -
-    server_cmd: >
-        vllm serve "Yanguan/DeepSeek-V3.2-Exp-bf16"
-        --host 0.0.0.0
-        --port $SERVER_PORT
-        --data-parallel-address $LOCAL_IP
-        --data-parallel-size 2
-        --data-parallel-size-local 1
-        --data-parallel-rpc-port 13389 
-        --tensor-parallel-size 16 
-        --seed 1024 
-        --enable-expert-parallel 
-        --max-num-seqs 16 
-        --max-model-len 17450 
-        --max-num-batched-tokens 17450 
-        --trust-remote-code 
-        --no-enable-prefix-caching 
-        --gpu-memory-utilization 0.9 
-
-  -
-    server_cmd: >
-        vllm serve "Yanguan/DeepSeek-V3.2-Exp-bf16"
-        --headless
-        --data-parallel-size 2 
-        --data-parallel-size-local 1 
-        --data-parallel-start-rank 1 
-        --data-parallel-address $MASTER_IP
-        --data-parallel-rpc-port 13389 
-        --tensor-parallel-size 16 
-        --seed 1024 
-        --max-num-seqs 16 
-        --max-model-len 17450 
-        --max-num-batched-tokens 17450 
-        --enable-expert-parallel 
-        --trust-remote-code 
-        --no-enable-prefix-caching 
-        --gpu-memory-utilization 0.92 
-benchmarks:

tests/e2e/nightly/single_node/models/test_deepseek_v3_2_w8a8.py

@@ -23,13 +23,10 @@ from vllm.utils.network_utils import get_open_port
 from tests.e2e.conftest import RemoteOpenAIServer
 from tools.aisbench import run_aisbench_cases
 
-MODELS = [
-    "vllm-ascend/DeepSeek-V3.2-Exp-W8A8",
-]
+MODELS = ["vllm-ascend/DeepSeek-V3.2-W8A8"]
 
 TENSOR_PARALLELS = [8]
 DATA_PARALLELS = [2]
-FULL_GRAPH = [True, False]
 
 prompts = [
     "San Francisco is a",
@@ -53,11 +50,11 @@ aisbench_cases = [{
     "dataset_path": "vllm-ascend/GSM8K-in3500-bs400",
     "request_conf": "vllm_api_stream_chat",
     "dataset_conf": "gsm8k/gsm8k_gen_0_shot_cot_str_perf",
-    "num_prompts": 16,
+    "num_prompts": 100,
     "max_out_len": 1500,
-    "batch_size": 8,
-    "request_rate": 0,
-    "baseline": 1,
+    "batch_size": 4,
+    "request_rate": 11.2,
+    "baseline": 120,
     "threshold": 0.97
 }]
 
@@ -66,24 +63,29 @@ aisbench_cases = [{
 @pytest.mark.parametrize("model", MODELS)
 @pytest.mark.parametrize("tp_size", TENSOR_PARALLELS)
 @pytest.mark.parametrize("dp_size", DATA_PARALLELS)
-@pytest.mark.parametrize("full_graph", FULL_GRAPH)
-async def test_models(model: str, tp_size: int, dp_size: int,
-                      full_graph: bool) -> None:
+async def test_models(model: str, tp_size: int, dp_size: int) -> None:
     port = get_open_port()
-    env_dict = {"HCCL_BUFFSIZE": "1024", "VLLM_ASCEND_ENABLE_MLAPO": "0"}
+    env_dict = {
+        "HCCL_OP_EXPANSION_MODE": "AIV",
+        "OMP_PROC_BIND": "false",
+        "OMP_NUM_THREADS": "1",
+        "HCCL_BUFFSIZE": "1024",
+        "VLLM_ASCEND_ENABLE_MLAPO": "1",
+        "PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
+        "VLLM_ASCEND_ENABLE_FLASHCOMM1": "0",
+    }
+
     server_args = [
-        "--no-enable-prefix-caching", "--enable-expert-parallel",
-        "--tensor-parallel-size",
+        "--enable-expert-parallel", "--tensor-parallel-size",
         str(tp_size), "--data-parallel-size",
         str(dp_size), "--port",
-        str(port), "--max-model-len", "16384", "--max-num-batched-tokens",
-        "16384", "--block-size", "16", "--trust-remote-code", "--quantization",
-        "ascend", "--gpu-memory-utilization", "0.9"
-    ]
-    if full_graph:
-        server_args += [
-            "--compilation-config",
-            '{"cudagraph_capture": [16], "cudagraph_model":"FULL_DECODE_ONLY"}'
+        str(port), "--max-model-len", "8192", "--max-num-batched-tokens",
+        "8192", "--max-num-seqs", "4", "--trust-remote-code", "--quantization",
+        "ascend", "--gpu-memory-utilization", "0.92", "--compilation-config",
+        '{"cudagraph_capture_sizes":[3, 6, 9, 12], "cudagraph_mode":"FULL_DECODE_ONLY"}',
+        "--speculative-config",
+        '{"num_speculative_tokens": 2, "method":"deepseek_mtp"}',
+        "--reasoning-parser", "deepseek_v3", "--tokenizer_mode", "deepseek_v32"
     ]
     request_keyword_args: dict[str, Any] = {
         **api_keyword_args,