[Test] Add test for deepseek-v3.2-exp-w8a8

zhangxinyuehfad · zhangxinyuehfad · commit 41ed3ee0c29d · 2025-11-04T23:34:50.000+08:00
Signed-off-by: hfadzxy &lt;starmoon_zhang@163.com&gt;
diff --git a/.github/workflows/_e2e_nightly_multi_node.yaml b/.github/workflows/_e2e_nightly_multi_node.yaml
@@ -103,7 +103,8 @@ jobs:
         - name: Decode kubeconfig from secrets
           run: |
             # Decode and save kubeconfig
-            echo "${{ secrets.KUBECONFIG_B64 }}" | base64 -d > $KUBECONFIG
+            # echo "${{ secrets.KUBECONFIG_B64 }}" | base64 -d > $KUBECONFIG
+            cp /root/.cache/.kube/kubeconfig.yaml "$KUBECONFIG"
 
         - name: Checkout code
           uses: actions/checkout@v4
diff --git a/.github/workflows/_e2e_nightly_single_node.yaml b/.github/workflows/_e2e_nightly_single_node.yaml
@@ -33,6 +33,9 @@ on:
       tests:
         required: true
         type: string
+      name:
+        required: false
+        type: string
 
 # Bash shells do not use ~/.profile or ~/.bashrc so these shells need to be explicitly
 # declared as "shell: bash -el {0}" on steps that need to be properly activated.
@@ -94,6 +97,28 @@ jobs:
           pip install -r requirements-dev.txt
           pip install -v -e .
 
+      - name: Install custom-ops & MLAPO (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/:${LD_LIBRARY_PATH}
+          wget https://vllm-ascend.obs.cn-north-4.myhuaweicloud.com/vllm-ascend/a3/custom_ops-1.0-cp311-cp311-linux_aarch64.whl
+          pip install custom_ops-1.0-cp311-cp311-linux_aarch64.whl
+
+          wget https://vllm-ascend.obs.cn-north-4.myhuaweicloud.com/vllm-ascend/a3/CANN-custom_ops-mlapo-linux.aarch64.run
+          chmod +x ./CANN-custom_ops-mlapo-linux.aarch64.run 
+          ./CANN-custom_ops-mlapo-linux.aarch64.run --quiet --install-path=/vllm-workspace/CANN
+          wget https://vllm-ascend.obs.cn-north-4.myhuaweicloud.com/vllm-ascend/a3/torch_npu-2.7.1%2Bgitb7c90d0-cp311-cp311-linux_aarch64.whl
+          pip install torch_npu-2.7.1+gitb7c90d0-cp311-cp311-linux_aarch64.whl
+          wget https://vllm-ascend.obs.cn-north-4.myhuaweicloud.com/vllm-ascend/a3/libopsproto_rt2.0.so
+          cp libopsproto_rt2.0.so /usr/local/Ascend/ascend-toolkit/8.2.RC1/opp/built-in/op_proto/lib/linux/aarch64/libopsproto_rt2.0.so
+          . /vllm-workspace/CANN/vendors/customize/bin/set_env.bash
+          export LD_PRELOAD=/vllm-workspace/CANN/vendors/customize/op_proto/lib/linux/aarch64/libcust_opsproto_rt2.0.so:${LD_PRELOAD}
+
       - name: Checkout aisbench repo and Install aisbench
         run: |
           git clone https://gitee.com/aisbench/benchmark.git
diff --git a/.github/workflows/vllm_ascend_test_nightly_a3.yaml b/.github/workflows/vllm_ascend_test_nightly_a3.yaml
@@ -28,6 +28,9 @@ on:
   pull_request: 
     branches:
       - 'main'
+  push: 
+    branches:
+      - 'main'
 
 # Bash shells do not use ~/.profile or ~/.bashrc so these shells need to be explicitly
 # declared as "shell: bash -el {0}" on steps that need to be properly activated.
@@ -43,74 +46,78 @@ concurrency:
 jobs:
   single-node-tests:
     name: single-node
-    if: github.event_name == 'schedule' || github.event_name == 'workflow_dispatch'
+    # if: github.event_name == 'schedule' || github.event_name == 'workflow_dispatch'
     strategy:
       fail-fast: false
       matrix:
         test_config:
-          - name: qwen3-32b-in8-a3
-            os: linux-aarch64-a3-4
-            tests: tests/e2e/nightly/models/test_qwen3_32b_int8.py
-          - name: qwen3-32b-int8-a3-feature-stack3
-            os: linux-aarch64-a3-4
-            tests: tests/e2e/nightly/features/test_qwen3_32b_int8_a3_feature_stack3.py
-          - name: qwen3-235b-a22b-w8a8-eplb
-            os: linux-aarch64-a3-16
-            tests: tests/e2e/nightly/models/test_qwen3_235b_a22b_w8a8_eplb.py
-          - name: deepseek-r1-w8a8-eplb
-            os: linux-aarch64-a3-16
-            tests: tests/e2e/nightly/models/test_deepseek_r1_w8a8_eplb.py
-          - name: qwen2-5-vl-7b
-            os: linux-aarch64-a3-4
-            tests: tests/e2e/nightly/models/test_qwen2_5_vl_7b.py
-          - name: qwen2-5-vl-32b
-            os: linux-aarch64-a3-4
-            tests: tests/e2e/nightly/models/test_qwen2_5_vl_32b.py
-          - name: qwen3-32b-int8-prefix-cache
-            os: linux-aarch64-a3-4
-            tests: tests/e2e/nightly/features/test_prefix_cache_qwen3_32b_int8.py
-          - name: deepseek-r1-0528-w8a8
-            os: linux-aarch64-a3-16
-            tests: tests/e2e/nightly/models/test_deepseek_r1_0528_w8a8.py
-          - name: deepseek-r1-0528-w8a8-prefix-cache
+          # - name: qwen3-32b-in8-a3
+          #   os: linux-aarch64-a3-4
+          #   tests: tests/e2e/nightly/models/test_qwen3_32b_int8.py
+          # - name: qwen3-32b-int8-a3-feature-stack3
+          #   os: linux-aarch64-a3-4
+          #   tests: tests/e2e/nightly/features/test_qwen3_32b_int8_a3_feature_stack3.py
+          # - name: qwen3-235b-a22b-w8a8-eplb
+          #   os: linux-aarch64-a3-16
+          #   tests: tests/e2e/nightly/models/test_qwen3_235b_a22b_w8a8_eplb.py
+          # - name: deepseek-r1-w8a8-eplb
+          #   os: linux-aarch64-a3-16
+          #   tests: tests/e2e/nightly/models/test_deepseek_r1_w8a8_eplb.py
+          # - name: qwen2-5-vl-7b
+          #   os: linux-aarch64-a3-4
+          #   tests: tests/e2e/nightly/models/test_qwen2_5_vl_7b.py
+          # - name: qwen2-5-vl-32b
+          #   os: linux-aarch64-a3-4
+          #   tests: tests/e2e/nightly/models/test_qwen2_5_vl_32b.py
+          # - name: qwen3-32b-int8-prefix-cache
+          #   os: linux-aarch64-a3-4
+          #   tests: tests/e2e/nightly/features/test_prefix_cache_qwen3_32b_int8.py
+          # - name: deepseek-r1-0528-w8a8
+          #   os: linux-aarch64-a3-16
+          #   tests: tests/e2e/nightly/models/test_deepseek_r1_0528_w8a8.py
+          # - name: deepseek-r1-0528-w8a8-prefix-cache
+          #   os: linux-aarch64-a3-16
+          #   tests: tests/e2e/nightly/features/test_prefix_cache_deepseek_r1_0528_w8a8.py
+          # - name: qwq-32b-a3
+          #   os: linux-aarch64-a3-4
+          #   tests: tests/e2e/nightly/models/test_qwq_32b.py
+          - name: deepseek3_2-exp-w8a8
             os: linux-aarch64-a3-16
-            tests: tests/e2e/nightly/features/test_prefix_cache_deepseek_r1_0528_w8a8.py
-          - name: qwq-32b-a3
-            os: linux-aarch64-a3-4
-            tests: tests/e2e/nightly/models/test_qwq_32b.py
+            tests: tests/e2e/nightly/models/test_deepseek_v3_2_exp_w8a8.py
     uses: ./.github/workflows/_e2e_nightly_single_node.yaml
     with:
       vllm: v0.11.0
       runner: ${{ matrix.test_config.os }}
       image: swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.3.rc1-a3-ubuntu22.04-py3.11
       tests: ${{ matrix.test_config.tests }}
+      name: ${{ matrix.test_config.name }}
 
   multi-node-tests:
     name: multi-node
-    needs: single-node-tests
-    if: always() && (github.event_name == 'schedule' || github.event_name == 'workflow_dispatch')
+    # needs: single-node-tests
+    # if: always() && (github.event_name == 'schedule' || github.event_name == 'workflow_dispatch')
     strategy:
       fail-fast: false
       max-parallel: 1
       matrix:
         test_config:
-          - name: multi-node-deepseek-pd
-            config_file_path: DeepSeek-V3.yaml
-            size: 2
-          - name: multi-node-qwen3-dp
-            config_file_path: Qwen3-235B-A3B.yaml
-            size: 2
-          - name: multi-node-deepseek3.2-exp-dp
-            config_file_path: tests/e2e/nightly/multi_node/config/models/DeepSeek3_2-Exp-W8A8.yaml
-            size: 2
-          - name: multi-node-dpsk-4node-pd
-            config_file_path: DeepSeek-R1-W8A8.yaml
-            size: 4
-          - name: multi-node-qwenw8a8-2node
-            config_file_path: Qwen3-235B-W8A8.yaml
-            size: 2
-          - name: multi-node-glm-2node
-            config_file_path: GLM-4_5.yaml
+          # - name: multi-node-deepseek-pd
+          #   config_file_path: DeepSeek-V3.yaml
+          #   size: 2
+          # - name: multi-node-qwen3-dp
+          #   config_file_path: Qwen3-235B-A3B.yaml
+          #   size: 2
+          # - name: multi-node-dpsk-4node-pd
+          #   config_file_path: DeepSeek-R1-W8A8.yaml
+          #   size: 4
+          # - name: multi-node-qwenw8a8-2node
+          #   config_file_path: Qwen3-235B-W8A8.yaml
+          #   size: 2
+          # - name: multi-node-glm-2node
+          #   config_file_path: GLM-4_5.yaml
+          #   size: 2
+          - name: multi-node-dpsk3.2-exp-dp
+            config_file_path: DeepSeek-V3_2-Exp-W8A8.yaml
             size: 2
     uses: ./.github/workflows/_e2e_nightly_multi_node.yaml
     with:
diff --git a/tests/e2e/nightly/models/test_deepseek_v3_2_exp_w8a8.py b/tests/e2e/nightly/models/test_deepseek_v3_2_exp_w8a8.py
@@ -0,0 +1,120 @@
+# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
+# Copyright 2023 The vLLM team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# This file is a part of the vllm-ascend project.
+#
+from typing import Any
+
+import openai
+import pytest
+from vllm.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",
+]
+
+TENSOR_PARALLELS = [8]
+DATA_PARALLELS = [2]
+FULL_GRAPH = [True, False]
+
+prompts = [
+    "San Francisco is a",
+]
+
+api_keyword_args = {
+    "max_tokens": 10,
+}
+
+aisbench_cases = [{
+    "case_type": "accuracy",
+    "dataset_path": "vllm-ascend/gsm8k-lite",
+    "request_conf": "vllm_api_general_chat",
+    "dataset_conf": "gsm8k/gsm8k_gen_0_shot_cot_chat_prompt",
+    "max_out_len": 4096,
+    "batch_size": 8,
+    "baseline": 95,
+    "threshold": 5
+}, {
+    "case_type": "performance",
+    "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,
+    "max_out_len": 1500,
+    "batch_size": 8,
+    "request_rate": 0,
+    "baseline": 1,
+    "threshold": 0.97
+}]
+
+
+@pytest.mark.asyncio
+@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:
+    port = get_open_port()
+    env_dict = {"HCCL_BUFFSIZE": "1024", "VLLM_ASCEND_ENABLE_MLAPO": "1"}
+    server_args = [
+        "--no-enable-prefix-caching",
+        "--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",
+        "--additional-config",
+        '{"ascend_scheduler_config":{"enabled":true},'
+        '"torchair_graph_config":{"enabled":true,"graph_batch_sizes":[16]}}',
+    ]
+    if full_graph:
+        server_args += [
+            "--compilation-config",
+            '{"cudagraph_capture": [16], "cudagraph_model":"FULL_DECODE_ONLY"}'
+        ]
+    request_keyword_args: dict[str, Any] = {
+        **api_keyword_args,
+    }
+    with RemoteOpenAIServer(model,
+                            server_args,
+                            server_port=port,
+                            env_dict=env_dict,
+                            auto_port=False) as server:
+        client = server.get_async_client()
+        batch = await client.completions.create(
+            model=model,
+            prompt=prompts,
+            **request_keyword_args,
+        )
+        choices: list[openai.types.CompletionChoice] = batch.choices
+        assert choices[0].text, "empty response"
+        # aisbench test
+        run_aisbench_cases(model, port, aisbench_cases)
diff --git a/tests/e2e/nightly/multi_node/config/models/DeepSeek-V3_2-Exp-W8A8.yaml b/tests/e2e/nightly/multi_node/config/models/DeepSeek-V3_2-Exp-W8A8.yaml
@@ -0,0 +1,55 @@
+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: 1
+
+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 
+        --additional-config '{"ascend_scheduler_config":{"enabled":true},"torchair_graph_config":{"enabled":true,"graph_batch_sizes":[16]}}'
+
+  -
+    server_cmd: >
+        vllm serve Yanguan/DeepSeek-V3.2-Exp-bf16 \
+        --host 0.0.0.0 
+        --port $SERVER_PORT 
+        --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 
+        --additional-config '{"ascend_scheduler_config":{"enabled":true},"torchair_graph_config":{"enabled":true,"graph_batch_sizes":[16]}}'
+benchmarks:
diff --git a/tests/e2e/nightly/multi_node/config/models/DeepSeek3_2-Exp-W8A8.yaml b/tests/e2e/nightly/multi_node/config/models/DeepSeek3_2-Exp-W8A8.yaml
diff --git a/tests/e2e/nightly/multi_node/scripts/run.sh b/tests/e2e/nightly/multi_node/scripts/run.sh
